gistfile1.py

from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation, Flatten
from keras.layers.convolutional import Convolution2D, MaxPooling2D
from keras.layers.normalization import BatchNormalization

#AlexNet with batch normalization in Keras 
#input image is 224x224

model = Sequential()
model.add(Convolution2D(64, 3, 11, 11, border_mode='full'))
model.add(BatchNormalization((64,226,226)))
model.add(Activation('relu'))
model.add(MaxPooling2D(poolsize=(3, 3)))

model.add(Convolution2D(128, 64, 7, 7, border_mode='full'))
model.add(BatchNormalization((128,115,115)))
model.add(Activation('relu'))
model.add(MaxPooling2D(poolsize=(3, 3)))

model.add(Convolution2D(192, 128, 3, 3, border_mode='full'))
model.add(BatchNormalization((128,112,112)))
model.add(Activation('relu'))
model.add(MaxPooling2D(poolsize=(3, 3)))

model.add(Convolution2D(256, 192, 3, 3, border_mode='full'))
model.add(BatchNormalization((128,108,108)))
model.add(Activation('relu'))
model.add(MaxPooling2D(poolsize=(3, 3)))

model.add(Flatten())
model.add(Dense(12*12*256, 4096, init='normal'))
model.add(BatchNormalization(4096))
model.add(Activation('relu'))
model.add(Dense(4096, 4096, init='normal'))
model.add(BatchNormalization(4096))
model.add(Activation('relu'))
model.add(Dense(4096, 1000, init='normal'))
model.add(BatchNormalization(1000))
model.add(Activation('softmax'))

srds_edge.png

I believe alexnet has two streams, and cannot be implemented with a sequential model and it must be implemented with the functional api. I only see one stream here.

@JonathanCMitchell - Possible because there are two variants of alexnet. The original one has two streams, but the caffenet version is a single stream. I think this is the caffenet version!

Is there some example codes for using this?
like input = mired()...
model.predict(input)...

Hi guys
I want to use alexnet for feature extraction.I was wondering if could tell me how to feed my as image into alexnet?

alexnet uses overlapping pooling, the first conv layer's pooling should use (3, 3) kernel with stride 2 according to the original paper.

model.add(MaxPooling2D(poolsize=(3, 3), strides=2))

why your Convolution2D have four customized parameters? currently version's keras only need 3 params..??

This link is awesome but its not in keras but tensorflow..
Here's the github link..

http://www.cs.toronto.edu/~guerzhoy/tf_alexnet/myalexnet_forward_newtf.py

You can find 2-stream AlexNet here: http://dandxy89.github.io/ImageModels/alexnet/