lists of operations for forward pass for alexnet and googlenet
Last active
August 29, 2015 14:17
-
-
Save moskewcz/a40f5f170c18c9b151f7 to your computer and use it in GitHub Desktop.
DCNN ops examples
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| data = NDA("data",num_img,3,227,227) # SOURCE num,chan,y,x | |
| conv1_filts = NDA("conv1_filts",96,3,11,11) # SOURCE out_chan,in_chan,y,x | |
| conv1_biases = NDA("conv1_biases",96) # SOURCE out_chan | |
| conv1 = NDA("conv1",num_img,96,55,55) # num,chan,y,x | |
| data_one_row_per_patch_buf = NDA("data_one_row_per_patch_buf",3025,363) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=data[i,:,:,:], dest=data_one_row_per_patch_buf, in_pad="0 0 0 0",stride="4 4" ) # one copy per output elem | |
| conv1 = data_one_row_per_patch_buf * transpose(reshape(conv1_filts,96,363)) # sgemm: MxNxK == 3025x363x96 | |
| ReLU(name="relu1",in_place=[conv1]) | |
| norm1 = NDA("norm1",num_img,96,55,55) # num,chan,y,x | |
| LRN(name="norm1",bots=[ conv1 ],tops=[ norm1 ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| pool1 = NDA("pool1",num_img,96,27,27) # num,chan,y,x | |
| Pooling(name="pool1",bots=[ norm1 ],tops=[ pool1 ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| conv2_filts = NDA("conv2_filts",256,96,5,5) # SOURCE out_chan,in_chan,y,x | |
| conv2_biases = NDA("conv2_biases",256) # SOURCE out_chan | |
| conv2 = NDA("conv2",num_img,256,27,27) # num,chan,y,x | |
| pool1_one_row_per_patch_buf = NDA("pool1_one_row_per_patch_buf",729,2400) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=pool1[i,:,:,:], dest=pool1_one_row_per_patch_buf, in_pad="2 2 2 2",stride="1 1" ) # one copy per output elem | |
| conv2 = pool1_one_row_per_patch_buf * transpose(reshape(conv2_filts,256,2400)) # sgemm: MxNxK == 729x2400x256 | |
| ReLU(name="relu2",in_place=[conv2]) | |
| norm2 = NDA("norm2",num_img,256,27,27) # num,chan,y,x | |
| LRN(name="norm2",bots=[ conv2 ],tops=[ norm2 ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| pool2 = NDA("pool2",num_img,256,13,13) # num,chan,y,x | |
| Pooling(name="pool2",bots=[ norm2 ],tops=[ pool2 ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| conv3_filts = NDA("conv3_filts",384,256,3,3) # SOURCE out_chan,in_chan,y,x | |
| conv3_biases = NDA("conv3_biases",384) # SOURCE out_chan | |
| conv3 = NDA("conv3",num_img,384,13,13) # num,chan,y,x | |
| pool2_one_row_per_patch_buf = NDA("pool2_one_row_per_patch_buf",169,2304) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=pool2[i,:,:,:], dest=pool2_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| conv3 = pool2_one_row_per_patch_buf * transpose(reshape(conv3_filts,384,2304)) # sgemm: MxNxK == 169x2304x384 | |
| ReLU(name="relu3",in_place=[conv3]) | |
| conv4_filts = NDA("conv4_filts",384,384,3,3) # SOURCE out_chan,in_chan,y,x | |
| conv4_biases = NDA("conv4_biases",384) # SOURCE out_chan | |
| conv4 = NDA("conv4",num_img,384,13,13) # num,chan,y,x | |
| conv3_one_row_per_patch_buf = NDA("conv3_one_row_per_patch_buf",169,3456) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=conv3[i,:,:,:], dest=conv3_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| conv4 = conv3_one_row_per_patch_buf * transpose(reshape(conv4_filts,384,3456)) # sgemm: MxNxK == 169x3456x384 | |
| ReLU(name="relu4",in_place=[conv4]) | |
| conv5_filts = NDA("conv5_filts",256,384,3,3) # SOURCE out_chan,in_chan,y,x | |
| conv5_biases = NDA("conv5_biases",256) # SOURCE out_chan | |
| conv5 = NDA("conv5",num_img,256,13,13) # num,chan,y,x | |
| conv4_one_row_per_patch_buf = NDA("conv4_one_row_per_patch_buf",169,3456) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=conv4[i,:,:,:], dest=conv4_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| conv5 = conv4_one_row_per_patch_buf * transpose(reshape(conv5_filts,256,3456)) # sgemm: MxNxK == 169x3456x256 | |
| ReLU(name="relu5",in_place=[conv5]) | |
| pool5 = NDA("pool5",num_img,256,6,6) # num,chan,y,x | |
| Pooling(name="pool5",bots=[ conv5 ],tops=[ pool5 ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| fc6_conv_filts = NDA("fc6_conv_filts",4096,256,6,6) # SOURCE out_chan,in_chan,y,x | |
| fc6_conv_biases = NDA("fc6_conv_biases",4096) # SOURCE out_chan | |
| fc6 = NDA("fc6",num_img,4096,1,1) # num,chan,y,x | |
| pool5_one_row_per_patch_buf = NDA("pool5_one_row_per_patch_buf",1,9216) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=pool5[i,:,:,:], dest=pool5_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| fc6 = pool5_one_row_per_patch_buf * transpose(reshape(fc6_conv_filts,4096,9216)) # sgemm: MxNxK == 1x9216x4096 | |
| ReLU(name="relu6",in_place=[fc6]) | |
| Dropout(name="drop6",in_place=[fc6]) | |
| fc7_conv_filts = NDA("fc7_conv_filts",4096,4096,1,1) # SOURCE out_chan,in_chan,y,x | |
| fc7_conv_biases = NDA("fc7_conv_biases",4096) # SOURCE out_chan | |
| fc7 = NDA("fc7",num_img,4096,1,1) # num,chan,y,x | |
| fc6_one_row_per_patch_buf = NDA("fc6_one_row_per_patch_buf",1,4096) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=fc6[i,:,:,:], dest=fc6_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| fc7 = fc6_one_row_per_patch_buf * transpose(reshape(fc7_conv_filts,4096,4096)) # sgemm: MxNxK == 1x4096x4096 | |
| ReLU(name="relu7",in_place=[fc7]) | |
| Dropout(name="drop7",in_place=[fc7]) | |
| fc8_conv_filts = NDA("fc8_conv_filts",1000,4096,1,1) # SOURCE out_chan,in_chan,y,x | |
| fc8_conv_biases = NDA("fc8_conv_biases",1000) # SOURCE out_chan | |
| fc8 = NDA("fc8",num_img,1000,1,1) # SINK num,chan,y,x | |
| fc7_one_row_per_patch_buf = NDA("fc7_one_row_per_patch_buf",1,4096) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=fc7[i,:,:,:], dest=fc7_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| fc8 = fc7_one_row_per_patch_buf * transpose(reshape(fc8_conv_filts,1000,4096)) # sgemm: MxNxK == 1x4096x1000 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| data = NDA("data",num_img,3,227,227) # SOURCE num,chan,y,x | |
| conv1_filts = NDA("conv1_filts",96,3,11,11) # SOURCE out_chan,in_chan,y,x | |
| conv1_biases = NDA("conv1_biases",96) # SOURCE out_chan | |
| conv1 = NDA("conv1",num_img,96,55,55) # num,chan,y,x | |
| Convolution(name="conv1",bots=[ data ],tops=[ conv1 ],filts=conv1_filts,biases=conv1_biases, | |
| in_pad="0 0 0 0",stride="4 4") | |
| ReLU(name="relu1",in_place=[conv1]) | |
| norm1 = NDA("norm1",num_img,96,55,55) # num,chan,y,x | |
| LRN(name="norm1",bots=[ conv1 ],tops=[ norm1 ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| pool1 = NDA("pool1",num_img,96,27,27) # num,chan,y,x | |
| Pooling(name="pool1",bots=[ norm1 ],tops=[ pool1 ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| conv2_filts = NDA("conv2_filts",256,96,5,5) # SOURCE out_chan,in_chan,y,x | |
| conv2_biases = NDA("conv2_biases",256) # SOURCE out_chan | |
| conv2 = NDA("conv2",num_img,256,27,27) # num,chan,y,x | |
| Convolution(name="conv2",bots=[ pool1 ],tops=[ conv2 ],filts=conv2_filts,biases=conv2_biases, | |
| in_pad="2 2 2 2",stride="1 1") | |
| ReLU(name="relu2",in_place=[conv2]) | |
| norm2 = NDA("norm2",num_img,256,27,27) # num,chan,y,x | |
| LRN(name="norm2",bots=[ conv2 ],tops=[ norm2 ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| pool2 = NDA("pool2",num_img,256,13,13) # num,chan,y,x | |
| Pooling(name="pool2",bots=[ norm2 ],tops=[ pool2 ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| conv3_filts = NDA("conv3_filts",384,256,3,3) # SOURCE out_chan,in_chan,y,x | |
| conv3_biases = NDA("conv3_biases",384) # SOURCE out_chan | |
| conv3 = NDA("conv3",num_img,384,13,13) # num,chan,y,x | |
| Convolution(name="conv3",bots=[ pool2 ],tops=[ conv3 ],filts=conv3_filts,biases=conv3_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu3",in_place=[conv3]) | |
| conv4_filts = NDA("conv4_filts",384,384,3,3) # SOURCE out_chan,in_chan,y,x | |
| conv4_biases = NDA("conv4_biases",384) # SOURCE out_chan | |
| conv4 = NDA("conv4",num_img,384,13,13) # num,chan,y,x | |
| Convolution(name="conv4",bots=[ conv3 ],tops=[ conv4 ],filts=conv4_filts,biases=conv4_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu4",in_place=[conv4]) | |
| conv5_filts = NDA("conv5_filts",256,384,3,3) # SOURCE out_chan,in_chan,y,x | |
| conv5_biases = NDA("conv5_biases",256) # SOURCE out_chan | |
| conv5 = NDA("conv5",num_img,256,13,13) # num,chan,y,x | |
| Convolution(name="conv5",bots=[ conv4 ],tops=[ conv5 ],filts=conv5_filts,biases=conv5_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu5",in_place=[conv5]) | |
| pool5 = NDA("pool5",num_img,256,6,6) # num,chan,y,x | |
| Pooling(name="pool5",bots=[ conv5 ],tops=[ pool5 ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| fc6_conv_filts = NDA("fc6_conv_filts",4096,256,6,6) # SOURCE out_chan,in_chan,y,x | |
| fc6_conv_biases = NDA("fc6_conv_biases",4096) # SOURCE out_chan | |
| fc6 = NDA("fc6",num_img,4096,1,1) # num,chan,y,x | |
| Convolution(name="fc6_conv",bots=[ pool5 ],tops=[ fc6 ],filts=fc6_conv_filts,biases=fc6_conv_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu6",in_place=[fc6]) | |
| Dropout(name="drop6",in_place=[fc6]) | |
| fc7_conv_filts = NDA("fc7_conv_filts",4096,4096,1,1) # SOURCE out_chan,in_chan,y,x | |
| fc7_conv_biases = NDA("fc7_conv_biases",4096) # SOURCE out_chan | |
| fc7 = NDA("fc7",num_img,4096,1,1) # num,chan,y,x | |
| Convolution(name="fc7_conv",bots=[ fc6 ],tops=[ fc7 ],filts=fc7_conv_filts,biases=fc7_conv_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu7",in_place=[fc7]) | |
| Dropout(name="drop7",in_place=[fc7]) | |
| fc8_conv_filts = NDA("fc8_conv_filts",1000,4096,1,1) # SOURCE out_chan,in_chan,y,x | |
| fc8_conv_biases = NDA("fc8_conv_biases",1000) # SOURCE out_chan | |
| fc8 = NDA("fc8",num_img,1000,1,1) # SINK num,chan,y,x | |
| Convolution(name="fc8_conv",bots=[ fc7 ],tops=[ fc8 ],filts=fc8_conv_filts,biases=fc8_conv_biases, | |
| in_pad="0 0 0 0",stride="1 1") |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| name: "AlexNet" | |
| layer { | |
| name: "data" | |
| type: "Data" | |
| top: "data" | |
| top: "label" | |
| include { | |
| phase: TRAIN | |
| } | |
| transform_param { | |
| mirror: true | |
| crop_size: 227 | |
| mean_file: "/scratch/datasets/data/ilsvrc12/imagenet_mean.binaryproto" | |
| } | |
| data_param { | |
| source: "/scratch/datasets/imagenet_classification/ilsvrc12_train_lmdb" | |
| batch_size: 256 | |
| backend: LMDB | |
| } | |
| } | |
| layer { | |
| name: "data" | |
| type: "Data" | |
| top: "data" | |
| top: "label" | |
| include { | |
| phase: TEST | |
| } | |
| transform_param { | |
| mirror: false | |
| crop_size: 227 | |
| mean_file: "/scratch/datasets/data/ilsvrc12/imagenet_mean.binaryproto" | |
| } | |
| data_param { | |
| source: "/scratch/datasets/imagenet_classification/ilsvrc12_val_lmdb" | |
| batch_size: 50 | |
| backend: LMDB | |
| } | |
| } | |
| layer { | |
| name: "conv1" | |
| type: "Convolution" | |
| bottom: "data" | |
| top: "conv1" | |
| param { | |
| lr_mult: 1 | |
| decay_mult: 1 | |
| } | |
| param { | |
| lr_mult: 2 | |
| decay_mult: 0 | |
| } | |
| convolution_param { | |
| num_output: 96 | |
| kernel_size: 11 | |
| stride: 4 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.01 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layer { | |
| name: "relu1" | |
| type: "ReLU" | |
| bottom: "conv1" | |
| top: "conv1" | |
| } | |
| layer { | |
| name: "norm1" | |
| type: "LRN" | |
| bottom: "conv1" | |
| top: "norm1" | |
| lrn_param { | |
| local_size: 5 | |
| alpha: 0.0001 | |
| beta: 0.75 | |
| } | |
| } | |
| layer { | |
| name: "pool1" | |
| type: "Pooling" | |
| bottom: "norm1" | |
| top: "pool1" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 2 | |
| } | |
| } | |
| layer { | |
| name: "conv2" | |
| type: "Convolution" | |
| bottom: "pool1" | |
| top: "conv2" | |
| param { | |
| lr_mult: 1 | |
| decay_mult: 1 | |
| } | |
| param { | |
| lr_mult: 2 | |
| decay_mult: 0 | |
| } | |
| convolution_param { | |
| num_output: 256 | |
| pad: 2 | |
| kernel_size: 5 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.01 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0.1 | |
| } | |
| } | |
| } | |
| layer { | |
| name: "relu2" | |
| type: "ReLU" | |
| bottom: "conv2" | |
| top: "conv2" | |
| } | |
| layer { | |
| name: "norm2" | |
| type: "LRN" | |
| bottom: "conv2" | |
| top: "norm2" | |
| lrn_param { | |
| local_size: 5 | |
| alpha: 0.0001 | |
| beta: 0.75 | |
| } | |
| } | |
| layer { | |
| name: "pool2" | |
| type: "Pooling" | |
| bottom: "norm2" | |
| top: "pool2" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 2 | |
| } | |
| } | |
| layer { | |
| name: "conv3" | |
| type: "Convolution" | |
| bottom: "pool2" | |
| top: "conv3" | |
| param { | |
| lr_mult: 1 | |
| decay_mult: 1 | |
| } | |
| param { | |
| lr_mult: 2 | |
| decay_mult: 0 | |
| } | |
| convolution_param { | |
| num_output: 384 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.01 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layer { | |
| name: "relu3" | |
| type: "ReLU" | |
| bottom: "conv3" | |
| top: "conv3" | |
| } | |
| layer { | |
| name: "conv4" | |
| type: "Convolution" | |
| bottom: "conv3" | |
| top: "conv4" | |
| param { | |
| lr_mult: 1 | |
| decay_mult: 1 | |
| } | |
| param { | |
| lr_mult: 2 | |
| decay_mult: 0 | |
| } | |
| convolution_param { | |
| num_output: 384 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.01 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0.1 | |
| } | |
| } | |
| } | |
| layer { | |
| name: "relu4" | |
| type: "ReLU" | |
| bottom: "conv4" | |
| top: "conv4" | |
| } | |
| layer { | |
| name: "conv5" | |
| type: "Convolution" | |
| bottom: "conv4" | |
| top: "conv5" | |
| param { | |
| lr_mult: 1 | |
| decay_mult: 1 | |
| } | |
| param { | |
| lr_mult: 2 | |
| decay_mult: 0 | |
| } | |
| convolution_param { | |
| num_output: 256 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.01 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0.1 | |
| } | |
| } | |
| } | |
| layer { | |
| name: "relu5" | |
| type: "ReLU" | |
| bottom: "conv5" | |
| top: "conv5" | |
| } | |
| layer { | |
| name: "pool5" | |
| type: "Pooling" | |
| bottom: "conv5" | |
| top: "pool5" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 2 | |
| } | |
| } | |
| layer { | |
| name: "fc6-conv" | |
| type: "Convolution" | |
| bottom: "pool5" | |
| top: "fc6" | |
| param { | |
| lr_mult: 1 | |
| decay_mult: 1 | |
| } | |
| param { | |
| lr_mult: 2 | |
| decay_mult: 0 | |
| } | |
| convolution_param { | |
| num_output: 4096 | |
| kernel_size: 6 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.005 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0.1 | |
| } | |
| } | |
| } | |
| layer { | |
| name: "relu6" | |
| type: "ReLU" | |
| bottom: "fc6" | |
| top: "fc6" | |
| } | |
| layer { | |
| name: "drop6" | |
| type: "Dropout" | |
| bottom: "fc6" | |
| top: "fc6" | |
| dropout_param { | |
| dropout_ratio: 0.5 | |
| } | |
| } | |
| layer { | |
| name: "fc7-conv" | |
| type: "Convolution" | |
| bottom: "fc6" | |
| top: "fc7" | |
| param { | |
| lr_mult: 1 | |
| decay_mult: 1 | |
| } | |
| param { | |
| lr_mult: 2 | |
| decay_mult: 0 | |
| } | |
| convolution_param { | |
| num_output: 4096 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.005 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0.1 | |
| } | |
| } | |
| } | |
| layer { | |
| name: "relu7" | |
| type: "ReLU" | |
| bottom: "fc7" | |
| top: "fc7" | |
| } | |
| layer { | |
| name: "drop7" | |
| type: "Dropout" | |
| bottom: "fc7" | |
| top: "fc7" | |
| dropout_param { | |
| dropout_ratio: 0.5 | |
| } | |
| } | |
| layer { | |
| name: "fc8-conv" | |
| type: "Convolution" | |
| bottom: "fc7" | |
| top: "fc8" | |
| param { | |
| lr_mult: 1 | |
| decay_mult: 1 | |
| } | |
| param { | |
| lr_mult: 2 | |
| decay_mult: 0 | |
| } | |
| convolution_param { | |
| num_output: 1000 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.01 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layer { | |
| name: "accuracy" | |
| type: "Accuracy" | |
| bottom: "fc8" | |
| bottom: "label" | |
| top: "accuracy" | |
| include { | |
| phase: TEST | |
| } | |
| } | |
| layer { | |
| name: "loss" | |
| type: "SoftmaxWithLoss" | |
| bottom: "fc8" | |
| bottom: "label" | |
| top: "loss" | |
