Unable to extract weights from model

[Venu Sharma]

hello,

   I was trying to extract weights from bvlc_alexnet model, specifically second convolution weights. 

nn.Sequential {

  [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> (9) -> (10) -> (11) -> (12) -> (13) -> (14) -> (15) -> (16) -> (17) -> (18) -> (19) -> (20) -> (21) -> (22) -> (23) -> (24) -> output]

  (1): cudnn.SpatialConvolution(3 -> 96, 11x11, 4,4)

  (2): cudnn.ReLU

  (3): cudnn.SpatialCrossMapLRN

  (4): cudnn.SpatialMaxPooling(3x3, 2,2)

  (5): cudnn.SpatialConvolution(96 -> 256, 5x5, 1,1, 2,2)

  (6): cudnn.ReLU

  (7): cudnn.SpatialCrossMapLRN

  (8): cudnn.SpatialMaxPooling(3x3, 2,2)

  (9): cudnn.SpatialConvolution(256 -> 384, 3x3, 1,1, 1,1)

  (10): cudnn.ReLU

  (11): cudnn.SpatialConvolution(384 -> 384, 3x3, 1,1, 1,1)

  (12): cudnn.ReLU

  (13): cudnn.SpatialConvolution(384 -> 256, 3x3, 1,1, 1,1)

  (14): cudnn.ReLU

  (15): cudnn.SpatialMaxPooling(3x3, 2,2)

  (16): nn.View(-1)

  (17): nn.Linear(9216 -> 4096)

  (18): cudnn.ReLU

  (19): nn.Dropout(0.500000)

  (20): nn.Linear(4096 -> 4096)

  (21): cudnn.ReLU

  (22): nn.Dropout(0.500000)

  (23): nn.Linear(4096 -> 1000)

  (24): cudnn.SoftMax

}

This model has groups as 2 enabled.

   .

   .

   .

  4 : 

    {

      dH : 2

      dW : 2

      kW : 3

      gradInput : CudaTensor - empty

      iSize : LongStorage - size: 4

      kH : 3

      _type : "torch.CudaTensor"

      padW : 0

      ceil_mode : true

      output : CudaTensor - empty

      padH : 0

      name : "pool1"

    }

  5 : 

    {

      padW : 2

      nInputPlane : 96

      output : CudaTensor - empty

      gradInput : CudaTensor - empty

      _type : "torch.CudaTensor"

      groups : 2

      gradBias : CudaTensor - size: 256

      dW : 1

      nOutputPlane : 256

      padH : 2

      kH : 5

      weight : CudaTensor - size: 256x48x5x5

      gradWeight : CudaTensor - size: 256x48x5x5

      bias : CudaTensor - size: 256

      dH : 1

      kW : 5

      name : "conv2"

    }

   .

   .

   .

My question isn't the weight of conv2 should be 256 x 96 x 5 x 5? It is 256 x 48 x 5 x 5 instead. Another doubt is since the input was 96 channels does that mean only on half of our input we are calculating weights? What happens to rest of the half?

Will it not create confusion while doing weights re-engineering as to where rest of the weight disappeared?

I see groups is implemented only in cudnn package(torch wrapper) whereas cunn as well as nn package doesn't have same. If I want to   convert this same model to run only on CPU how m I suppose to go about doing this.

I'm novice to torch any suggestion regarding this would be of immense help.

Regards

Venu