siavashk / imagenet-autoencoder Goto Github PK

Please see the attached screenshot and focus on the things in the red rectangle. The left one are the weights of decoder filters in the last layer. The right one are the weights of encoder filters in the first layer. They get updated every epoch. I added below display code to https://github.com/siavashk/imagenet-autoencoder/blob/master/scripts/train.lua#L44:

eweight = ae.net.modules[1].weight
dweight = ae.net.modules[19].weight
eweight_view = eweight:float():view(12, 3, 5, 5)
dweight_view = dweight:float():view(12, 3, 5, 5)
de = image.toDisplayTensor{input=eweight_view,
padding=2,
nrow=math.floor(math.sqrt(12)),
symmetric=true}
dd = image.toDisplayTensor{input=dweight_view,
padding=2,
nrow=math.floor(math.sqrt(12)),
symmetric=true}
win1 = image.display{image=dd, win=win1, legend='Decoder filters', zoom=2}
win2 = image.display{image=de, win=win2, legend='Encoder filters', zoom=2}

I change kernel size to 5 so we can get a clearer looking. You can find the decoder layer learned something meaningful, but the encoder layer learned just noise. If I stop the training at this time and test the net. I am pretty sure the learned image would be very similar with the original one according to my previous experiments. Can you give me a little comment on this? Is the encoder part normal?

Hi @siavashk
I have a question, please see line #28:
https://github.com/siavashk/imagenet-autoencoder/blob/master/core/autoencoder.lua#L28

self.net:add(nn.SpatialConvolution(12, 12, 3, 3, 1, 1, 1, 1))

You see, 12 in and 12 out. So we have the same number of the input and the output, what effect will this layer actually do?

Thanks,
Yingjun

self.net:add(nn.Reshape(24 * 16 * 16))
self.net:add(nn.Linear(24 * 16 * 16, 1568))
self.net:add(nn.Linear(1568, 24 * 16 * 16))
self.net:add(nn.Reshape(24, 16, 16))

I removed them and seems network can still be trained.

Hi Siavash,
I made some changes to imagenet-autoencoder and trained it. After some rounds, I get 0.003 for training error and 0.013 for validation error. I feel the gap between them is big, is the autoencoder in overfitting status?