| } |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| data = NDA("data",num_img,3,224,224) # SOURCE num,chan,y,x | |
| conv1_filts = NDA("conv1_filts",64,3,7,7) # SOURCE out_chan,in_chan,y,x | |
| conv1_biases = NDA("conv1_biases",64) # SOURCE out_chan | |
| conv1 = NDA("conv1",num_img,64,112,112) # num,chan,y,x | |
| Convolution(name="conv1",bots=[ data ],tops=[ conv1 ],filts=conv1_filts,biases=conv1_biases, | |
| in_pad="3 3 3 3",stride="2 2") | |
| ReLU(name="relu1",in_place=[conv1]) | |
| pool1 = NDA("pool1",num_img,64,56,56) # num,chan,y,x | |
| Pooling(name="pool1",bots=[ conv1 ],tops=[ pool1 ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| norm1 = NDA("norm1",num_img,64,56,56) # num,chan,y,x | |
| LRN(name="norm1",bots=[ pool1 ],tops=[ norm1 ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| reduction2_filts = NDA("reduction2_filts",64,64,1,1) # SOURCE out_chan,in_chan,y,x | |
| reduction2_biases = NDA("reduction2_biases",64) # SOURCE out_chan | |
| reduction2 = NDA("reduction2",num_img,64,56,56) # num,chan,y,x | |
| Convolution(name="reduction2",bots=[ norm1 ],tops=[ reduction2 ],filts=reduction2_filts,biases=reduction2_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_reduction2",in_place=[reduction2]) | |
| conv2_filts = NDA("conv2_filts",192,64,3,3) # SOURCE out_chan,in_chan,y,x | |
| conv2_biases = NDA("conv2_biases",192) # SOURCE out_chan | |
| conv2 = NDA("conv2",num_img,192,56,56) # num,chan,y,x | |
| Convolution(name="conv2",bots=[ reduction2 ],tops=[ conv2 ],filts=conv2_filts,biases=conv2_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu2",in_place=[conv2]) | |
| norm2 = NDA("norm2",num_img,192,56,56) # num,chan,y,x | |
| LRN(name="norm2",bots=[ conv2 ],tops=[ norm2 ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| pool2 = NDA("pool2",num_img,192,28,28) # num,chan,y,x | |
| Pooling(name="pool2",bots=[ norm2 ],tops=[ pool2 ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| icp1_reduction1_filts = NDA("icp1_reduction1_filts",96,192,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp1_reduction1_biases = NDA("icp1_reduction1_biases",96) # SOURCE out_chan | |
| icp1_reduction1 = NDA("icp1_reduction1",num_img,96,28,28) # num,chan,y,x | |
| Convolution(name="icp1_reduction1",bots=[ pool2 ],tops=[ icp1_reduction1 ],filts=icp1_reduction1_filts,biases=icp1_reduction1_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp1_reduction1",in_place=[icp1_reduction1]) | |
| icp1_out1_filts = NDA("icp1_out1_filts",128,96,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp1_out1_biases = NDA("icp1_out1_biases",128) # SOURCE out_chan | |
| icp1_out1 = NDA("icp1_out1",num_img,128,28,28) # num,chan,y,x | |
| Convolution(name="icp1_out1",bots=[ icp1_reduction1 ],tops=[ icp1_out1 ],filts=icp1_out1_filts,biases=icp1_out1_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu_icp1_out1",in_place=[icp1_out1]) | |
| icp1_reduction2_filts = NDA("icp1_reduction2_filts",16,192,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp1_reduction2_biases = NDA("icp1_reduction2_biases",16) # SOURCE out_chan | |
| icp1_reduction2 = NDA("icp1_reduction2",num_img,16,28,28) # num,chan,y,x | |
| Convolution(name="icp1_reduction2",bots=[ pool2 ],tops=[ icp1_reduction2 ],filts=icp1_reduction2_filts,biases=icp1_reduction2_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp1_reduction2",in_place=[icp1_reduction2]) | |
| icp1_out2_filts = NDA("icp1_out2_filts",32,16,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp1_out2_biases = NDA("icp1_out2_biases",32) # SOURCE out_chan | |
| icp1_out2 = NDA("icp1_out2",num_img,32,28,28) # num,chan,y,x | |
| Convolution(name="icp1_out2",bots=[ icp1_reduction2 ],tops=[ icp1_out2 ],filts=icp1_out2_filts,biases=icp1_out2_biases, | |
| in_pad="2 2 2 2",stride="1 1") | |
| ReLU(name="relu_icp1_out2",in_place=[icp1_out2]) | |
| icp1_pool = NDA("icp1_pool",num_img,192,28,28) # num,chan,y,x | |
| Pooling(name="icp1_pool",bots=[ pool2 ],tops=[ icp1_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp1_out3_filts = NDA("icp1_out3_filts",32,192,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp1_out3_biases = NDA("icp1_out3_biases",32) # SOURCE out_chan | |
| icp1_out3 = NDA("icp1_out3",num_img,32,28,28) # num,chan,y,x | |
| Convolution(name="icp1_out3",bots=[ icp1_pool ],tops=[ icp1_out3 ],filts=icp1_out3_filts,biases=icp1_out3_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp1_out3",in_place=[icp1_out3]) | |
| icp1_out0_filts = NDA("icp1_out0_filts",64,192,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp1_out0_biases = NDA("icp1_out0_biases",64) # SOURCE out_chan | |
| icp1_out0 = NDA("icp1_out0",num_img,64,28,28) # num,chan,y,x | |
| Convolution(name="icp1_out0",bots=[ pool2 ],tops=[ icp1_out0 ],filts=icp1_out0_filts,biases=icp1_out0_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp1_out0",in_place=[icp1_out0]) | |
| icp2_in = NDA("icp2_in",num_img,256,28,28) # num,chan,y,x | |
| Concat(name="icp2_in",bots=[ icp1_out0, icp1_out1, icp1_out2, icp1_out3 ],tops=[ icp2_in ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp2_reduction1_filts = NDA("icp2_reduction1_filts",128,256,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp2_reduction1_biases = NDA("icp2_reduction1_biases",128) # SOURCE out_chan | |
| icp2_reduction1 = NDA("icp2_reduction1",num_img,128,28,28) # num,chan,y,x | |
| Convolution(name="icp2_reduction1",bots=[ icp2_in ],tops=[ icp2_reduction1 ],filts=icp2_reduction1_filts,biases=icp2_reduction1_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp2_reduction1",in_place=[icp2_reduction1]) | |
| icp2_out1_filts = NDA("icp2_out1_filts",192,128,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp2_out1_biases = NDA("icp2_out1_biases",192) # SOURCE out_chan | |
| icp2_out1 = NDA("icp2_out1",num_img,192,28,28) # num,chan,y,x | |
| Convolution(name="icp2_out1",bots=[ icp2_reduction1 ],tops=[ icp2_out1 ],filts=icp2_out1_filts,biases=icp2_out1_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu_icp2_out1",in_place=[icp2_out1]) | |
| icp2_reduction2_filts = NDA("icp2_reduction2_filts",32,256,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp2_reduction2_biases = NDA("icp2_reduction2_biases",32) # SOURCE out_chan | |
| icp2_reduction2 = NDA("icp2_reduction2",num_img,32,28,28) # num,chan,y,x | |
| Convolution(name="icp2_reduction2",bots=[ icp2_in ],tops=[ icp2_reduction2 ],filts=icp2_reduction2_filts,biases=icp2_reduction2_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp2_reduction2",in_place=[icp2_reduction2]) | |
| icp2_out2_filts = NDA("icp2_out2_filts",96,32,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp2_out2_biases = NDA("icp2_out2_biases",96) # SOURCE out_chan | |
| icp2_out2 = NDA("icp2_out2",num_img,96,28,28) # num,chan,y,x | |
| Convolution(name="icp2_out2",bots=[ icp2_reduction2 ],tops=[ icp2_out2 ],filts=icp2_out2_filts,biases=icp2_out2_biases, | |
| in_pad="2 2 2 2",stride="1 1") | |
| ReLU(name="relu_icp2_out2",in_place=[icp2_out2]) | |
| icp2_pool = NDA("icp2_pool",num_img,256,28,28) # num,chan,y,x | |
| Pooling(name="icp2_pool",bots=[ icp2_in ],tops=[ icp2_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp2_out3_filts = NDA("icp2_out3_filts",64,256,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp2_out3_biases = NDA("icp2_out3_biases",64) # SOURCE out_chan | |
| icp2_out3 = NDA("icp2_out3",num_img,64,28,28) # num,chan,y,x | |
| Convolution(name="icp2_out3",bots=[ icp2_pool ],tops=[ icp2_out3 ],filts=icp2_out3_filts,biases=icp2_out3_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp2_out3",in_place=[icp2_out3]) | |
| icp2_out0_filts = NDA("icp2_out0_filts",128,256,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp2_out0_biases = NDA("icp2_out0_biases",128) # SOURCE out_chan | |
| icp2_out0 = NDA("icp2_out0",num_img,128,28,28) # num,chan,y,x | |
| Convolution(name="icp2_out0",bots=[ icp2_in ],tops=[ icp2_out0 ],filts=icp2_out0_filts,biases=icp2_out0_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp2_out0",in_place=[icp2_out0]) | |
| icp2_out = NDA("icp2_out",num_img,480,28,28) # num,chan,y,x | |
| Concat(name="icp2_out",bots=[ icp2_out0, icp2_out1, icp2_out2, icp2_out3 ],tops=[ icp2_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp3_in = NDA("icp3_in",num_img,480,14,14) # num,chan,y,x | |
| Pooling(name="icp3_in",bots=[ icp2_out ],tops=[ icp3_in ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| icp3_reduction1_filts = NDA("icp3_reduction1_filts",96,480,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp3_reduction1_biases = NDA("icp3_reduction1_biases",96) # SOURCE out_chan | |
| icp3_reduction1 = NDA("icp3_reduction1",num_img,96,14,14) # num,chan,y,x | |
| Convolution(name="icp3_reduction1",bots=[ icp3_in ],tops=[ icp3_reduction1 ],filts=icp3_reduction1_filts,biases=icp3_reduction1_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp3_reduction1",in_place=[icp3_reduction1]) | |
| icp3_out1_filts = NDA("icp3_out1_filts",208,96,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp3_out1_biases = NDA("icp3_out1_biases",208) # SOURCE out_chan | |
| icp3_out1 = NDA("icp3_out1",num_img,208,14,14) # num,chan,y,x | |
| Convolution(name="icp3_out1",bots=[ icp3_reduction1 ],tops=[ icp3_out1 ],filts=icp3_out1_filts,biases=icp3_out1_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu_icp3_out1",in_place=[icp3_out1]) | |
| icp3_reduction2_filts = NDA("icp3_reduction2_filts",16,480,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp3_reduction2_biases = NDA("icp3_reduction2_biases",16) # SOURCE out_chan | |
| icp3_reduction2 = NDA("icp3_reduction2",num_img,16,14,14) # num,chan,y,x | |
| Convolution(name="icp3_reduction2",bots=[ icp3_in ],tops=[ icp3_reduction2 ],filts=icp3_reduction2_filts,biases=icp3_reduction2_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp3_reduction2",in_place=[icp3_reduction2]) | |
| icp3_out2_filts = NDA("icp3_out2_filts",48,16,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp3_out2_biases = NDA("icp3_out2_biases",48) # SOURCE out_chan | |
| icp3_out2 = NDA("icp3_out2",num_img,48,14,14) # num,chan,y,x | |
| Convolution(name="icp3_out2",bots=[ icp3_reduction2 ],tops=[ icp3_out2 ],filts=icp3_out2_filts,biases=icp3_out2_biases, | |
| in_pad="2 2 2 2",stride="1 1") | |
| ReLU(name="relu_icp3_out2",in_place=[icp3_out2]) | |
| icp3_pool = NDA("icp3_pool",num_img,480,14,14) # num,chan,y,x | |
| Pooling(name="icp3_pool",bots=[ icp3_in ],tops=[ icp3_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp3_out3_filts = NDA("icp3_out3_filts",64,480,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp3_out3_biases = NDA("icp3_out3_biases",64) # SOURCE out_chan | |
| icp3_out3 = NDA("icp3_out3",num_img,64,14,14) # num,chan,y,x | |
| Convolution(name="icp3_out3",bots=[ icp3_pool ],tops=[ icp3_out3 ],filts=icp3_out3_filts,biases=icp3_out3_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp3_out3",in_place=[icp3_out3]) | |
| icp3_out0_filts = NDA("icp3_out0_filts",192,480,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp3_out0_biases = NDA("icp3_out0_biases",192) # SOURCE out_chan | |
| icp3_out0 = NDA("icp3_out0",num_img,192,14,14) # num,chan,y,x | |
| Convolution(name="icp3_out0",bots=[ icp3_in ],tops=[ icp3_out0 ],filts=icp3_out0_filts,biases=icp3_out0_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp3_out0",in_place=[icp3_out0]) | |
| icp3_out = NDA("icp3_out",num_img,512,14,14) # num,chan,y,x | |
| Concat(name="icp3_out",bots=[ icp3_out0, icp3_out1, icp3_out2, icp3_out3 ],tops=[ icp3_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| cls1_pool = NDA("cls1_pool",num_img,512,4,4) # num,chan,y,x | |
| Pooling(name="cls1_pool",bots=[ icp3_out ],tops=[ cls1_pool ], | |
| in_pad="0 0 0 0",stride="3 3") | |
| cls1_reduction_filts = NDA("cls1_reduction_filts",128,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| cls1_reduction_biases = NDA("cls1_reduction_biases",128) # SOURCE out_chan | |
| cls1_reduction = NDA("cls1_reduction",num_img,128,4,4) # num,chan,y,x | |
| Convolution(name="cls1_reduction",bots=[ cls1_pool ],tops=[ cls1_reduction ],filts=cls1_reduction_filts,biases=cls1_reduction_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_cls1_reduction",in_place=[cls1_reduction]) | |
| cls1_fc1_filts = NDA("cls1_fc1_filts",1024,128,4,4) # SOURCE out_chan,in_chan,y,x | |
| cls1_fc1_biases = NDA("cls1_fc1_biases",1024) # SOURCE out_chan | |
| cls1_fc1 = NDA("cls1_fc1",num_img,1024,1,1) # num,chan,y,x | |
| InnerProduct(name="cls1_fc1",bots=[ cls1_reduction ],tops=[ cls1_fc1 ],filts=cls1_fc1_filts,biases=cls1_fc1_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_cls1_fc1",in_place=[cls1_fc1]) | |
| Dropout(name="cls1_drop",in_place=[cls1_fc1]) | |
| cls1_fc2__filts = NDA("cls1_fc2__filts",33,1024,1,1) # SOURCE out_chan,in_chan,y,x | |
| cls1_fc2__biases = NDA("cls1_fc2__biases",33) # SOURCE out_chan | |
| cls1_fc2_ = NDA("cls1_fc2_",num_img,33,1,1) # SINK num,chan,y,x | |
| InnerProduct(name="cls1_fc2_",bots=[ cls1_fc1 ],tops=[ cls1_fc2_ ],filts=cls1_fc2__filts,biases=cls1_fc2__biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp4_reduction1_filts = NDA("icp4_reduction1_filts",112,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp4_reduction1_biases = NDA("icp4_reduction1_biases",112) # SOURCE out_chan | |
| icp4_reduction1 = NDA("icp4_reduction1",num_img,112,14,14) # num,chan,y,x | |
| Convolution(name="icp4_reduction1",bots=[ icp3_out ],tops=[ icp4_reduction1 ],filts=icp4_reduction1_filts,biases=icp4_reduction1_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp4_reduction1",in_place=[icp4_reduction1]) | |
| icp4_out1_filts = NDA("icp4_out1_filts",224,112,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp4_out1_biases = NDA("icp4_out1_biases",224) # SOURCE out_chan | |
| icp4_out1 = NDA("icp4_out1",num_img,224,14,14) # num,chan,y,x | |
| Convolution(name="icp4_out1",bots=[ icp4_reduction1 ],tops=[ icp4_out1 ],filts=icp4_out1_filts,biases=icp4_out1_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu_icp4_out1",in_place=[icp4_out1]) | |
| icp4_reduction2_filts = NDA("icp4_reduction2_filts",24,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp4_reduction2_biases = NDA("icp4_reduction2_biases",24) # SOURCE out_chan | |
| icp4_reduction2 = NDA("icp4_reduction2",num_img,24,14,14) # num,chan,y,x | |
| Convolution(name="icp4_reduction2",bots=[ icp3_out ],tops=[ icp4_reduction2 ],filts=icp4_reduction2_filts,biases=icp4_reduction2_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp4_reduction2",in_place=[icp4_reduction2]) | |
| icp4_out2_filts = NDA("icp4_out2_filts",64,24,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp4_out2_biases = NDA("icp4_out2_biases",64) # SOURCE out_chan | |
| icp4_out2 = NDA("icp4_out2",num_img,64,14,14) # num,chan,y,x | |
| Convolution(name="icp4_out2",bots=[ icp4_reduction2 ],tops=[ icp4_out2 ],filts=icp4_out2_filts,biases=icp4_out2_biases, | |
| in_pad="2 2 2 2",stride="1 1") | |
| ReLU(name="relu_icp4_out2",in_place=[icp4_out2]) | |
| icp4_pool = NDA("icp4_pool",num_img,512,14,14) # num,chan,y,x | |
| Pooling(name="icp4_pool",bots=[ icp3_out ],tops=[ icp4_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp4_out3_filts = NDA("icp4_out3_filts",64,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp4_out3_biases = NDA("icp4_out3_biases",64) # SOURCE out_chan | |
| icp4_out3 = NDA("icp4_out3",num_img,64,14,14) # num,chan,y,x | |
| Convolution(name="icp4_out3",bots=[ icp4_pool ],tops=[ icp4_out3 ],filts=icp4_out3_filts,biases=icp4_out3_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp4_out3",in_place=[icp4_out3]) | |
| icp4_out0_filts = NDA("icp4_out0_filts",160,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp4_out0_biases = NDA("icp4_out0_biases",160) # SOURCE out_chan | |
| icp4_out0 = NDA("icp4_out0",num_img,160,14,14) # num,chan,y,x | |
| Convolution(name="icp4_out0",bots=[ icp3_out ],tops=[ icp4_out0 ],filts=icp4_out0_filts,biases=icp4_out0_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp4_out0",in_place=[icp4_out0]) | |
| icp4_out = NDA("icp4_out",num_img,512,14,14) # num,chan,y,x | |
| Concat(name="icp4_out",bots=[ icp4_out0, icp4_out1, icp4_out2, icp4_out3 ],tops=[ icp4_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp5_reduction1_filts = NDA("icp5_reduction1_filts",128,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp5_reduction1_biases = NDA("icp5_reduction1_biases",128) # SOURCE out_chan | |
