torch_alexnet.md

AlexNet trained in Torch7

Trained by me using https://github.com/soumith/imagenet-multiGPU.torch, achieves 56.7% top1 center crop accuracy on ImageNet validation set. Tested here: https://github.com/szagoruyko/imagenet-validation.torch

Download link: https://yadi.sk/d/uX6id_yZoU8FC (247 MB)

Load as:

net = torch.load'./alexnet_torch.t7':unpack()

Input image size is 227.

Preprocessing

Separate mean std per channel is saved with the network:

> print(net.transform)
{
  mean :
    {
      1 : 0.48462227599918
      2 : 0.45624044862054
      3 : 0.40588363755159
    }
  std :
    {
      1 : 0.22889466674951
      2 : 0.22446679341259
      3 : 0.22495548344775
    }
}

Can be loaded without CUDA support too.

Printout:

nn.Sequential {
  (1): nn.Concat {
    input
      |`-> (1): nn.Sequential {
      |      (1): nn.SpatialConvolution(3 -> 48, 11x11, 4,4, 2,2)
      |      (2): nn.SpatialBatchNormalization
      |      (3): nn.ReLU
      |      (4): nn.SpatialMaxPooling(3,3,2,2)
      |      (5): nn.SpatialConvolution(48 -> 128, 5x5, 1,1, 2,2)
      |      (6): nn.SpatialBatchNormalization
      |      (7): nn.ReLU
      |      (8): nn.SpatialMaxPooling(3,3,2,2)
      |      (9): nn.SpatialConvolution(128 -> 192, 3x3, 1,1, 1,1)
      |      (10): nn.SpatialBatchNormalization
      |      (11): nn.ReLU
      |      (12): nn.SpatialConvolution(192 -> 192, 3x3, 1,1, 1,1)
      |      (13): nn.SpatialBatchNormalization
      |      (14): nn.ReLU
      |      (15): nn.SpatialConvolution(192 -> 128, 3x3, 1,1, 1,1)
      |      (16): nn.SpatialBatchNormalization
      |      (17): nn.ReLU
      |      (18): nn.SpatialMaxPooling(3,3,2,2)
      |    }
      |`-> (2): nn.Sequential {
             (1): nn.SpatialConvolution(3 -> 48, 11x11, 4,4, 2,2)
             (2): nn.SpatialBatchNormalization
             (3): nn.ReLU
             (4): nn.SpatialMaxPooling(3,3,2,2)
             (5): nn.SpatialConvolution(48 -> 128, 5x5, 1,1, 2,2)
             (6): nn.SpatialBatchNormalization
             (7): nn.ReLU
             (8): nn.SpatialMaxPooling(3,3,2,2)
             (9): nn.SpatialConvolution(128 -> 192, 3x3, 1,1, 1,1)
             (10): nn.SpatialBatchNormalization
             (11): nn.ReLU
             (12): nn.SpatialConvolution(192 -> 192, 3x3, 1,1, 1,1)
             (13): nn.SpatialBatchNormalization
             (14): nn.ReLU
             (15): nn.SpatialConvolution(192 -> 128, 3x3, 1,1, 1,1)
             (16): nn.SpatialBatchNormalization
             (17): nn.ReLU
             (18): nn.SpatialMaxPooling(3,3,2,2)
           }
       ... -> output
  }
  (2): nn.Sequential {
    (1): nn.View(9216)
    (2): nn.Dropout(0.500000)
    (3): nn.Linear(9216 -> 4096)
    (4): nn.BatchNormalization
    (5): nn.Threshold
    (6): nn.Dropout(0.500000)
    (7): nn.Linear(4096 -> 4096)
    (8): nn.BatchNormalization
    (9): nn.Threshold
    (10): nn.Linear(4096 -> 1000)
  }
}