| icp5_reduction1 = NDA("icp5_reduction1",num_img,128,14,14) # num,chan,y,x | |
| Convolution(name="icp5_reduction1",bots=[ icp4_out ],tops=[ icp5_reduction1 ],filts=icp5_reduction1_filts,biases=icp5_reduction1_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp5_reduction1",in_place=[icp5_reduction1]) | |
| icp5_out1_filts = NDA("icp5_out1_filts",256,128,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp5_out1_biases = NDA("icp5_out1_biases",256) # SOURCE out_chan | |
| icp5_out1 = NDA("icp5_out1",num_img,256,14,14) # num,chan,y,x | |
| Convolution(name="icp5_out1",bots=[ icp5_reduction1 ],tops=[ icp5_out1 ],filts=icp5_out1_filts,biases=icp5_out1_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu_icp5_out1",in_place=[icp5_out1]) | |
| icp5_reduction2_filts = NDA("icp5_reduction2_filts",24,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp5_reduction2_biases = NDA("icp5_reduction2_biases",24) # SOURCE out_chan | |
| icp5_reduction2 = NDA("icp5_reduction2",num_img,24,14,14) # num,chan,y,x | |
| Convolution(name="icp5_reduction2",bots=[ icp4_out ],tops=[ icp5_reduction2 ],filts=icp5_reduction2_filts,biases=icp5_reduction2_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp5_reduction2",in_place=[icp5_reduction2]) | |
| icp5_out2_filts = NDA("icp5_out2_filts",64,24,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp5_out2_biases = NDA("icp5_out2_biases",64) # SOURCE out_chan | |
| icp5_out2 = NDA("icp5_out2",num_img,64,14,14) # num,chan,y,x | |
| Convolution(name="icp5_out2",bots=[ icp5_reduction2 ],tops=[ icp5_out2 ],filts=icp5_out2_filts,biases=icp5_out2_biases, | |
| in_pad="2 2 2 2",stride="1 1") | |
| ReLU(name="relu_icp5_out2",in_place=[icp5_out2]) | |
| icp5_pool = NDA("icp5_pool",num_img,512,14,14) # num,chan,y,x | |
| Pooling(name="icp5_pool",bots=[ icp4_out ],tops=[ icp5_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp5_out3_filts = NDA("icp5_out3_filts",64,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp5_out3_biases = NDA("icp5_out3_biases",64) # SOURCE out_chan | |
| icp5_out3 = NDA("icp5_out3",num_img,64,14,14) # num,chan,y,x | |
| Convolution(name="icp5_out3",bots=[ icp5_pool ],tops=[ icp5_out3 ],filts=icp5_out3_filts,biases=icp5_out3_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp5_out3",in_place=[icp5_out3]) | |
| icp5_out0_filts = NDA("icp5_out0_filts",128,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp5_out0_biases = NDA("icp5_out0_biases",128) # SOURCE out_chan | |
| icp5_out0 = NDA("icp5_out0",num_img,128,14,14) # num,chan,y,x | |
| Convolution(name="icp5_out0",bots=[ icp4_out ],tops=[ icp5_out0 ],filts=icp5_out0_filts,biases=icp5_out0_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp5_out0",in_place=[icp5_out0]) | |
| icp5_out = NDA("icp5_out",num_img,512,14,14) # num,chan,y,x | |
| Concat(name="icp5_out",bots=[ icp5_out0, icp5_out1, icp5_out2, icp5_out3 ],tops=[ icp5_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp6_reduction1_filts = NDA("icp6_reduction1_filts",144,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp6_reduction1_biases = NDA("icp6_reduction1_biases",144) # SOURCE out_chan | |
| icp6_reduction1 = NDA("icp6_reduction1",num_img,144,14,14) # num,chan,y,x | |
| Convolution(name="icp6_reduction1",bots=[ icp5_out ],tops=[ icp6_reduction1 ],filts=icp6_reduction1_filts,biases=icp6_reduction1_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp6_reduction1",in_place=[icp6_reduction1]) | |
| icp6_out1_filts = NDA("icp6_out1_filts",288,144,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp6_out1_biases = NDA("icp6_out1_biases",288) # SOURCE out_chan | |
| icp6_out1 = NDA("icp6_out1",num_img,288,14,14) # num,chan,y,x | |
| Convolution(name="icp6_out1",bots=[ icp6_reduction1 ],tops=[ icp6_out1 ],filts=icp6_out1_filts,biases=icp6_out1_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu_icp6_out1",in_place=[icp6_out1]) | |
| icp6_reduction2_filts = NDA("icp6_reduction2_filts",32,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp6_reduction2_biases = NDA("icp6_reduction2_biases",32) # SOURCE out_chan | |
| icp6_reduction2 = NDA("icp6_reduction2",num_img,32,14,14) # num,chan,y,x | |
| Convolution(name="icp6_reduction2",bots=[ icp5_out ],tops=[ icp6_reduction2 ],filts=icp6_reduction2_filts,biases=icp6_reduction2_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp6_reduction2",in_place=[icp6_reduction2]) | |
| icp6_out2_filts = NDA("icp6_out2_filts",64,32,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp6_out2_biases = NDA("icp6_out2_biases",64) # SOURCE out_chan | |
| icp6_out2 = NDA("icp6_out2",num_img,64,14,14) # num,chan,y,x | |
| Convolution(name="icp6_out2",bots=[ icp6_reduction2 ],tops=[ icp6_out2 ],filts=icp6_out2_filts,biases=icp6_out2_biases, | |
| in_pad="2 2 2 2",stride="1 1") | |
| ReLU(name="relu_icp6_out2",in_place=[icp6_out2]) | |
| icp6_pool = NDA("icp6_pool",num_img,512,14,14) # num,chan,y,x | |
| Pooling(name="icp6_pool",bots=[ icp5_out ],tops=[ icp6_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp6_out3_filts = NDA("icp6_out3_filts",64,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp6_out3_biases = NDA("icp6_out3_biases",64) # SOURCE out_chan | |
| icp6_out3 = NDA("icp6_out3",num_img,64,14,14) # num,chan,y,x | |
| Convolution(name="icp6_out3",bots=[ icp6_pool ],tops=[ icp6_out3 ],filts=icp6_out3_filts,biases=icp6_out3_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp6_out3",in_place=[icp6_out3]) | |
| icp6_out0_filts = NDA("icp6_out0_filts",112,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp6_out0_biases = NDA("icp6_out0_biases",112) # SOURCE out_chan | |
| icp6_out0 = NDA("icp6_out0",num_img,112,14,14) # num,chan,y,x | |
| Convolution(name="icp6_out0",bots=[ icp5_out ],tops=[ icp6_out0 ],filts=icp6_out0_filts,biases=icp6_out0_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp6_out0",in_place=[icp6_out0]) | |
| icp6_out = NDA("icp6_out",num_img,528,14,14) # num,chan,y,x | |
| Concat(name="icp6_out",bots=[ icp6_out0, icp6_out1, icp6_out2, icp6_out3 ],tops=[ icp6_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| cls2_pool = NDA("cls2_pool",num_img,528,4,4) # num,chan,y,x | |
| Pooling(name="cls2_pool",bots=[ icp6_out ],tops=[ cls2_pool ], | |
| in_pad="0 0 0 0",stride="3 3") | |
| cls2_reduction_filts = NDA("cls2_reduction_filts",128,528,1,1) # SOURCE out_chan,in_chan,y,x | |
| cls2_reduction_biases = NDA("cls2_reduction_biases",128) # SOURCE out_chan | |
| cls2_reduction = NDA("cls2_reduction",num_img,128,4,4) # num,chan,y,x | |
| Convolution(name="cls2_reduction",bots=[ cls2_pool ],tops=[ cls2_reduction ],filts=cls2_reduction_filts,biases=cls2_reduction_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_cls2_reduction",in_place=[cls2_reduction]) | |
| cls2_fc1_filts = NDA("cls2_fc1_filts",1024,128,4,4) # SOURCE out_chan,in_chan,y,x | |
| cls2_fc1_biases = NDA("cls2_fc1_biases",1024) # SOURCE out_chan | |
| cls2_fc1 = NDA("cls2_fc1",num_img,1024,1,1) # num,chan,y,x | |
| InnerProduct(name="cls2_fc1",bots=[ cls2_reduction ],tops=[ cls2_fc1 ],filts=cls2_fc1_filts,biases=cls2_fc1_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_cls2_fc1",in_place=[cls2_fc1]) | |
| Dropout(name="cls2_drop",in_place=[cls2_fc1]) | |
| cls2_fc2__filts = NDA("cls2_fc2__filts",33,1024,1,1) # SOURCE out_chan,in_chan,y,x | |
| cls2_fc2__biases = NDA("cls2_fc2__biases",33) # SOURCE out_chan | |
| cls2_fc2_ = NDA("cls2_fc2_",num_img,33,1,1) # SINK num,chan,y,x | |
| InnerProduct(name="cls2_fc2_",bots=[ cls2_fc1 ],tops=[ cls2_fc2_ ],filts=cls2_fc2__filts,biases=cls2_fc2__biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp7_reduction1_filts = NDA("icp7_reduction1_filts",160,528,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp7_reduction1_biases = NDA("icp7_reduction1_biases",160) # SOURCE out_chan | |
| icp7_reduction1 = NDA("icp7_reduction1",num_img,160,14,14) # num,chan,y,x | |
| Convolution(name="icp7_reduction1",bots=[ icp6_out ],tops=[ icp7_reduction1 ],filts=icp7_reduction1_filts,biases=icp7_reduction1_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp7_reduction1",in_place=[icp7_reduction1]) | |
| icp7_out1_filts = NDA("icp7_out1_filts",320,160,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp7_out1_biases = NDA("icp7_out1_biases",320) # SOURCE out_chan | |
| icp7_out1 = NDA("icp7_out1",num_img,320,14,14) # num,chan,y,x | |
| Convolution(name="icp7_out1",bots=[ icp7_reduction1 ],tops=[ icp7_out1 ],filts=icp7_out1_filts,biases=icp7_out1_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu_icp7_out1",in_place=[icp7_out1]) | |
| icp7_reduction2_filts = NDA("icp7_reduction2_filts",32,528,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp7_reduction2_biases = NDA("icp7_reduction2_biases",32) # SOURCE out_chan | |
| icp7_reduction2 = NDA("icp7_reduction2",num_img,32,14,14) # num,chan,y,x | |
| Convolution(name="icp7_reduction2",bots=[ icp6_out ],tops=[ icp7_reduction2 ],filts=icp7_reduction2_filts,biases=icp7_reduction2_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp7_reduction2",in_place=[icp7_reduction2]) | |
| icp7_out2_filts = NDA("icp7_out2_filts",128,32,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp7_out2_biases = NDA("icp7_out2_biases",128) # SOURCE out_chan | |
| icp7_out2 = NDA("icp7_out2",num_img,128,14,14) # num,chan,y,x | |
| Convolution(name="icp7_out2",bots=[ icp7_reduction2 ],tops=[ icp7_out2 ],filts=icp7_out2_filts,biases=icp7_out2_biases, | |
| in_pad="2 2 2 2",stride="1 1") | |
| ReLU(name="relu_icp7_out2",in_place=[icp7_out2]) | |
| icp7_pool = NDA("icp7_pool",num_img,528,14,14) # num,chan,y,x | |
| Pooling(name="icp7_pool",bots=[ icp6_out ],tops=[ icp7_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp7_out3_filts = NDA("icp7_out3_filts",128,528,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp7_out3_biases = NDA("icp7_out3_biases",128) # SOURCE out_chan | |
| icp7_out3 = NDA("icp7_out3",num_img,128,14,14) # num,chan,y,x | |
| Convolution(name="icp7_out3",bots=[ icp7_pool ],tops=[ icp7_out3 ],filts=icp7_out3_filts,biases=icp7_out3_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp7_out3",in_place=[icp7_out3]) | |
| icp7_out0_filts = NDA("icp7_out0_filts",256,528,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp7_out0_biases = NDA("icp7_out0_biases",256) # SOURCE out_chan | |
| icp7_out0 = NDA("icp7_out0",num_img,256,14,14) # num,chan,y,x | |
| Convolution(name="icp7_out0",bots=[ icp6_out ],tops=[ icp7_out0 ],filts=icp7_out0_filts,biases=icp7_out0_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp7_out0",in_place=[icp7_out0]) | |
| icp7_out = NDA("icp7_out",num_img,832,14,14) # num,chan,y,x | |
| Concat(name="icp7_out",bots=[ icp7_out0, icp7_out1, icp7_out2, icp7_out3 ],tops=[ icp7_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp8_in = NDA("icp8_in",num_img,832,7,7) # num,chan,y,x | |
| Pooling(name="icp8_in",bots=[ icp7_out ],tops=[ icp8_in ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| icp8_reduction1_filts = NDA("icp8_reduction1_filts",160,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp8_reduction1_biases = NDA("icp8_reduction1_biases",160) # SOURCE out_chan | |
| icp8_reduction1 = NDA("icp8_reduction1",num_img,160,7,7) # num,chan,y,x | |
| Convolution(name="icp8_reduction1",bots=[ icp8_in ],tops=[ icp8_reduction1 ],filts=icp8_reduction1_filts,biases=icp8_reduction1_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp8_reduction1",in_place=[icp8_reduction1]) | |
| icp8_out1_filts = NDA("icp8_out1_filts",320,160,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp8_out1_biases = NDA("icp8_out1_biases",320) # SOURCE out_chan | |
| icp8_out1 = NDA("icp8_out1",num_img,320,7,7) # num,chan,y,x | |
| Convolution(name="icp8_out1",bots=[ icp8_reduction1 ],tops=[ icp8_out1 ],filts=icp8_out1_filts,biases=icp8_out1_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu_icp8_out1",in_place=[icp8_out1]) | |
| icp8_reduction2_filts = NDA("icp8_reduction2_filts",32,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp8_reduction2_biases = NDA("icp8_reduction2_biases",32) # SOURCE out_chan | |
| icp8_reduction2 = NDA("icp8_reduction2",num_img,32,7,7) # num,chan,y,x | |
| Convolution(name="icp8_reduction2",bots=[ icp8_in ],tops=[ icp8_reduction2 ],filts=icp8_reduction2_filts,biases=icp8_reduction2_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp8_reduction2",in_place=[icp8_reduction2]) | |
| icp8_out2_filts = NDA("icp8_out2_filts",128,32,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp8_out2_biases = NDA("icp8_out2_biases",128) # SOURCE out_chan | |
| icp8_out2 = NDA("icp8_out2",num_img,128,7,7) # num,chan,y,x | |
| Convolution(name="icp8_out2",bots=[ icp8_reduction2 ],tops=[ icp8_out2 ],filts=icp8_out2_filts,biases=icp8_out2_biases, | |
| in_pad="2 2 2 2",stride="1 1") | |
| ReLU(name="relu_icp8_out2",in_place=[icp8_out2]) | |
| icp8_pool = NDA("icp8_pool",num_img,832,7,7) # num,chan,y,x | |
| Pooling(name="icp8_pool",bots=[ icp8_in ],tops=[ icp8_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp8_out3_filts = NDA("icp8_out3_filts",128,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp8_out3_biases = NDA("icp8_out3_biases",128) # SOURCE out_chan | |
| icp8_out3 = NDA("icp8_out3",num_img,128,7,7) # num,chan,y,x | |
| Convolution(name="icp8_out3",bots=[ icp8_pool ],tops=[ icp8_out3 ],filts=icp8_out3_filts,biases=icp8_out3_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp8_out3",in_place=[icp8_out3]) | |
| icp8_out0_filts = NDA("icp8_out0_filts",256,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp8_out0_biases = NDA("icp8_out0_biases",256) # SOURCE out_chan | |
| icp8_out0 = NDA("icp8_out0",num_img,256,7,7) # num,chan,y,x | |
| Convolution(name="icp8_out0",bots=[ icp8_in ],tops=[ icp8_out0 ],filts=icp8_out0_filts,biases=icp8_out0_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp8_out0",in_place=[icp8_out0]) | |
| icp8_out = NDA("icp8_out",num_img,832,7,7) # num,chan,y,x | |
| Concat(name="icp8_out",bots=[ icp8_out0, icp8_out1, icp8_out2, icp8_out3 ],tops=[ icp8_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp9_reduction1_filts = NDA("icp9_reduction1_filts",192,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp9_reduction1_biases = NDA("icp9_reduction1_biases",192) # SOURCE out_chan | |
| icp9_reduction1 = NDA("icp9_reduction1",num_img,192,7,7) # num,chan,y,x | |
| Convolution(name="icp9_reduction1",bots=[ icp8_out ],tops=[ icp9_reduction1 ],filts=icp9_reduction1_filts,biases=icp9_reduction1_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp9_reduction1",in_place=[icp9_reduction1]) | |
| icp9_out1_filts = NDA("icp9_out1_filts",384,192,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp9_out1_biases = NDA("icp9_out1_biases",384) # SOURCE out_chan | |
| icp9_out1 = NDA("icp9_out1",num_img,384,7,7) # num,chan,y,x | |
| Convolution(name="icp9_out1",bots=[ icp9_reduction1 ],tops=[ icp9_out1 ],filts=icp9_out1_filts,biases=icp9_out1_biases, | |
| in_pad="1 1 1 1",stride="1 1") | |
| ReLU(name="relu_icp9_out1",in_place=[icp9_out1]) | |
| icp9_reduction2_filts = NDA("icp9_reduction2_filts",48,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp9_reduction2_biases = NDA("icp9_reduction2_biases",48) # SOURCE out_chan | |
| icp9_reduction2 = NDA("icp9_reduction2",num_img,48,7,7) # num,chan,y,x | |
| Convolution(name="icp9_reduction2",bots=[ icp8_out ],tops=[ icp9_reduction2 ],filts=icp9_reduction2_filts,biases=icp9_reduction2_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp9_reduction2",in_place=[icp9_reduction2]) | |
| icp9_out2_filts = NDA("icp9_out2_filts",128,48,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp9_out2_biases = NDA("icp9_out2_biases",128) # SOURCE out_chan | |
| icp9_out2 = NDA("icp9_out2",num_img,128,7,7) # num,chan,y,x | |
| Convolution(name="icp9_out2",bots=[ icp9_reduction2 ],tops=[ icp9_out2 ],filts=icp9_out2_filts,biases=icp9_out2_biases, | |
| in_pad="2 2 2 2",stride="1 1") | |
| ReLU(name="relu_icp9_out2",in_place=[icp9_out2]) | |
| icp9_pool = NDA("icp9_pool",num_img,832,7,7) # num,chan,y,x | |
| Pooling(name="icp9_pool",bots=[ icp8_out ],tops=[ icp9_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp9_out3_filts = NDA("icp9_out3_filts",128,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp9_out3_biases = NDA("icp9_out3_biases",128) # SOURCE out_chan | |
| icp9_out3 = NDA("icp9_out3",num_img,128,7,7) # num,chan,y,x | |
| Convolution(name="icp9_out3",bots=[ icp9_pool ],tops=[ icp9_out3 ],filts=icp9_out3_filts,biases=icp9_out3_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp9_out3",in_place=[icp9_out3]) | |
| icp9_out0_filts = NDA("icp9_out0_filts",384,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp9_out0_biases = NDA("icp9_out0_biases",384) # SOURCE out_chan | |
| icp9_out0 = NDA("icp9_out0",num_img,384,7,7) # num,chan,y,x | |
| Convolution(name="icp9_out0",bots=[ icp8_out ],tops=[ icp9_out0 ],filts=icp9_out0_filts,biases=icp9_out0_biases, | |
| in_pad="0 0 0 0",stride="1 1") | |
| ReLU(name="relu_icp9_out0",in_place=[icp9_out0]) | |
| icp9_out = NDA("icp9_out",num_img,1024,7,7) # num,chan,y,x | |
| Concat(name="icp9_out",bots=[ icp9_out0, icp9_out1, icp9_out2, icp9_out3 ],tops=[ icp9_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| cls3_pool = NDA("cls3_pool",num_img,1024,1,1) # num,chan,y,x | |
| Pooling(name="cls3_pool",bots=[ icp9_out ],tops=[ cls3_pool ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| Dropout(name="cls3_drop",in_place=[cls3_pool]) | |
| cls3_fc__filts = NDA("cls3_fc__filts",33,1024,1,1) # SOURCE out_chan,in_chan,y,x | |
| cls3_fc__biases = NDA("cls3_fc__biases",33) # SOURCE out_chan | |
| cls3_fc_ = NDA("cls3_fc_",num_img,33,1,1) # SINK num,chan,y,x | |
| InnerProduct(name="cls3_fc_",bots=[ cls3_pool ],tops=[ cls3_fc_ ],filts=cls3_fc__filts,biases=cls3_fc__biases, | |
| in_pad="0 0 0 0",stride="1 1") |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| data = NDA("data",num_img,3,224,224) # SOURCE num,chan,y,x | |
| conv1_filts = NDA("conv1_filts",64,3,7,7) # SOURCE out_chan,in_chan,y,x | |
| conv1_biases = NDA("conv1_biases",64) # SOURCE out_chan | |
| conv1 = NDA("conv1",num_img,64,112,112) # num,chan,y,x | |
| data_one_row_per_patch_buf = NDA("data_one_row_per_patch_buf",12544,147) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=data[i,:,:,:], dest=data_one_row_per_patch_buf, in_pad="3 3 3 3",stride="2 2" ) # one copy per output elem | |
| conv1 = data_one_row_per_patch_buf * transpose(reshape(conv1_filts,64,147)) # sgemm: MxNxK == 12544x147x64 | |
| ReLU(name="relu1",in_place=[conv1]) | |
| pool1 = NDA("pool1",num_img,64,56,56) # num,chan,y,x | |
| Pooling(name="pool1",bots=[ conv1 ],tops=[ pool1 ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| norm1 = NDA("norm1",num_img,64,56,56) # num,chan,y,x | |
| LRN(name="norm1",bots=[ pool1 ],tops=[ norm1 ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| reduction2_filts = NDA("reduction2_filts",64,64,1,1) # SOURCE out_chan,in_chan,y,x | |
| reduction2_biases = NDA("reduction2_biases",64) # SOURCE out_chan | |
| reduction2 = NDA("reduction2",num_img,64,56,56) # num,chan,y,x | |
| norm1_one_row_per_patch_buf = NDA("norm1_one_row_per_patch_buf",3136,64) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=norm1[i,:,:,:], dest=norm1_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| reduction2 = norm1_one_row_per_patch_buf * transpose(reshape(reduction2_filts,64,64)) # sgemm: MxNxK == 3136x64x64 | |
| ReLU(name="relu_reduction2",in_place=[reduction2]) | |
| conv2_filts = NDA("conv2_filts",192,64,3,3) # SOURCE out_chan,in_chan,y,x | |
| conv2_biases = NDA("conv2_biases",192) # SOURCE out_chan | |
| conv2 = NDA("conv2",num_img,192,56,56) # num,chan,y,x | |
| reduction2_one_row_per_patch_buf = NDA("reduction2_one_row_per_patch_buf",3136,576) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=reduction2[i,:,:,:], dest=reduction2_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| conv2 = reduction2_one_row_per_patch_buf * transpose(reshape(conv2_filts,192,576)) # sgemm: MxNxK == 3136x576x192 | |
| ReLU(name="relu2",in_place=[conv2]) | |
| norm2 = NDA("norm2",num_img,192,56,56) # num,chan,y,x | |
| LRN(name="norm2",bots=[ conv2 ],tops=[ norm2 ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| pool2 = NDA("pool2",num_img,192,28,28) # num,chan,y,x | |
| Pooling(name="pool2",bots=[ norm2 ],tops=[ pool2 ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| icp1_reduction1_filts = NDA("icp1_reduction1_filts",96,192,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp1_reduction1_biases = NDA("icp1_reduction1_biases",96) # SOURCE out_chan | |
| icp1_reduction1 = NDA("icp1_reduction1",num_img,96,28,28) # num,chan,y,x | |
| pool2_one_row_per_patch_buf = NDA("pool2_one_row_per_patch_buf",784,192) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=pool2[i,:,:,:], dest=pool2_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp1_reduction1 = pool2_one_row_per_patch_buf * transpose(reshape(icp1_reduction1_filts,96,192)) # sgemm: MxNxK == 784x192x96 | |
| ReLU(name="relu_icp1_reduction1",in_place=[icp1_reduction1]) | |
| icp1_out1_filts = NDA("icp1_out1_filts",128,96,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp1_out1_biases = NDA("icp1_out1_biases",128) # SOURCE out_chan | |
| icp1_out1 = NDA("icp1_out1",num_img,128,28,28) # num,chan,y,x | |
| icp1_reduction1_one_row_per_patch_buf = NDA("icp1_reduction1_one_row_per_patch_buf",784,864) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp1_reduction1[i,:,:,:], dest=icp1_reduction1_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| icp1_out1 = icp1_reduction1_one_row_per_patch_buf * transpose(reshape(icp1_out1_filts,128,864)) # sgemm: MxNxK == 784x864x128 | |
| ReLU(name="relu_icp1_out1",in_place=[icp1_out1]) | |
| icp1_reduction2_filts = NDA("icp1_reduction2_filts",16,192,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp1_reduction2_biases = NDA("icp1_reduction2_biases",16) # SOURCE out_chan | |
| icp1_reduction2 = NDA("icp1_reduction2",num_img,16,28,28) # num,chan,y,x | |
| pool2_one_row_per_patch_buf = NDA("pool2_one_row_per_patch_buf",784,192) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=pool2[i,:,:,:], dest=pool2_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp1_reduction2 = pool2_one_row_per_patch_buf * transpose(reshape(icp1_reduction2_filts,16,192)) # sgemm: MxNxK == 784x192x16 | |
| ReLU(name="relu_icp1_reduction2",in_place=[icp1_reduction2]) | |
| icp1_out2_filts = NDA("icp1_out2_filts",32,16,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp1_out2_biases = NDA("icp1_out2_biases",32) # SOURCE out_chan | |
| icp1_out2 = NDA("icp1_out2",num_img,32,28,28) # num,chan,y,x | |
| icp1_reduction2_one_row_per_patch_buf = NDA("icp1_reduction2_one_row_per_patch_buf",784,400) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp1_reduction2[i,:,:,:], dest=icp1_reduction2_one_row_per_patch_buf, in_pad="2 2 2 2",stride="1 1" ) # one copy per output elem | |
| icp1_out2 = icp1_reduction2_one_row_per_patch_buf * transpose(reshape(icp1_out2_filts,32,400)) # sgemm: MxNxK == 784x400x32 | |
| ReLU(name="relu_icp1_out2",in_place=[icp1_out2]) | |
| icp1_pool = NDA("icp1_pool",num_img,192,28,28) # num,chan,y,x | |
| Pooling(name="icp1_pool",bots=[ pool2 ],tops=[ icp1_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp1_out3_filts = NDA("icp1_out3_filts",32,192,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp1_out3_biases = NDA("icp1_out3_biases",32) # SOURCE out_chan | |
| icp1_out3 = NDA("icp1_out3",num_img,32,28,28) # num,chan,y,x | |
| icp1_pool_one_row_per_patch_buf = NDA("icp1_pool_one_row_per_patch_buf",784,192) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp1_pool[i,:,:,:], dest=icp1_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp1_out3 = icp1_pool_one_row_per_patch_buf * transpose(reshape(icp1_out3_filts,32,192)) # sgemm: MxNxK == 784x192x32 | |
| ReLU(name="relu_icp1_out3",in_place=[icp1_out3]) | |
| icp1_out0_filts = NDA("icp1_out0_filts",64,192,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp1_out0_biases = NDA("icp1_out0_biases",64) # SOURCE out_chan | |
| icp1_out0 = NDA("icp1_out0",num_img,64,28,28) # num,chan,y,x | |
| pool2_one_row_per_patch_buf = NDA("pool2_one_row_per_patch_buf",784,192) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=pool2[i,:,:,:], dest=pool2_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp1_out0 = pool2_one_row_per_patch_buf * transpose(reshape(icp1_out0_filts,64,192)) # sgemm: MxNxK == 784x192x64 | |
| ReLU(name="relu_icp1_out0",in_place=[icp1_out0]) | |
| icp2_in = NDA("icp2_in",num_img,256,28,28) # num,chan,y,x | |
| Concat(name="icp2_in",bots=[ icp1_out0, icp1_out1, icp1_out2, icp1_out3 ],tops=[ icp2_in ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp2_reduction1_filts = NDA("icp2_reduction1_filts",128,256,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp2_reduction1_biases = NDA("icp2_reduction1_biases",128) # SOURCE out_chan | |
| icp2_reduction1 = NDA("icp2_reduction1",num_img,128,28,28) # num,chan,y,x | |
| icp2_in_one_row_per_patch_buf = NDA("icp2_in_one_row_per_patch_buf",784,256) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp2_in[i,:,:,:], dest=icp2_in_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp2_reduction1 = icp2_in_one_row_per_patch_buf * transpose(reshape(icp2_reduction1_filts,128,256)) # sgemm: MxNxK == 784x256x128 | |
| ReLU(name="relu_icp2_reduction1",in_place=[icp2_reduction1]) | |
| icp2_out1_filts = NDA("icp2_out1_filts",192,128,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp2_out1_biases = NDA("icp2_out1_biases",192) # SOURCE out_chan | |
| icp2_out1 = NDA("icp2_out1",num_img,192,28,28) # num,chan,y,x | |
| icp2_reduction1_one_row_per_patch_buf = NDA("icp2_reduction1_one_row_per_patch_buf",784,1152) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp2_reduction1[i,:,:,:], dest=icp2_reduction1_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| icp2_out1 = icp2_reduction1_one_row_per_patch_buf * transpose(reshape(icp2_out1_filts,192,1152)) # sgemm: MxNxK == 784x1152x192 | |
| ReLU(name="relu_icp2_out1",in_place=[icp2_out1]) | |
| icp2_reduction2_filts = NDA("icp2_reduction2_filts",32,256,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp2_reduction2_biases = NDA("icp2_reduction2_biases",32) # SOURCE out_chan | |
| icp2_reduction2 = NDA("icp2_reduction2",num_img,32,28,28) # num,chan,y,x | |
| icp2_in_one_row_per_patch_buf = NDA("icp2_in_one_row_per_patch_buf",784,256) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp2_in[i,:,:,:], dest=icp2_in_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp2_reduction2 = icp2_in_one_row_per_patch_buf * transpose(reshape(icp2_reduction2_filts,32,256)) # sgemm: MxNxK == 784x256x32 | |
| ReLU(name="relu_icp2_reduction2",in_place=[icp2_reduction2]) | |
| icp2_out2_filts = NDA("icp2_out2_filts",96,32,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp2_out2_biases = NDA("icp2_out2_biases",96) # SOURCE out_chan | |
| icp2_out2 = NDA("icp2_out2",num_img,96,28,28) # num,chan,y,x | |
| icp2_reduction2_one_row_per_patch_buf = NDA("icp2_reduction2_one_row_per_patch_buf",784,800) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp2_reduction2[i,:,:,:], dest=icp2_reduction2_one_row_per_patch_buf, in_pad="2 2 2 2",stride="1 1" ) # one copy per output elem | |
| icp2_out2 = icp2_reduction2_one_row_per_patch_buf * transpose(reshape(icp2_out2_filts,96,800)) # sgemm: MxNxK == 784x800x96 | |
| ReLU(name="relu_icp2_out2",in_place=[icp2_out2]) | |
| icp2_pool = NDA("icp2_pool",num_img,256,28,28) # num,chan,y,x | |
| Pooling(name="icp2_pool",bots=[ icp2_in ],tops=[ icp2_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp2_out3_filts = NDA("icp2_out3_filts",64,256,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp2_out3_biases = NDA("icp2_out3_biases",64) # SOURCE out_chan | |
| icp2_out3 = NDA("icp2_out3",num_img,64,28,28) # num,chan,y,x | |
| icp2_pool_one_row_per_patch_buf = NDA("icp2_pool_one_row_per_patch_buf",784,256) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp2_pool[i,:,:,:], dest=icp2_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp2_out3 = icp2_pool_one_row_per_patch_buf * transpose(reshape(icp2_out3_filts,64,256)) # sgemm: MxNxK == 784x256x64 | |
| ReLU(name="relu_icp2_out3",in_place=[icp2_out3]) | |
| icp2_out0_filts = NDA("icp2_out0_filts",128,256,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp2_out0_biases = NDA("icp2_out0_biases",128) # SOURCE out_chan | |
| icp2_out0 = NDA("icp2_out0",num_img,128,28,28) # num,chan,y,x | |
| icp2_in_one_row_per_patch_buf = NDA("icp2_in_one_row_per_patch_buf",784,256) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp2_in[i,:,:,:], dest=icp2_in_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp2_out0 = icp2_in_one_row_per_patch_buf * transpose(reshape(icp2_out0_filts,128,256)) # sgemm: MxNxK == 784x256x128 | |
| ReLU(name="relu_icp2_out0",in_place=[icp2_out0]) | |
| icp2_out = NDA("icp2_out",num_img,480,28,28) # num,chan,y,x | |
| Concat(name="icp2_out",bots=[ icp2_out0, icp2_out1, icp2_out2, icp2_out3 ],tops=[ icp2_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp3_in = NDA("icp3_in",num_img,480,14,14) # num,chan,y,x | |
| Pooling(name="icp3_in",bots=[ icp2_out ],tops=[ icp3_in ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| icp3_reduction1_filts = NDA("icp3_reduction1_filts",96,480,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp3_reduction1_biases = NDA("icp3_reduction1_biases",96) # SOURCE out_chan | |
| icp3_reduction1 = NDA("icp3_reduction1",num_img,96,14,14) # num,chan,y,x | |
| icp3_in_one_row_per_patch_buf = NDA("icp3_in_one_row_per_patch_buf",196,480) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp3_in[i,:,:,:], dest=icp3_in_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp3_reduction1 = icp3_in_one_row_per_patch_buf * transpose(reshape(icp3_reduction1_filts,96,480)) # sgemm: MxNxK == 196x480x96 | |
| ReLU(name="relu_icp3_reduction1",in_place=[icp3_reduction1]) | |
| icp3_out1_filts = NDA("icp3_out1_filts",208,96,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp3_out1_biases = NDA("icp3_out1_biases",208) # SOURCE out_chan | |
| icp3_out1 = NDA("icp3_out1",num_img,208,14,14) # num,chan,y,x | |
| icp3_reduction1_one_row_per_patch_buf = NDA("icp3_reduction1_one_row_per_patch_buf",196,864) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp3_reduction1[i,:,:,:], dest=icp3_reduction1_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| icp3_out1 = icp3_reduction1_one_row_per_patch_buf * transpose(reshape(icp3_out1_filts,208,864)) # sgemm: MxNxK == 196x864x208 | |
| ReLU(name="relu_icp3_out1",in_place=[icp3_out1]) | |
| icp3_reduction2_filts = NDA("icp3_reduction2_filts",16,480,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp3_reduction2_biases = NDA("icp3_reduction2_biases",16) # SOURCE out_chan | |
| icp3_reduction2 = NDA("icp3_reduction2",num_img,16,14,14) # num,chan,y,x | |
| icp3_in_one_row_per_patch_buf = NDA("icp3_in_one_row_per_patch_buf",196,480) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp3_in[i,:,:,:], dest=icp3_in_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp3_reduction2 = icp3_in_one_row_per_patch_buf * transpose(reshape(icp3_reduction2_filts,16,480)) # sgemm: MxNxK == 196x480x16 | |
| ReLU(name="relu_icp3_reduction2",in_place=[icp3_reduction2]) | |
| icp3_out2_filts = NDA("icp3_out2_filts",48,16,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp3_out2_biases = NDA("icp3_out2_biases",48) # SOURCE out_chan | |
| icp3_out2 = NDA("icp3_out2",num_img,48,14,14) # num,chan,y,x | |
| icp3_reduction2_one_row_per_patch_buf = NDA("icp3_reduction2_one_row_per_patch_buf",196,400) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp3_reduction2[i,:,:,:], dest=icp3_reduction2_one_row_per_patch_buf, in_pad="2 2 2 2",stride="1 1" ) # one copy per output elem | |
| icp3_out2 = icp3_reduction2_one_row_per_patch_buf * transpose(reshape(icp3_out2_filts,48,400)) # sgemm: MxNxK == 196x400x48 | |
| ReLU(name="relu_icp3_out2",in_place=[icp3_out2]) | |
| icp3_pool = NDA("icp3_pool",num_img,480,14,14) # num,chan,y,x | |
| Pooling(name="icp3_pool",bots=[ icp3_in ],tops=[ icp3_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp3_out3_filts = NDA("icp3_out3_filts",64,480,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp3_out3_biases = NDA("icp3_out3_biases",64) # SOURCE out_chan | |
| icp3_out3 = NDA("icp3_out3",num_img,64,14,14) # num,chan,y,x | |
| icp3_pool_one_row_per_patch_buf = NDA("icp3_pool_one_row_per_patch_buf",196,480) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp3_pool[i,:,:,:], dest=icp3_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp3_out3 = icp3_pool_one_row_per_patch_buf * transpose(reshape(icp3_out3_filts,64,480)) # sgemm: MxNxK == 196x480x64 | |
| ReLU(name="relu_icp3_out3",in_place=[icp3_out3]) | |
| icp3_out0_filts = NDA("icp3_out0_filts",192,480,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp3_out0_biases = NDA("icp3_out0_biases",192) # SOURCE out_chan | |
| icp3_out0 = NDA("icp3_out0",num_img,192,14,14) # num,chan,y,x | |
| icp3_in_one_row_per_patch_buf = NDA("icp3_in_one_row_per_patch_buf",196,480) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp3_in[i,:,:,:], dest=icp3_in_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp3_out0 = icp3_in_one_row_per_patch_buf * transpose(reshape(icp3_out0_filts,192,480)) # sgemm: MxNxK == 196x480x192 | |
| ReLU(name="relu_icp3_out0",in_place=[icp3_out0]) | |
| icp3_out = NDA("icp3_out",num_img,512,14,14) # num,chan,y,x | |
| Concat(name="icp3_out",bots=[ icp3_out0, icp3_out1, icp3_out2, icp3_out3 ],tops=[ icp3_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| cls1_pool = NDA("cls1_pool",num_img,512,4,4) # num,chan,y,x | |
| Pooling(name="cls1_pool",bots=[ icp3_out ],tops=[ cls1_pool ], | |
| in_pad="0 0 0 0",stride="3 3") | |
| cls1_reduction_filts = NDA("cls1_reduction_filts",128,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| cls1_reduction_biases = NDA("cls1_reduction_biases",128) # SOURCE out_chan | |
| cls1_reduction = NDA("cls1_reduction",num_img,128,4,4) # num,chan,y,x | |
| cls1_pool_one_row_per_patch_buf = NDA("cls1_pool_one_row_per_patch_buf",16,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=cls1_pool[i,:,:,:], dest=cls1_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| cls1_reduction = cls1_pool_one_row_per_patch_buf * transpose(reshape(cls1_reduction_filts,128,512)) # sgemm: MxNxK == 16x512x128 | |
| ReLU(name="relu_cls1_reduction",in_place=[cls1_reduction]) | |
| cls1_fc1_filts = NDA("cls1_fc1_filts",1024,128,4,4) # SOURCE out_chan,in_chan,y,x | |
| cls1_fc1_biases = NDA("cls1_fc1_biases",1024) # SOURCE out_chan | |
| cls1_fc1 = NDA("cls1_fc1",num_img,1024,1,1) # num,chan,y,x | |
| cls1_reduction_one_row_per_patch_buf = NDA("cls1_reduction_one_row_per_patch_buf",1,2048) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=cls1_reduction[i,:,:,:], dest=cls1_reduction_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| cls1_fc1 = cls1_reduction_one_row_per_patch_buf * transpose(reshape(cls1_fc1_filts,1024,2048)) # sgemm: MxNxK == 1x2048x1024 | |
| ReLU(name="relu_cls1_fc1",in_place=[cls1_fc1]) | |
| Dropout(name="cls1_drop",in_place=[cls1_fc1]) | |
| cls1_fc2__filts = NDA("cls1_fc2__filts",33,1024,1,1) # SOURCE out_chan,in_chan,y,x | |
| cls1_fc2__biases = NDA("cls1_fc2__biases",33) # SOURCE out_chan | |
| cls1_fc2_ = NDA("cls1_fc2_",num_img,33,1,1) # SINK num,chan,y,x | |
| cls1_fc1_one_row_per_patch_buf = NDA("cls1_fc1_one_row_per_patch_buf",1,1024) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=cls1_fc1[i,:,:,:], dest=cls1_fc1_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| cls1_fc2_ = cls1_fc1_one_row_per_patch_buf * transpose(reshape(cls1_fc2__filts,33,1024)) # sgemm: MxNxK == 1x1024x33 | |
| icp4_reduction1_filts = NDA("icp4_reduction1_filts",112,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp4_reduction1_biases = NDA("icp4_reduction1_biases",112) # SOURCE out_chan | |
| icp4_reduction1 = NDA("icp4_reduction1",num_img,112,14,14) # num,chan,y,x | |
| icp3_out_one_row_per_patch_buf = NDA("icp3_out_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp3_out[i,:,:,:], dest=icp3_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp4_reduction1 = icp3_out_one_row_per_patch_buf * transpose(reshape(icp4_reduction1_filts,112,512)) # sgemm: MxNxK == 196x512x112 | |
| ReLU(name="relu_icp4_reduction1",in_place=[icp4_reduction1]) | |
| icp4_out1_filts = NDA("icp4_out1_filts",224,112,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp4_out1_biases = NDA("icp4_out1_biases",224) # SOURCE out_chan | |
| icp4_out1 = NDA("icp4_out1",num_img,224,14,14) # num,chan,y,x | |
| icp4_reduction1_one_row_per_patch_buf = NDA("icp4_reduction1_one_row_per_patch_buf",196,1008) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp4_reduction1[i,:,:,:], dest=icp4_reduction1_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| icp4_out1 = icp4_reduction1_one_row_per_patch_buf * transpose(reshape(icp4_out1_filts,224,1008)) # sgemm: MxNxK == 196x1008x224 | |
| ReLU(name="relu_icp4_out1",in_place=[icp4_out1]) | |
| icp4_reduction2_filts = NDA("icp4_reduction2_filts",24,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp4_reduction2_biases = NDA("icp4_reduction2_biases",24) # SOURCE out_chan | |
| icp4_reduction2 = NDA("icp4_reduction2",num_img,24,14,14) # num,chan,y,x | |
| icp3_out_one_row_per_patch_buf = NDA("icp3_out_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp3_out[i,:,:,:], dest=icp3_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp4_reduction2 = icp3_out_one_row_per_patch_buf * transpose(reshape(icp4_reduction2_filts,24,512)) # sgemm: MxNxK == 196x512x24 | |
| ReLU(name="relu_icp4_reduction2",in_place=[icp4_reduction2]) | |
| icp4_out2_filts = NDA("icp4_out2_filts",64,24,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp4_out2_biases = NDA("icp4_out2_biases",64) # SOURCE out_chan | |
| icp4_out2 = NDA("icp4_out2",num_img,64,14,14) # num,chan,y,x | |
| icp4_reduction2_one_row_per_patch_buf = NDA("icp4_reduction2_one_row_per_patch_buf",196,600) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp4_reduction2[i,:,:,:], dest=icp4_reduction2_one_row_per_patch_buf, in_pad="2 2 2 2",stride="1 1" ) # one copy per output elem | |
| icp4_out2 = icp4_reduction2_one_row_per_patch_buf * transpose(reshape(icp4_out2_filts,64,600)) # sgemm: MxNxK == 196x600x64 | |
| ReLU(name="relu_icp4_out2",in_place=[icp4_out2]) | |
| icp4_pool = NDA("icp4_pool",num_img,512,14,14) # num,chan,y,x | |
| Pooling(name="icp4_pool",bots=[ icp3_out ],tops=[ icp4_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp4_out3_filts = NDA("icp4_out3_filts",64,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp4_out3_biases = NDA("icp4_out3_biases",64) # SOURCE out_chan | |
| icp4_out3 = NDA("icp4_out3",num_img,64,14,14) # num,chan,y,x | |
| icp4_pool_one_row_per_patch_buf = NDA("icp4_pool_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp4_pool[i,:,:,:], dest=icp4_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp4_out3 = icp4_pool_one_row_per_patch_buf * transpose(reshape(icp4_out3_filts,64,512)) # sgemm: MxNxK == 196x512x64 | |
| ReLU(name="relu_icp4_out3",in_place=[icp4_out3]) | |
| icp4_out0_filts = NDA("icp4_out0_filts",160,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp4_out0_biases = NDA("icp4_out0_biases",160) # SOURCE out_chan | |
| icp4_out0 = NDA("icp4_out0",num_img,160,14,14) # num,chan,y,x | |
| icp3_out_one_row_per_patch_buf = NDA("icp3_out_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp3_out[i,:,:,:], dest=icp3_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp4_out0 = icp3_out_one_row_per_patch_buf * transpose(reshape(icp4_out0_filts,160,512)) # sgemm: MxNxK == 196x512x160 | |
| ReLU(name="relu_icp4_out0",in_place=[icp4_out0]) | |
| icp4_out = NDA("icp4_out",num_img,512,14,14) # num,chan,y,x | |
| Concat(name="icp4_out",bots=[ icp4_out0, icp4_out1, icp4_out2, icp4_out3 ],tops=[ icp4_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp5_reduction1_filts = NDA("icp5_reduction1_filts",128,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp5_reduction1_biases = NDA("icp5_reduction1_biases",128) # SOURCE out_chan | |
| icp5_reduction1 = NDA("icp5_reduction1",num_img,128,14,14) # num,chan,y,x | |
| icp4_out_one_row_per_patch_buf = NDA("icp4_out_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp4_out[i,:,:,:], dest=icp4_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp5_reduction1 = icp4_out_one_row_per_patch_buf * transpose(reshape(icp5_reduction1_filts,128,512)) # sgemm: MxNxK == 196x512x128 | |
| ReLU(name="relu_icp5_reduction1",in_place=[icp5_reduction1]) | |
| icp5_out1_filts = NDA("icp5_out1_filts",256,128,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp5_out1_biases = NDA("icp5_out1_biases",256) # SOURCE out_chan | |
| icp5_out1 = NDA("icp5_out1",num_img,256,14,14) # num,chan,y,x | |
| icp5_reduction1_one_row_per_patch_buf = NDA("icp5_reduction1_one_row_per_patch_buf",196,1152) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp5_reduction1[i,:,:,:], dest=icp5_reduction1_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| icp5_out1 = icp5_reduction1_one_row_per_patch_buf * transpose(reshape(icp5_out1_filts,256,1152)) # sgemm: MxNxK == 196x1152x256 | |
| ReLU(name="relu_icp5_out1",in_place=[icp5_out1]) | |
| icp5_reduction2_filts = NDA("icp5_reduction2_filts",24,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp5_reduction2_biases = NDA("icp5_reduction2_biases",24) # SOURCE out_chan | |
| icp5_reduction2 = NDA("icp5_reduction2",num_img,24,14,14) # num,chan,y,x | |
| icp4_out_one_row_per_patch_buf = NDA("icp4_out_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp4_out[i,:,:,:], dest=icp4_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp5_reduction2 = icp4_out_one_row_per_patch_buf * transpose(reshape(icp5_reduction2_filts,24,512)) # sgemm: MxNxK == 196x512x24 | |
| ReLU(name="relu_icp5_reduction2",in_place=[icp5_reduction2]) | |
| icp5_out2_filts = NDA("icp5_out2_filts",64,24,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp5_out2_biases = NDA("icp5_out2_biases",64) # SOURCE out_chan | |
| icp5_out2 = NDA("icp5_out2",num_img,64,14,14) # num,chan,y,x | |
| icp5_reduction2_one_row_per_patch_buf = NDA("icp5_reduction2_one_row_per_patch_buf",196,600) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp5_reduction2[i,:,:,:], dest=icp5_reduction2_one_row_per_patch_buf, in_pad="2 2 2 2",stride="1 1" ) # one copy per output elem | |
| icp5_out2 = icp5_reduction2_one_row_per_patch_buf * transpose(reshape(icp5_out2_filts,64,600)) # sgemm: MxNxK == 196x600x64 | |
| ReLU(name="relu_icp5_out2",in_place=[icp5_out2]) | |
| icp5_pool = NDA("icp5_pool",num_img,512,14,14) # num,chan,y,x | |
| Pooling(name="icp5_pool",bots=[ icp4_out ],tops=[ icp5_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp5_out3_filts = NDA("icp5_out3_filts",64,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp5_out3_biases = NDA("icp5_out3_biases",64) # SOURCE out_chan | |
| icp5_out3 = NDA("icp5_out3",num_img,64,14,14) # num,chan,y,x | |
| icp5_pool_one_row_per_patch_buf = NDA("icp5_pool_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp5_pool[i,:,:,:], dest=icp5_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp5_out3 = icp5_pool_one_row_per_patch_buf * transpose(reshape(icp5_out3_filts,64,512)) # sgemm: MxNxK == 196x512x64 | |
| ReLU(name="relu_icp5_out3",in_place=[icp5_out3]) | |
| icp5_out0_filts = NDA("icp5_out0_filts",128,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp5_out0_biases = NDA("icp5_out0_biases",128) # SOURCE out_chan | |
| icp5_out0 = NDA("icp5_out0",num_img,128,14,14) # num,chan,y,x | |
| icp4_out_one_row_per_patch_buf = NDA("icp4_out_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp4_out[i,:,:,:], dest=icp4_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp5_out0 = icp4_out_one_row_per_patch_buf * transpose(reshape(icp5_out0_filts,128,512)) # sgemm: MxNxK == 196x512x128 | |
| ReLU(name="relu_icp5_out0",in_place=[icp5_out0]) | |
| icp5_out = NDA("icp5_out",num_img,512,14,14) # num,chan,y,x | |
| Concat(name="icp5_out",bots=[ icp5_out0, icp5_out1, icp5_out2, icp5_out3 ],tops=[ icp5_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp6_reduction1_filts = NDA("icp6_reduction1_filts",144,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp6_reduction1_biases = NDA("icp6_reduction1_biases",144) # SOURCE out_chan | |
| icp6_reduction1 = NDA("icp6_reduction1",num_img,144,14,14) # num,chan,y,x | |
| icp5_out_one_row_per_patch_buf = NDA("icp5_out_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp5_out[i,:,:,:], dest=icp5_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp6_reduction1 = icp5_out_one_row_per_patch_buf * transpose(reshape(icp6_reduction1_filts,144,512)) # sgemm: MxNxK == 196x512x144 | |
| ReLU(name="relu_icp6_reduction1",in_place=[icp6_reduction1]) | |
| icp6_out1_filts = NDA("icp6_out1_filts",288,144,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp6_out1_biases = NDA("icp6_out1_biases",288) # SOURCE out_chan | |
| icp6_out1 = NDA("icp6_out1",num_img,288,14,14) # num,chan,y,x | |
| icp6_reduction1_one_row_per_patch_buf = NDA("icp6_reduction1_one_row_per_patch_buf",196,1296) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp6_reduction1[i,:,:,:], dest=icp6_reduction1_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| icp6_out1 = icp6_reduction1_one_row_per_patch_buf * transpose(reshape(icp6_out1_filts,288,1296)) # sgemm: MxNxK == 196x1296x288 | |
| ReLU(name="relu_icp6_out1",in_place=[icp6_out1]) | |
| icp6_reduction2_filts = NDA("icp6_reduction2_filts",32,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp6_reduction2_biases = NDA("icp6_reduction2_biases",32) # SOURCE out_chan | |
| icp6_reduction2 = NDA("icp6_reduction2",num_img,32,14,14) # num,chan,y,x | |
| icp5_out_one_row_per_patch_buf = NDA("icp5_out_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp5_out[i,:,:,:], dest=icp5_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp6_reduction2 = icp5_out_one_row_per_patch_buf * transpose(reshape(icp6_reduction2_filts,32,512)) # sgemm: MxNxK == 196x512x32 | |
| ReLU(name="relu_icp6_reduction2",in_place=[icp6_reduction2]) | |
| icp6_out2_filts = NDA("icp6_out2_filts",64,32,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp6_out2_biases = NDA("icp6_out2_biases",64) # SOURCE out_chan | |
| icp6_out2 = NDA("icp6_out2",num_img,64,14,14) # num,chan,y,x | |
| icp6_reduction2_one_row_per_patch_buf = NDA("icp6_reduction2_one_row_per_patch_buf",196,800) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp6_reduction2[i,:,:,:], dest=icp6_reduction2_one_row_per_patch_buf, in_pad="2 2 2 2",stride="1 1" ) # one copy per output elem | |
| icp6_out2 = icp6_reduction2_one_row_per_patch_buf * transpose(reshape(icp6_out2_filts,64,800)) # sgemm: MxNxK == 196x800x64 | |
| ReLU(name="relu_icp6_out2",in_place=[icp6_out2]) | |
| icp6_pool = NDA("icp6_pool",num_img,512,14,14) # num,chan,y,x | |
| Pooling(name="icp6_pool",bots=[ icp5_out ],tops=[ icp6_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp6_out3_filts = NDA("icp6_out3_filts",64,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp6_out3_biases = NDA("icp6_out3_biases",64) # SOURCE out_chan | |
| icp6_out3 = NDA("icp6_out3",num_img,64,14,14) # num,chan,y,x | |
| icp6_pool_one_row_per_patch_buf = NDA("icp6_pool_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp6_pool[i,:,:,:], dest=icp6_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp6_out3 = icp6_pool_one_row_per_patch_buf * transpose(reshape(icp6_out3_filts,64,512)) # sgemm: MxNxK == 196x512x64 | |
| ReLU(name="relu_icp6_out3",in_place=[icp6_out3]) | |
| icp6_out0_filts = NDA("icp6_out0_filts",112,512,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp6_out0_biases = NDA("icp6_out0_biases",112) # SOURCE out_chan | |
| icp6_out0 = NDA("icp6_out0",num_img,112,14,14) # num,chan,y,x | |
| icp5_out_one_row_per_patch_buf = NDA("icp5_out_one_row_per_patch_buf",196,512) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp5_out[i,:,:,:], dest=icp5_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp6_out0 = icp5_out_one_row_per_patch_buf * transpose(reshape(icp6_out0_filts,112,512)) # sgemm: MxNxK == 196x512x112 | |
| ReLU(name="relu_icp6_out0",in_place=[icp6_out0]) | |
| icp6_out = NDA("icp6_out",num_img,528,14,14) # num,chan,y,x | |
| Concat(name="icp6_out",bots=[ icp6_out0, icp6_out1, icp6_out2, icp6_out3 ],tops=[ icp6_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| cls2_pool = NDA("cls2_pool",num_img,528,4,4) # num,chan,y,x | |
| Pooling(name="cls2_pool",bots=[ icp6_out ],tops=[ cls2_pool ], | |
| in_pad="0 0 0 0",stride="3 3") | |
| cls2_reduction_filts = NDA("cls2_reduction_filts",128,528,1,1) # SOURCE out_chan,in_chan,y,x | |
| cls2_reduction_biases = NDA("cls2_reduction_biases",128) # SOURCE out_chan | |
| cls2_reduction = NDA("cls2_reduction",num_img,128,4,4) # num,chan,y,x | |
| cls2_pool_one_row_per_patch_buf = NDA("cls2_pool_one_row_per_patch_buf",16,528) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=cls2_pool[i,:,:,:], dest=cls2_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| cls2_reduction = cls2_pool_one_row_per_patch_buf * transpose(reshape(cls2_reduction_filts,128,528)) # sgemm: MxNxK == 16x528x128 | |
| ReLU(name="relu_cls2_reduction",in_place=[cls2_reduction]) | |
| cls2_fc1_filts = NDA("cls2_fc1_filts",1024,128,4,4) # SOURCE out_chan,in_chan,y,x | |
| cls2_fc1_biases = NDA("cls2_fc1_biases",1024) # SOURCE out_chan | |
| cls2_fc1 = NDA("cls2_fc1",num_img,1024,1,1) # num,chan,y,x | |
| cls2_reduction_one_row_per_patch_buf = NDA("cls2_reduction_one_row_per_patch_buf",1,2048) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=cls2_reduction[i,:,:,:], dest=cls2_reduction_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| cls2_fc1 = cls2_reduction_one_row_per_patch_buf * transpose(reshape(cls2_fc1_filts,1024,2048)) # sgemm: MxNxK == 1x2048x1024 | |
| ReLU(name="relu_cls2_fc1",in_place=[cls2_fc1]) | |
| Dropout(name="cls2_drop",in_place=[cls2_fc1]) | |
| cls2_fc2__filts = NDA("cls2_fc2__filts",33,1024,1,1) # SOURCE out_chan,in_chan,y,x | |
| cls2_fc2__biases = NDA("cls2_fc2__biases",33) # SOURCE out_chan | |
| cls2_fc2_ = NDA("cls2_fc2_",num_img,33,1,1) # SINK num,chan,y,x | |
| cls2_fc1_one_row_per_patch_buf = NDA("cls2_fc1_one_row_per_patch_buf",1,1024) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=cls2_fc1[i,:,:,:], dest=cls2_fc1_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| cls2_fc2_ = cls2_fc1_one_row_per_patch_buf * transpose(reshape(cls2_fc2__filts,33,1024)) # sgemm: MxNxK == 1x1024x33 | |
| icp7_reduction1_filts = NDA("icp7_reduction1_filts",160,528,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp7_reduction1_biases = NDA("icp7_reduction1_biases",160) # SOURCE out_chan | |
| icp7_reduction1 = NDA("icp7_reduction1",num_img,160,14,14) # num,chan,y,x | |
| icp6_out_one_row_per_patch_buf = NDA("icp6_out_one_row_per_patch_buf",196,528) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp6_out[i,:,:,:], dest=icp6_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp7_reduction1 = icp6_out_one_row_per_patch_buf * transpose(reshape(icp7_reduction1_filts,160,528)) # sgemm: MxNxK == 196x528x160 | |
| ReLU(name="relu_icp7_reduction1",in_place=[icp7_reduction1]) | |
| icp7_out1_filts = NDA("icp7_out1_filts",320,160,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp7_out1_biases = NDA("icp7_out1_biases",320) # SOURCE out_chan | |
| icp7_out1 = NDA("icp7_out1",num_img,320,14,14) # num,chan,y,x | |
| icp7_reduction1_one_row_per_patch_buf = NDA("icp7_reduction1_one_row_per_patch_buf",196,1440) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp7_reduction1[i,:,:,:], dest=icp7_reduction1_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| icp7_out1 = icp7_reduction1_one_row_per_patch_buf * transpose(reshape(icp7_out1_filts,320,1440)) # sgemm: MxNxK == 196x1440x320 | |
| ReLU(name="relu_icp7_out1",in_place=[icp7_out1]) | |
| icp7_reduction2_filts = NDA("icp7_reduction2_filts",32,528,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp7_reduction2_biases = NDA("icp7_reduction2_biases",32) # SOURCE out_chan | |
| icp7_reduction2 = NDA("icp7_reduction2",num_img,32,14,14) # num,chan,y,x | |
| icp6_out_one_row_per_patch_buf = NDA("icp6_out_one_row_per_patch_buf",196,528) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp6_out[i,:,:,:], dest=icp6_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp7_reduction2 = icp6_out_one_row_per_patch_buf * transpose(reshape(icp7_reduction2_filts,32,528)) # sgemm: MxNxK == 196x528x32 | |
| ReLU(name="relu_icp7_reduction2",in_place=[icp7_reduction2]) | |
| icp7_out2_filts = NDA("icp7_out2_filts",128,32,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp7_out2_biases = NDA("icp7_out2_biases",128) # SOURCE out_chan | |
| icp7_out2 = NDA("icp7_out2",num_img,128,14,14) # num,chan,y,x | |
| icp7_reduction2_one_row_per_patch_buf = NDA("icp7_reduction2_one_row_per_patch_buf",196,800) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp7_reduction2[i,:,:,:], dest=icp7_reduction2_one_row_per_patch_buf, in_pad="2 2 2 2",stride="1 1" ) # one copy per output elem | |
| icp7_out2 = icp7_reduction2_one_row_per_patch_buf * transpose(reshape(icp7_out2_filts,128,800)) # sgemm: MxNxK == 196x800x128 | |
| ReLU(name="relu_icp7_out2",in_place=[icp7_out2]) | |
| icp7_pool = NDA("icp7_pool",num_img,528,14,14) # num,chan,y,x | |
| Pooling(name="icp7_pool",bots=[ icp6_out ],tops=[ icp7_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp7_out3_filts = NDA("icp7_out3_filts",128,528,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp7_out3_biases = NDA("icp7_out3_biases",128) # SOURCE out_chan | |
| icp7_out3 = NDA("icp7_out3",num_img,128,14,14) # num,chan,y,x | |
| icp7_pool_one_row_per_patch_buf = NDA("icp7_pool_one_row_per_patch_buf",196,528) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp7_pool[i,:,:,:], dest=icp7_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp7_out3 = icp7_pool_one_row_per_patch_buf * transpose(reshape(icp7_out3_filts,128,528)) # sgemm: MxNxK == 196x528x128 | |
| ReLU(name="relu_icp7_out3",in_place=[icp7_out3]) | |
| icp7_out0_filts = NDA("icp7_out0_filts",256,528,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp7_out0_biases = NDA("icp7_out0_biases",256) # SOURCE out_chan | |
| icp7_out0 = NDA("icp7_out0",num_img,256,14,14) # num,chan,y,x | |
| icp6_out_one_row_per_patch_buf = NDA("icp6_out_one_row_per_patch_buf",196,528) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp6_out[i,:,:,:], dest=icp6_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp7_out0 = icp6_out_one_row_per_patch_buf * transpose(reshape(icp7_out0_filts,256,528)) # sgemm: MxNxK == 196x528x256 | |
| ReLU(name="relu_icp7_out0",in_place=[icp7_out0]) | |
| icp7_out = NDA("icp7_out",num_img,832,14,14) # num,chan,y,x | |
| Concat(name="icp7_out",bots=[ icp7_out0, icp7_out1, icp7_out2, icp7_out3 ],tops=[ icp7_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp8_in = NDA("icp8_in",num_img,832,7,7) # num,chan,y,x | |
| Pooling(name="icp8_in",bots=[ icp7_out ],tops=[ icp8_in ], | |
| in_pad="0 0 0 0",stride="2 2") | |
| icp8_reduction1_filts = NDA("icp8_reduction1_filts",160,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp8_reduction1_biases = NDA("icp8_reduction1_biases",160) # SOURCE out_chan | |
| icp8_reduction1 = NDA("icp8_reduction1",num_img,160,7,7) # num,chan,y,x | |
| icp8_in_one_row_per_patch_buf = NDA("icp8_in_one_row_per_patch_buf",49,832) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp8_in[i,:,:,:], dest=icp8_in_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp8_reduction1 = icp8_in_one_row_per_patch_buf * transpose(reshape(icp8_reduction1_filts,160,832)) # sgemm: MxNxK == 49x832x160 | |
| ReLU(name="relu_icp8_reduction1",in_place=[icp8_reduction1]) | |
| icp8_out1_filts = NDA("icp8_out1_filts",320,160,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp8_out1_biases = NDA("icp8_out1_biases",320) # SOURCE out_chan | |
| icp8_out1 = NDA("icp8_out1",num_img,320,7,7) # num,chan,y,x | |
| icp8_reduction1_one_row_per_patch_buf = NDA("icp8_reduction1_one_row_per_patch_buf",49,1440) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp8_reduction1[i,:,:,:], dest=icp8_reduction1_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| icp8_out1 = icp8_reduction1_one_row_per_patch_buf * transpose(reshape(icp8_out1_filts,320,1440)) # sgemm: MxNxK == 49x1440x320 | |
| ReLU(name="relu_icp8_out1",in_place=[icp8_out1]) | |
| icp8_reduction2_filts = NDA("icp8_reduction2_filts",32,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp8_reduction2_biases = NDA("icp8_reduction2_biases",32) # SOURCE out_chan | |
| icp8_reduction2 = NDA("icp8_reduction2",num_img,32,7,7) # num,chan,y,x | |
| icp8_in_one_row_per_patch_buf = NDA("icp8_in_one_row_per_patch_buf",49,832) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp8_in[i,:,:,:], dest=icp8_in_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp8_reduction2 = icp8_in_one_row_per_patch_buf * transpose(reshape(icp8_reduction2_filts,32,832)) # sgemm: MxNxK == 49x832x32 | |
| ReLU(name="relu_icp8_reduction2",in_place=[icp8_reduction2]) | |
| icp8_out2_filts = NDA("icp8_out2_filts",128,32,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp8_out2_biases = NDA("icp8_out2_biases",128) # SOURCE out_chan | |
| icp8_out2 = NDA("icp8_out2",num_img,128,7,7) # num,chan,y,x | |
| icp8_reduction2_one_row_per_patch_buf = NDA("icp8_reduction2_one_row_per_patch_buf",49,800) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp8_reduction2[i,:,:,:], dest=icp8_reduction2_one_row_per_patch_buf, in_pad="2 2 2 2",stride="1 1" ) # one copy per output elem | |
| icp8_out2 = icp8_reduction2_one_row_per_patch_buf * transpose(reshape(icp8_out2_filts,128,800)) # sgemm: MxNxK == 49x800x128 | |
| ReLU(name="relu_icp8_out2",in_place=[icp8_out2]) | |
| icp8_pool = NDA("icp8_pool",num_img,832,7,7) # num,chan,y,x | |
| Pooling(name="icp8_pool",bots=[ icp8_in ],tops=[ icp8_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp8_out3_filts = NDA("icp8_out3_filts",128,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp8_out3_biases = NDA("icp8_out3_biases",128) # SOURCE out_chan | |
| icp8_out3 = NDA("icp8_out3",num_img,128,7,7) # num,chan,y,x | |
| icp8_pool_one_row_per_patch_buf = NDA("icp8_pool_one_row_per_patch_buf",49,832) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp8_pool[i,:,:,:], dest=icp8_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp8_out3 = icp8_pool_one_row_per_patch_buf * transpose(reshape(icp8_out3_filts,128,832)) # sgemm: MxNxK == 49x832x128 | |
| ReLU(name="relu_icp8_out3",in_place=[icp8_out3]) | |
| icp8_out0_filts = NDA("icp8_out0_filts",256,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp8_out0_biases = NDA("icp8_out0_biases",256) # SOURCE out_chan | |
| icp8_out0 = NDA("icp8_out0",num_img,256,7,7) # num,chan,y,x | |
| icp8_in_one_row_per_patch_buf = NDA("icp8_in_one_row_per_patch_buf",49,832) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp8_in[i,:,:,:], dest=icp8_in_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp8_out0 = icp8_in_one_row_per_patch_buf * transpose(reshape(icp8_out0_filts,256,832)) # sgemm: MxNxK == 49x832x256 | |
| ReLU(name="relu_icp8_out0",in_place=[icp8_out0]) | |
| icp8_out = NDA("icp8_out",num_img,832,7,7) # num,chan,y,x | |
| Concat(name="icp8_out",bots=[ icp8_out0, icp8_out1, icp8_out2, icp8_out3 ],tops=[ icp8_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| icp9_reduction1_filts = NDA("icp9_reduction1_filts",192,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp9_reduction1_biases = NDA("icp9_reduction1_biases",192) # SOURCE out_chan | |
| icp9_reduction1 = NDA("icp9_reduction1",num_img,192,7,7) # num,chan,y,x | |
| icp8_out_one_row_per_patch_buf = NDA("icp8_out_one_row_per_patch_buf",49,832) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp8_out[i,:,:,:], dest=icp8_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp9_reduction1 = icp8_out_one_row_per_patch_buf * transpose(reshape(icp9_reduction1_filts,192,832)) # sgemm: MxNxK == 49x832x192 | |
| ReLU(name="relu_icp9_reduction1",in_place=[icp9_reduction1]) | |
| icp9_out1_filts = NDA("icp9_out1_filts",384,192,3,3) # SOURCE out_chan,in_chan,y,x | |
| icp9_out1_biases = NDA("icp9_out1_biases",384) # SOURCE out_chan | |
| icp9_out1 = NDA("icp9_out1",num_img,384,7,7) # num,chan,y,x | |
| icp9_reduction1_one_row_per_patch_buf = NDA("icp9_reduction1_one_row_per_patch_buf",49,1728) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp9_reduction1[i,:,:,:], dest=icp9_reduction1_one_row_per_patch_buf, in_pad="1 1 1 1",stride="1 1" ) # one copy per output elem | |
| icp9_out1 = icp9_reduction1_one_row_per_patch_buf * transpose(reshape(icp9_out1_filts,384,1728)) # sgemm: MxNxK == 49x1728x384 | |
| ReLU(name="relu_icp9_out1",in_place=[icp9_out1]) | |
| icp9_reduction2_filts = NDA("icp9_reduction2_filts",48,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp9_reduction2_biases = NDA("icp9_reduction2_biases",48) # SOURCE out_chan | |
| icp9_reduction2 = NDA("icp9_reduction2",num_img,48,7,7) # num,chan,y,x | |
| icp8_out_one_row_per_patch_buf = NDA("icp8_out_one_row_per_patch_buf",49,832) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp8_out[i,:,:,:], dest=icp8_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp9_reduction2 = icp8_out_one_row_per_patch_buf * transpose(reshape(icp9_reduction2_filts,48,832)) # sgemm: MxNxK == 49x832x48 | |
| ReLU(name="relu_icp9_reduction2",in_place=[icp9_reduction2]) | |
| icp9_out2_filts = NDA("icp9_out2_filts",128,48,5,5) # SOURCE out_chan,in_chan,y,x | |
| icp9_out2_biases = NDA("icp9_out2_biases",128) # SOURCE out_chan | |
| icp9_out2 = NDA("icp9_out2",num_img,128,7,7) # num,chan,y,x | |
| icp9_reduction2_one_row_per_patch_buf = NDA("icp9_reduction2_one_row_per_patch_buf",49,1200) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp9_reduction2[i,:,:,:], dest=icp9_reduction2_one_row_per_patch_buf, in_pad="2 2 2 2",stride="1 1" ) # one copy per output elem | |
| icp9_out2 = icp9_reduction2_one_row_per_patch_buf * transpose(reshape(icp9_out2_filts,128,1200)) # sgemm: MxNxK == 49x1200x128 | |
| ReLU(name="relu_icp9_out2",in_place=[icp9_out2]) | |
| icp9_pool = NDA("icp9_pool",num_img,832,7,7) # num,chan,y,x | |
| Pooling(name="icp9_pool",bots=[ icp8_out ],tops=[ icp9_pool ], | |
| in_pad="1 1 1 1",stride="1 1") | |
| icp9_out3_filts = NDA("icp9_out3_filts",128,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp9_out3_biases = NDA("icp9_out3_biases",128) # SOURCE out_chan | |
| icp9_out3 = NDA("icp9_out3",num_img,128,7,7) # num,chan,y,x | |
| icp9_pool_one_row_per_patch_buf = NDA("icp9_pool_one_row_per_patch_buf",49,832) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp9_pool[i,:,:,:], dest=icp9_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp9_out3 = icp9_pool_one_row_per_patch_buf * transpose(reshape(icp9_out3_filts,128,832)) # sgemm: MxNxK == 49x832x128 | |
| ReLU(name="relu_icp9_out3",in_place=[icp9_out3]) | |
| icp9_out0_filts = NDA("icp9_out0_filts",384,832,1,1) # SOURCE out_chan,in_chan,y,x | |
| icp9_out0_biases = NDA("icp9_out0_biases",384) # SOURCE out_chan | |
| icp9_out0 = NDA("icp9_out0",num_img,384,7,7) # num,chan,y,x | |
| icp8_out_one_row_per_patch_buf = NDA("icp8_out_one_row_per_patch_buf",49,832) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=icp8_out[i,:,:,:], dest=icp8_out_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| icp9_out0 = icp8_out_one_row_per_patch_buf * transpose(reshape(icp9_out0_filts,384,832)) # sgemm: MxNxK == 49x832x384 | |
| ReLU(name="relu_icp9_out0",in_place=[icp9_out0]) | |
| icp9_out = NDA("icp9_out",num_img,1024,7,7) # num,chan,y,x | |
| Concat(name="icp9_out",bots=[ icp9_out0, icp9_out1, icp9_out2, icp9_out3 ],tops=[ icp9_out ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| cls3_pool = NDA("cls3_pool",num_img,1024,1,1) # num,chan,y,x | |
| Pooling(name="cls3_pool",bots=[ icp9_out ],tops=[ cls3_pool ], | |
| in_pad="0 0 0 0",stride="1 1") | |
| Dropout(name="cls3_drop",in_place=[cls3_pool]) | |
| cls3_fc__filts = NDA("cls3_fc__filts",33,1024,1,1) # SOURCE out_chan,in_chan,y,x | |
| cls3_fc__biases = NDA("cls3_fc__biases",33) # SOURCE out_chan | |
| cls3_fc_ = NDA("cls3_fc_",num_img,33,1,1) # SINK num,chan,y,x | |
| cls3_pool_one_row_per_patch_buf = NDA("cls3_pool_one_row_per_patch_buf",1,1024) | |
| for i in range(0,num_img): | |
| patches_to_rows( src=cls3_pool[i,:,:,:], dest=cls3_pool_one_row_per_patch_buf, in_pad="0 0 0 0",stride="1 1" ) # one copy per output elem | |
| cls3_fc_ = cls3_pool_one_row_per_patch_buf * transpose(reshape(cls3_fc__filts,33,1024)) # sgemm: MxNxK == 1x1024x33 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| name: "GoogLeNet" | |
| layers { | |
| name: "data" | |
| type: DATA | |
| top: "data" | |
| top: "label" | |
| data_param { | |
| #source: "train_lmdb" | |
| #source: "/scratch/forresti/ilsvrc2012/ilsvrc2012_train_256x256_lmdb" | |
| #source: "/nscratch/forresti/FlickrLogos-32/FlickrLogos-32_trainval_lmdb" | |
| #source: "/nscratch/forresti/FlickrLogos-32/FlickrLogos-32_trainval_logosonly_lmdb" | |
| source: "/scratch/datasets/imagenet_classification/ilsvrc12_train_lmdb" | |
| backend: LMDB | |
| batch_size: 32 | |
| } | |
| transform_param { | |
| crop_size: 224 | |
| #mean_file: "network5_mean.binaryproto" | |
| #mean_file: "/home/eecs/forresti/caffe_depthMax_and_hog/data/ilsvrc12/imagenet_mean.binaryproto" | |
| mean_file: "/scratch/datasets/data/ilsvrc12/imagenet_mean.binaryproto" | |
| mirror: true | |
| } | |
| include: { phase: TRAIN } | |
| } | |
| layers { | |
| name: "data" | |
| type: DATA | |
| top: "data" | |
| top: "label" | |
| data_param { | |
| #source: "test_lmdb" | |
| #source: "/scratch/forresti/ilsvrc2012/ilsvrc2012_val_256x256_lmdb" | |
| #source: "/nscratch/forresti/FlickrLogos-32/FlickrLogos-32_test_lmdb" | |
| #source: "/nscratch/forresti/FlickrLogos-32/FlickrLogos-32_test_logosonly_lmdb" | |
| source: "/scratch/datasets/imagenet_classification/ilsvrc12_train_lmdb" | |
| backend: LMDB | |
| #batch_size: 32 | |
| batch_size: 25 | |
| } | |
| transform_param { | |
| crop_size: 224 | |
| #mean_file: "network5_mean.binaryproto" | |
| #mean_file: "/home/eecs/forresti/caffe_depthMax_and_hog/data/ilsvrc12/imagenet_mean.binaryproto" | |
| mean_file: "/scratch/datasets/data/ilsvrc12/imagenet_mean.binaryproto" | |
| mirror: false | |
| } | |
| include: { phase: TEST } | |
| } | |
| layers { | |
| name: "conv1" | |
| type: CONVOLUTION | |
| bottom: "data" | |
| top: "conv1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 64 | |
| kernel_size: 7 | |
| stride: 2 | |
| pad: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.015 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu1" | |
| type: RELU | |
| bottom: "conv1" | |
| top: "conv1" | |
| } | |
| layers { | |
| name: "pool1" | |
| type: POOLING | |
| bottom: "conv1" | |
| top: "pool1" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 2 | |
| pad: 0 | |
| } | |
| } | |
| layers { | |
| name: "norm1" | |
| type: LRN | |
| bottom: "pool1" | |
| top: "norm1" | |
| lrn_param { | |
| local_size: 5 | |
| alpha: 0.0001 | |
| beta: 0.75 | |
| } | |
| } | |
| layers { | |
| name: "reduction2" | |
| type: CONVOLUTION | |
| bottom: "norm1" | |
| top: "reduction2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 64 | |
| pad: 0 | |
| kernel_size: 1 | |
| group: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_reduction2" | |
| type: RELU | |
| bottom: "reduction2" | |
| top: "reduction2" | |
| } | |
| layers { | |
| name: "conv2" | |
| type: CONVOLUTION | |
| bottom: "reduction2" | |
| top: "conv2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 192 | |
| pad: 1 | |
| kernel_size: 3 | |
| group: 1 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.02 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu2" | |
| type: RELU | |
| bottom: "conv2" | |
| top: "conv2" | |
| } | |
| layers { | |
| name: "norm2" | |
| type: LRN | |
| bottom: "conv2" | |
| top: "norm2" | |
| lrn_param { | |
| local_size: 5 | |
| alpha: 0.0001 | |
| beta: 0.75 | |
| } | |
| } | |
| layers { | |
| name: "pool2" | |
| type: POOLING | |
| bottom: "norm2" | |
| top: "pool2" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 2 | |
| pad: 0 | |
| } | |
| } | |
| # Inception module 1 *************** | |
| layers { | |
| name: "icp1_reduction1" | |
| type: CONVOLUTION | |
| bottom: "pool2" | |
| top: "icp1_reduction1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 96 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp1_reduction1" | |
| type: RELU | |
| bottom: "icp1_reduction1" | |
| top: "icp1_reduction1" | |
| } | |
| layers { | |
| name: "icp1_reduction2" | |
| type: CONVOLUTION | |
| bottom: "pool2" | |
| top: "icp1_reduction2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 16 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp1_reduction2" | |
| type: RELU | |
| bottom: "icp1_reduction2" | |
| top: "icp1_reduction2" | |
| } | |
| layers { | |
| name: "icp1_pool" | |
| type: POOLING | |
| bottom: "pool2" | |
| top: "icp1_pool" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 1 | |
| pad: 1 | |
| } | |
| } | |
| # *********** | |
| layers { | |
| name: "icp1_out0" | |
| type: CONVOLUTION | |
| bottom: "pool2" | |
| top: "icp1_out0" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 64 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp1_out0" | |
| type: RELU | |
| bottom: "icp1_out0" | |
| top: "icp1_out0" | |
| } | |
| layers { | |
| name: "icp1_out1" | |
| type: CONVOLUTION | |
| bottom: "icp1_reduction1" | |
| top: "icp1_out1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.04 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp1_out1" | |
| type: RELU | |
| bottom: "icp1_out1" | |
| top: "icp1_out1" | |
| } | |
| layers { | |
| name: "icp1_out2" | |
| type: CONVOLUTION | |
| bottom: "icp1_reduction2" | |
| top: "icp1_out2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 32 | |
| pad: 2 | |
| kernel_size: 5 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.08 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp1_out2" | |
| type: RELU | |
| bottom: "icp1_out2" | |
| top: "icp1_out2" | |
| } | |
| layers { | |
| name: "icp1_out3" | |
| type: CONVOLUTION | |
| bottom: "icp1_pool" | |
| top: "icp1_out3" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 32 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp1_out3" | |
| type: RELU | |
| bottom: "icp1_out3" | |
| top: "icp1_out3" | |
| } | |
| # Concat them together | |
| layers { | |
| name: "icp2_in" | |
| type: CONCAT | |
| bottom: "icp1_out0" | |
| bottom: "icp1_out1" | |
| bottom: "icp1_out2" | |
| bottom: "icp1_out3" | |
| top: "icp2_in" | |
| } | |
| # Inception module 2 *************** | |
| layers { | |
| name: "icp2_reduction1" | |
| type: CONVOLUTION | |
| bottom: "icp2_in" | |
| top: "icp2_reduction1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp2_reduction1" | |
| type: RELU | |
| bottom: "icp2_reduction1" | |
| top: "icp2_reduction1" | |
| } | |
| layers { | |
| name: "icp2_reduction2" | |
| type: CONVOLUTION | |
| bottom: "icp2_in" | |
| top: "icp2_reduction2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 32 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp2_reduction2" | |
| type: RELU | |
| bottom: "icp2_reduction2" | |
| top: "icp2_reduction2" | |
| } | |
| layers { | |
| name: "icp2_pool" | |
| type: POOLING | |
| bottom: "icp2_in" | |
| top: "icp2_pool" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 1 | |
| pad: 1 | |
| } | |
| } | |
| # *********** | |
| layers { | |
| name: "icp2_out0" | |
| type: CONVOLUTION | |
| bottom: "icp2_in" | |
| top: "icp2_out0" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp2_out0" | |
| type: RELU | |
| bottom: "icp2_out0" | |
| top: "icp2_out0" | |
| } | |
| layers { | |
| name: "icp2_out1" | |
| type: CONVOLUTION | |
| bottom: "icp2_reduction1" | |
| top: "icp2_out1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 192 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.04 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp2_out1" | |
| type: RELU | |
| bottom: "icp2_out1" | |
| top: "icp2_out1" | |
| } | |
| layers { | |
| name: "icp2_out2" | |
| type: CONVOLUTION | |
| bottom: "icp2_reduction2" | |
| top: "icp2_out2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 96 | |
| pad: 2 | |
| kernel_size: 5 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.08 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp2_out2" | |
| type: RELU | |
| bottom: "icp2_out2" | |
| top: "icp2_out2" | |
| } | |
| layers { | |
| name: "icp2_out3" | |
| type: CONVOLUTION | |
| bottom: "icp2_pool" | |
| top: "icp2_out3" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 64 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp2_out3" | |
| type: RELU | |
| bottom: "icp2_out3" | |
| top: "icp2_out3" | |
| } | |
| # Concat them together | |
| layers { | |
| name: "icp2_out" | |
| type: CONCAT | |
| bottom: "icp2_out0" | |
| bottom: "icp2_out1" | |
| bottom: "icp2_out2" | |
| bottom: "icp2_out3" | |
| top: "icp2_out" | |
| } | |
| layers { | |
| name: "icp3_in" | |
| type: POOLING | |
| bottom: "icp2_out" | |
| top: "icp3_in" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 2 | |
| pad: 0 | |
| } | |
| } | |
| # Inception module 3 *************** | |
| layers { | |
| name: "icp3_reduction1" | |
| type: CONVOLUTION | |
| bottom: "icp3_in" | |
| top: "icp3_reduction1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 96 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp3_reduction1" | |
| type: RELU | |
| bottom: "icp3_reduction1" | |
| top: "icp3_reduction1" | |
| } | |
| layers { | |
| name: "icp3_reduction2" | |
| type: CONVOLUTION | |
| bottom: "icp3_in" | |
| top: "icp3_reduction2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 16 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp3_reduction2" | |
| type: RELU | |
| bottom: "icp3_reduction2" | |
| top: "icp3_reduction2" | |
| } | |
| layers { | |
| name: "icp3_pool" | |
| type: POOLING | |
| bottom: "icp3_in" | |
| top: "icp3_pool" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 1 | |
| pad: 1 | |
| } | |
| } | |
| # *********** | |
| layers { | |
| name: "icp3_out0" | |
| type: CONVOLUTION | |
| bottom: "icp3_in" | |
| top: "icp3_out0" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 192 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp3_out0" | |
| type: RELU | |
| bottom: "icp3_out0" | |
| top: "icp3_out0" | |
| } | |
| layers { | |
| name: "icp3_out1" | |
| type: CONVOLUTION | |
| bottom: "icp3_reduction1" | |
| top: "icp3_out1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 208 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.04 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp3_out1" | |
| type: RELU | |
| bottom: "icp3_out1" | |
| top: "icp3_out1" | |
| } | |
| layers { | |
| name: "icp3_out2" | |
| type: CONVOLUTION | |
| bottom: "icp3_reduction2" | |
| top: "icp3_out2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 48 | |
| pad: 2 | |
| kernel_size: 5 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.08 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp3_out2" | |
| type: RELU | |
| bottom: "icp3_out2" | |
| top: "icp3_out2" | |
| } | |
| layers { | |
| name: "icp3_out3" | |
| type: CONVOLUTION | |
| bottom: "icp3_pool" | |
| top: "icp3_out3" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 64 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp3_out3" | |
| type: RELU | |
| bottom: "icp3_out3" | |
| top: "icp3_out3" | |
| } | |
| # Concat them together | |
| layers { | |
| name: "icp3_out" | |
| type: CONCAT | |
| bottom: "icp3_out0" | |
| bottom: "icp3_out1" | |
| bottom: "icp3_out2" | |
| bottom: "icp3_out3" | |
| top: "icp3_out" | |
| } | |
| # first classification branch ************ | |
| layers { | |
| name: "cls1_pool" | |
| type: POOLING | |
| bottom: "icp3_out" | |
| top: "cls1_pool" | |
| pooling_param { | |
| pool: AVE | |
| kernel_size: 5 | |
| stride: 3 | |
| pad: 0 | |
| # this padding is somewhat special | |
| } | |
| } | |
| layers { | |
| name: "cls1_reduction" | |
| type: CONVOLUTION | |
| bottom: "cls1_pool" | |
| top: "cls1_reduction" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_cls1_reduction" | |
| type: RELU | |
| bottom: "cls1_reduction" | |
| top: "cls1_reduction" | |
| } | |
| layers { | |
| name: "cls1_fc1" | |
| type: INNER_PRODUCT | |
| bottom: "cls1_reduction" | |
| top: "cls1_fc1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| inner_product_param { | |
| num_output: 1024 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.01 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_cls1_fc1" | |
| type: RELU | |
| bottom: "cls1_fc1" | |
| top: "cls1_fc1" | |
| } | |
| layers { | |
| name: "cls1_drop" | |
| type: DROPOUT | |
| bottom: "cls1_fc1" | |
| top: "cls1_fc1" | |
| dropout_param { | |
| dropout_ratio: 0.7 | |
| } | |
| } | |
| layers { | |
| name: "cls1_fc2_" | |
| type: INNER_PRODUCT | |
| bottom: "cls1_fc1" | |
| top: "cls1_fc2_" | |
| #blobs_lr: 1 | |
| #blobs_lr: 2 | |
| # blobs_lr is set to higher than for other layers, because this layer is starting from random while the others are already trained | |
| blobs_lr: 10 | |
| blobs_lr: 20 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| inner_product_param { | |
| #num_output: 1000 | |
| num_output: 33 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.01 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "loss1" | |
| type: SOFTMAX_LOSS | |
| bottom: "cls1_fc2_" | |
| bottom: "label" | |
| top: "loss1" | |
| loss_weight: 1 | |
| } | |
| # Inception module 4 *************** | |
| layers { | |
| name: "icp4_reduction1" | |
| type: CONVOLUTION | |
| bottom: "icp3_out" | |
| top: "icp4_reduction1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 112 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp4_reduction1" | |
| type: RELU | |
| bottom: "icp4_reduction1" | |
| top: "icp4_reduction1" | |
| } | |
| layers { | |
| name: "icp4_reduction2" | |
| type: CONVOLUTION | |
| bottom: "icp3_out" | |
| top: "icp4_reduction2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 24 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp4_reduction2" | |
| type: RELU | |
| bottom: "icp4_reduction2" | |
| top: "icp4_reduction2" | |
| } | |
| layers { | |
| name: "icp4_pool" | |
| type: POOLING | |
| bottom: "icp3_out" | |
| top: "icp4_pool" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 1 | |
| pad: 1 | |
| } | |
| } | |
| # *********** | |
| layers { | |
| name: "icp4_out0" | |
| type: CONVOLUTION | |
| bottom: "icp3_out" | |
| top: "icp4_out0" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 160 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp4_out0" | |
| type: RELU | |
| bottom: "icp4_out0" | |
| top: "icp4_out0" | |
| } | |
| layers { | |
| name: "icp4_out1" | |
| type: CONVOLUTION | |
| bottom: "icp4_reduction1" | |
| top: "icp4_out1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 224 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.04 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp4_out1" | |
| type: RELU | |
| bottom: "icp4_out1" | |
| top: "icp4_out1" | |
| } | |
| layers { | |
| name: "icp4_out2" | |
| type: CONVOLUTION | |
| bottom: "icp4_reduction2" | |
| top: "icp4_out2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 64 | |
| pad: 2 | |
| kernel_size: 5 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.08 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp4_out2" | |
| type: RELU | |
| bottom: "icp4_out2" | |
| top: "icp4_out2" | |
| } | |
| layers { | |
| name: "icp4_out3" | |
| type: CONVOLUTION | |
| bottom: "icp4_pool" | |
| top: "icp4_out3" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 64 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp4_out3" | |
| type: RELU | |
| bottom: "icp4_out3" | |
| top: "icp4_out3" | |
| } | |
| # Concat them together | |
| layers { | |
| name: "icp4_out" | |
| type: CONCAT | |
| bottom: "icp4_out0" | |
| bottom: "icp4_out1" | |
| bottom: "icp4_out2" | |
| bottom: "icp4_out3" | |
| top: "icp4_out" | |
| } | |
| # Inception module 5 *************** | |
| layers { | |
| name: "icp5_reduction1" | |
| type: CONVOLUTION | |
| bottom: "icp4_out" | |
| top: "icp5_reduction1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp5_reduction1" | |
| type: RELU | |
| bottom: "icp5_reduction1" | |
| top: "icp5_reduction1" | |
| } | |
| layers { | |
| name: "icp5_reduction2" | |
| type: CONVOLUTION | |
| bottom: "icp4_out" | |
| top: "icp5_reduction2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 24 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp5_reduction2" | |
| type: RELU | |
| bottom: "icp5_reduction2" | |
| top: "icp5_reduction2" | |
| } | |
| layers { | |
| name: "icp5_pool" | |
| type: POOLING | |
| bottom: "icp4_out" | |
| top: "icp5_pool" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 1 | |
| pad: 1 | |
| } | |
| } | |
| # *********** | |
| layers { | |
| name: "icp5_out0" | |
| type: CONVOLUTION | |
| bottom: "icp4_out" | |
| top: "icp5_out0" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp5_out0" | |
| type: RELU | |
| bottom: "icp5_out0" | |
| top: "icp5_out0" | |
| } | |
| layers { | |
| name: "icp5_out1" | |
| type: CONVOLUTION | |
| bottom: "icp5_reduction1" | |
| top: "icp5_out1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 256 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.04 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp5_out1" | |
| type: RELU | |
| bottom: "icp5_out1" | |
| top: "icp5_out1" | |
| } | |
| layers { | |
| name: "icp5_out2" | |
| type: CONVOLUTION | |
| bottom: "icp5_reduction2" | |
| top: "icp5_out2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 64 | |
| pad: 2 | |
| kernel_size: 5 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.08 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp5_out2" | |
| type: RELU | |
| bottom: "icp5_out2" | |
| top: "icp5_out2" | |
| } | |
| layers { | |
| name: "icp5_out3" | |
| type: CONVOLUTION | |
| bottom: "icp5_pool" | |
| top: "icp5_out3" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 64 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp5_out3" | |
| type: RELU | |
| bottom: "icp5_out3" | |
| top: "icp5_out3" | |
| } | |
| # Concat them together | |
| layers { | |
| name: "icp5_out" | |
| type: CONCAT | |
| bottom: "icp5_out0" | |
| bottom: "icp5_out1" | |
| bottom: "icp5_out2" | |
| bottom: "icp5_out3" | |
| top: "icp5_out" | |
| } | |
| # Inception module 6 *************** | |
| layers { | |
| name: "icp6_reduction1" | |
| type: CONVOLUTION | |
| bottom: "icp5_out" | |
| top: "icp6_reduction1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 144 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp6_reduction1" | |
| type: RELU | |
| bottom: "icp6_reduction1" | |
| top: "icp6_reduction1" | |
| } | |
| layers { | |
| name: "icp6_reduction2" | |
| type: CONVOLUTION | |
| bottom: "icp5_out" | |
| top: "icp6_reduction2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 32 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp6_reduction2" | |
| type: RELU | |
| bottom: "icp6_reduction2" | |
| top: "icp6_reduction2" | |
| } | |
| layers { | |
| name: "icp6_pool" | |
| type: POOLING | |
| bottom: "icp5_out" | |
| top: "icp6_pool" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 1 | |
| pad: 1 | |
| } | |
| } | |
| # *********** | |
| layers { | |
| name: "icp6_out0" | |
| type: CONVOLUTION | |
| bottom: "icp5_out" | |
| top: "icp6_out0" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 112 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp6_out0" | |
| type: RELU | |
| bottom: "icp6_out0" | |
| top: "icp6_out0" | |
| } | |
| layers { | |
| name: "icp6_out1" | |
| type: CONVOLUTION | |
| bottom: "icp6_reduction1" | |
| top: "icp6_out1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 288 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.04 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp6_out1" | |
| type: RELU | |
| bottom: "icp6_out1" | |
| top: "icp6_out1" | |
| } | |
| layers { | |
| name: "icp6_out2" | |
| type: CONVOLUTION | |
| bottom: "icp6_reduction2" | |
| top: "icp6_out2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 64 | |
| pad: 2 | |
| kernel_size: 5 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.08 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp6_out2" | |
| type: RELU | |
| bottom: "icp6_out2" | |
| top: "icp6_out2" | |
| } | |
| layers { | |
| name: "icp6_out3" | |
| type: CONVOLUTION | |
| bottom: "icp6_pool" | |
| top: "icp6_out3" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 64 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp6_out3" | |
| type: RELU | |
| bottom: "icp6_out3" | |
| top: "icp6_out3" | |
| } | |
| # Concat them together | |
| layers { | |
| name: "icp6_out" | |
| type: CONCAT | |
| bottom: "icp6_out0" | |
| bottom: "icp6_out1" | |
| bottom: "icp6_out2" | |
| bottom: "icp6_out3" | |
| top: "icp6_out" | |
| } | |
| # second classification branch ************ | |
| layers { | |
| name: "cls2_pool" | |
| type: POOLING | |
| bottom: "icp6_out" | |
| top: "cls2_pool" | |
| pooling_param { | |
| pool: AVE | |
| kernel_size: 5 | |
| stride: 3 | |
| pad: 0 | |
| # this padding is somewhat special | |
| } | |
| } | |
| layers { | |
| name: "cls2_reduction" | |
| type: CONVOLUTION | |
| bottom: "cls2_pool" | |
| top: "cls2_reduction" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_cls2_reduction" | |
| type: RELU | |
| bottom: "cls2_reduction" | |
| top: "cls2_reduction" | |
| } | |
| layers { | |
| name: "cls2_fc1" | |
| type: INNER_PRODUCT | |
| bottom: "cls2_reduction" | |
| top: "cls2_fc1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| inner_product_param { | |
| num_output: 1024 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.01 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_cls2_fc1" | |
| type: RELU | |
| bottom: "cls2_fc1" | |
| top: "cls2_fc1" | |
| } | |
| layers { | |
| name: "cls2_drop" | |
| type: DROPOUT | |
| bottom: "cls2_fc1" | |
| top: "cls2_fc1" | |
| dropout_param { | |
| dropout_ratio: 0.7 | |
| } | |
| } | |
| layers { | |
| name: "cls2_fc2_" | |
| type: INNER_PRODUCT | |
| bottom: "cls2_fc1" | |
| top: "cls2_fc2_" | |
| #blobs_lr: 1 | |
| #blobs_lr: 2 | |
| # blobs_lr is set to higher than for other layers, because this layer is starting from random while the others are already trained | |
| blobs_lr: 10 | |
| blobs_lr: 20 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| inner_product_param { | |
| #num_output: 1000 | |
| num_output: 33 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.01 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "loss2" | |
| type: SOFTMAX_LOSS | |
| bottom: "cls2_fc2_" | |
| bottom: "label" | |
| top: "loss2" | |
| loss_weight: 1 | |
| } | |
| # Inception module 7 *************** | |
| layers { | |
| name: "icp7_reduction1" | |
| type: CONVOLUTION | |
| bottom: "icp6_out" | |
| top: "icp7_reduction1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 160 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp7_reduction1" | |
| type: RELU | |
| bottom: "icp7_reduction1" | |
| top: "icp7_reduction1" | |
| } | |
| layers { | |
| name: "icp7_reduction2" | |
| type: CONVOLUTION | |
| bottom: "icp6_out" | |
| top: "icp7_reduction2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 32 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp7_reduction2" | |
| type: RELU | |
| bottom: "icp7_reduction2" | |
| top: "icp7_reduction2" | |
| } | |
| layers { | |
| name: "icp7_pool" | |
| type: POOLING | |
| bottom: "icp6_out" | |
| top: "icp7_pool" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 1 | |
| pad: 1 | |
| } | |
| } | |
| # *********** | |
| layers { | |
| name: "icp7_out0" | |
| type: CONVOLUTION | |
| bottom: "icp6_out" | |
| top: "icp7_out0" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 256 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp7_out0" | |
| type: RELU | |
| bottom: "icp7_out0" | |
| top: "icp7_out0" | |
| } | |
| layers { | |
| name: "icp7_out1" | |
| type: CONVOLUTION | |
| bottom: "icp7_reduction1" | |
| top: "icp7_out1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 320 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.04 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp7_out1" | |
| type: RELU | |
| bottom: "icp7_out1" | |
| top: "icp7_out1" | |
| } | |
| layers { | |
| name: "icp7_out2" | |
| type: CONVOLUTION | |
| bottom: "icp7_reduction2" | |
| top: "icp7_out2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 2 | |
| kernel_size: 5 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.08 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp7_out2" | |
| type: RELU | |
| bottom: "icp7_out2" | |
| top: "icp7_out2" | |
| } | |
| layers { | |
| name: "icp7_out3" | |
| type: CONVOLUTION | |
| bottom: "icp7_pool" | |
| top: "icp7_out3" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp7_out3" | |
| type: RELU | |
| bottom: "icp7_out3" | |
| top: "icp7_out3" | |
| } | |
| # Concat them together | |
| layers { | |
| name: "icp7_out" | |
| type: CONCAT | |
| bottom: "icp7_out0" | |
| bottom: "icp7_out1" | |
| bottom: "icp7_out2" | |
| bottom: "icp7_out3" | |
| top: "icp7_out" | |
| } | |
| layers { | |
| name: "icp8_in" | |
| type: POOLING | |
| bottom: "icp7_out" | |
| top: "icp8_in" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 2 | |
| pad: 0 | |
| } | |
| } | |
| # Inception module 8 *************** | |
| layers { | |
| name: "icp8_reduction1" | |
| type: CONVOLUTION | |
| bottom: "icp8_in" | |
| top: "icp8_reduction1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 160 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp8_reduction1" | |
| type: RELU | |
| bottom: "icp8_reduction1" | |
| top: "icp8_reduction1" | |
| } | |
| layers { | |
| name: "icp8_reduction2" | |
| type: CONVOLUTION | |
| bottom: "icp8_in" | |
| top: "icp8_reduction2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 32 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp8_reduction2" | |
| type: RELU | |
| bottom: "icp8_reduction2" | |
| top: "icp8_reduction2" | |
| } | |
| layers { | |
| name: "icp8_pool" | |
| type: POOLING | |
| bottom: "icp8_in" | |
| top: "icp8_pool" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 1 | |
| pad: 1 | |
| } | |
| } | |
| # *********** | |
| layers { | |
| name: "icp8_out0" | |
| type: CONVOLUTION | |
| bottom: "icp8_in" | |
| top: "icp8_out0" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 256 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp8_out0" | |
| type: RELU | |
| bottom: "icp8_out0" | |
| top: "icp8_out0" | |
| } | |
| layers { | |
| name: "icp8_out1" | |
| type: CONVOLUTION | |
| bottom: "icp8_reduction1" | |
| top: "icp8_out1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 320 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.04 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp8_out1" | |
| type: RELU | |
| bottom: "icp8_out1" | |
| top: "icp8_out1" | |
| } | |
| layers { | |
| name: "icp8_out2" | |
| type: CONVOLUTION | |
| bottom: "icp8_reduction2" | |
| top: "icp8_out2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 2 | |
| kernel_size: 5 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.08 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp8_out2" | |
| type: RELU | |
| bottom: "icp8_out2" | |
| top: "icp8_out2" | |
| } | |
| layers { | |
| name: "icp8_out3" | |
| type: CONVOLUTION | |
| bottom: "icp8_pool" | |
| top: "icp8_out3" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp8_out3" | |
| type: RELU | |
| bottom: "icp8_out3" | |
| top: "icp8_out3" | |
| } | |
| # Concat them together | |
| layers { | |
| name: "icp8_out" | |
| type: CONCAT | |
| bottom: "icp8_out0" | |
| bottom: "icp8_out1" | |
| bottom: "icp8_out2" | |
| bottom: "icp8_out3" | |
| top: "icp8_out" | |
| } | |
| # Inception module 9 *************** | |
| layers { | |
| name: "icp9_reduction1" | |
| type: CONVOLUTION | |
| bottom: "icp8_out" | |
| top: "icp9_reduction1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 192 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp9_reduction1" | |
| type: RELU | |
| bottom: "icp9_reduction1" | |
| top: "icp9_reduction1" | |
| } | |
| layers { | |
| name: "icp9_reduction2" | |
| type: CONVOLUTION | |
| bottom: "icp8_out" | |
| top: "icp9_reduction2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 48 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp9_reduction2" | |
| type: RELU | |
| bottom: "icp9_reduction2" | |
| top: "icp9_reduction2" | |
| } | |
| layers { | |
| name: "icp9_pool" | |
| type: POOLING | |
| bottom: "icp8_out" | |
| top: "icp9_pool" | |
| pooling_param { | |
| pool: MAX | |
| kernel_size: 3 | |
| stride: 1 | |
| pad: 1 | |
| } | |
| } | |
| # *********** | |
| layers { | |
| name: "icp9_out0" | |
| type: CONVOLUTION | |
| bottom: "icp8_out" | |
| top: "icp9_out0" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 384 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp9_out0" | |
| type: RELU | |
| bottom: "icp9_out0" | |
| top: "icp9_out0" | |
| } | |
| layers { | |
| name: "icp9_out1" | |
| type: CONVOLUTION | |
| bottom: "icp9_reduction1" | |
| top: "icp9_out1" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 384 | |
| pad: 1 | |
| kernel_size: 3 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.04 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp9_out1" | |
| type: RELU | |
| bottom: "icp9_out1" | |
| top: "icp9_out1" | |
| } | |
| layers { | |
| name: "icp9_out2" | |
| type: CONVOLUTION | |
| bottom: "icp9_reduction2" | |
| top: "icp9_out2" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 2 | |
| kernel_size: 5 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.08 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp9_out2" | |
| type: RELU | |
| bottom: "icp9_out2" | |
| top: "icp9_out2" | |
| } | |
| layers { | |
| name: "icp9_out3" | |
| type: CONVOLUTION | |
| bottom: "icp9_pool" | |
| top: "icp9_out3" | |
| blobs_lr: 1 | |
| blobs_lr: 2 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| convolution_param { | |
| num_output: 128 | |
| pad: 0 | |
| kernel_size: 1 | |
| weight_filler { | |
| type: "xavier" | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "relu_icp9_out3" | |
| type: RELU | |
| bottom: "icp9_out3" | |
| top: "icp9_out3" | |
| } | |
| # Concat them together | |
| layers { | |
| name: "icp9_out" | |
| type: CONCAT | |
| bottom: "icp9_out0" | |
| bottom: "icp9_out1" | |
| bottom: "icp9_out2" | |
| bottom: "icp9_out3" | |
| top: "icp9_out" | |
| } | |
| # third classification branch | |
| layers { | |
| name: "cls3_pool" | |
| type: POOLING | |
| bottom: "icp9_out" | |
| top: "cls3_pool" | |
| pooling_param { | |
| pool: AVE | |
| kernel_size: 7 | |
| stride: 1 | |
| pad: 0 | |
| # This padding is somewhat special | |
| } | |
| } | |
| layers { | |
| name: "cls3_drop" | |
| type: DROPOUT | |
| bottom: "cls3_pool" | |
| top: "cls3_pool" | |
| dropout_param { | |
| dropout_ratio: 0.4 | |
| } | |
| } | |
| layers { | |
| name: "cls3_fc_" | |
| type: INNER_PRODUCT | |
| bottom: "cls3_pool" | |
| top: "cls3_fc_" | |
| #blobs_lr: 1 | |
| #blobs_lr: 2 | |
| # blobs_lr is set to higher than for other layers, because this layer is starting from random while the others are already trained | |
| blobs_lr: 10 | |
| blobs_lr: 20 | |
| weight_decay: 1 | |
| weight_decay: 0 | |
| inner_product_param { | |
| num_output: 33 | |
| weight_filler { | |
| type: "gaussian" | |
| std: 0.01 | |
| } | |
| bias_filler { | |
| type: "constant" | |
| value: 0 | |
| } | |
| } | |
| } | |
| layers { | |
| name: "loss3" | |
| type: SOFTMAX_LOSS | |
| bottom: "cls3_fc_" | |
| bottom: "label" | |
| top: "loss3" | |
| loss_weight: 1 | |
| } | |
| layers { | |
| name: "accuracy1" | |
| type: ACCURACY | |
| bottom: "cls3_fc_" | |
| bottom: "label" | |
| top: "accuracy1" | |
| include: { phase: TEST } | |
| } | |
| layers { | |
| name: "accuracy5" | |
| type: ACCURACY | |
| bottom: "cls3_fc_" | |
| bottom: "label" | |
| top: "accuracy5" | |
| include: { phase: TEST } | |
| accuracy_param { | |
| top_k: 5 | |
| } | |
| } |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| # from up one dir from gist (i.e. run/tr1) | |
| boda cnet_ana --in-model=alexnet_ng_conv --in-sz=227 --print-ops=1 --print-ops-fn=alexnet_ng_conv.py && cp alexnet_ng_conv.py ex_gist/alexnet_ng_conv.txt | |
| boda cnet_ana --in-model=googlenet_ft_fl --in-sz=224 --print-ops=1 --print-ops-fn=googlenet_ft_fl.py && cp googlenet_ft_fl.py ex_gist/googlenet.txt | |
| boda cnet_ana --in-model=alexnet_ng_conv --in-sz=227 --print-ops=1 --expand-ops=1 --print-ops-fn=alexnet_ng_conv.sgemm.py && cp alexnet_ng_conv.sgemm.py ex_gist | |
| boda cnet_ana --in-model=googlenet_ft_fl --in-sz=224 --print-ops=1 --expand-ops=1 --print-ops-fn=googlenet_ft_fl.sgemm.py && cp googlenet_ft_fl.sgemm.py ex_gist |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment