Unknown layer type: table

[yuvalaviel]

Hi,

Trying to convert deepdesc .t7 net from https://github.com/etrulls/deepdesc-release
After countless efforts, I ran into an issue I couldn't solve:
Apparently the network model has a layer type "table" which is not implemented in Caffe?
Unfortunately, I'm not familiar with Torch, so I can't inspect this 'table' layer.

The model is here:
https://github.com/etrulls/deepdesc-release/blob/master/models/CNN3_p8_n8_split4_073000.t7

Any help will be appreciated.
Working on Ubuntu 16.04.

[yuvalaviel]

Actually, just printing the layers in Torch, I can't see any 'table' layer:

th> require 'nn'
th> data = torch.load( 'models/CNN3_p8_n8_split4_073000.t7' )
th> print(data.desc)
nn.Sequential {
[input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> (9) -> (10) -> (11) -> (12) -> output]
(1): nn.SpatialConvolution(1 -> 32, 7x7)
(2): nn.Tanh
(3): nn.Sequential {
[input -> (1) -> (2) -> (3) -> (4) -> output]
(1): nn.Square
(2): nn.SpatialAveragePooling(2x2, 2,2)
(3): nn.MulConstant
(4): nn.Sqrt
}
(4): nn.SpatialSubtractiveNormalization
(5): nn.SpatialConvolution(32 -> 64, 6x6)
(6): nn.Tanh
(7): nn.Sequential {
[input -> (1) -> (2) -> (3) -> (4) -> output]
(1): nn.Square
(2): nn.SpatialAveragePooling(3x3, 3,3)
(3): nn.MulConstant
(4): nn.Sqrt
}
(8): nn.SpatialSubtractiveNormalization
(9): nn.SpatialConvolution(64 -> 128, 5x5)
(10): nn.Tanh
(11): nn.Sequential {
[input -> (1) -> (2) -> (3) -> (4) -> output]
(1): nn.Square
(2): nn.SpatialAveragePooling(4x4, 4,4)
(3): nn.MulConstant
(4): nn.Sqrt
}
(12): nn.Reshape(128)
}

[ptriantd]

@yuvalaviel I am having a similar issue. Did you manage to solve it eventually?

My output is:

th> print(net_fast)
{
  1 : 
    {
      tmp_in : CudaTensor - empty
      tmp_out : CudaTensor - empty
      fgradInput : CudaTensor - empty
      output : CudaTensor - empty
      modules : 
        {
          1 : 
            {
              padW : 1
              nInputPlane : 1
              output : CudaTensor - empty
              _type : "torch.CudaTensor"
              shared : true
              dH : 1
              dW : 1
              nOutputPlane : 64
              padH : 1
              kH : 3
              finput : CudaTensor - empty
              weight : CudaTensor - size: 64x1x3x3
              tmp_out : CudaTensor - empty
              kW : 3
              bias : CudaTensor - size: 64
              groups : 1
              fgradInput : CudaTensor - empty
              tmp_in : CudaTensor - empty
            }
          2 : 
            {
              inplace : true
              tmp_out : CudaTensor - empty
              output : CudaTensor - empty
              _type : "torch.CudaTensor"
              tmp_in : CudaTensor - empty
              finput : CudaTensor - empty
              fgradInput : CudaTensor - empty
            }
          3 : 
            {
              padW : 1
              nInputPlane : 64
              output : CudaTensor - empty
              _type : "torch.CudaTensor"
              shared : true
              dH : 1
              dW : 1
              nOutputPlane : 64
              padH : 1
              kH : 3
              finput : CudaTensor - empty
              weight : CudaTensor - size: 64x64x3x3
              tmp_out : CudaTensor - empty
              kW : 3
              bias : CudaTensor - size: 64
              groups : 1
              fgradInput : CudaTensor - empty
              tmp_in : CudaTensor - empty
            }
          4 : 
            {
              inplace : true
              tmp_out : CudaTensor - empty
              output : CudaTensor - empty
              _type : "torch.CudaTensor"
              tmp_in : CudaTensor - empty
              finput : CudaTensor - empty
              fgradInput : CudaTensor - empty
            }
          5 : 
            {
              padW : 1
              nInputPlane : 64
              output : CudaTensor - empty
              _type : "torch.CudaTensor"
              shared : true
              dH : 1
              dW : 1
              nOutputPlane : 64
              padH : 1
              kH : 3
              finput : CudaTensor - empty
              weight : CudaTensor - size: 64x64x3x3
              tmp_out : CudaTensor - empty
              kW : 3
              bias : CudaTensor - size: 64
              groups : 1
              fgradInput : CudaTensor - empty
              tmp_in : CudaTensor - empty
            }
          6 : 
            {
              inplace : true
              tmp_out : CudaTensor - empty
              output : CudaTensor - empty
              _type : "torch.CudaTensor"
              tmp_in : CudaTensor - empty
              finput : CudaTensor - empty
              fgradInput : CudaTensor - empty
            }
          7 : 
            {
              padW : 1
              nInputPlane : 64
              output : CudaTensor - empty
              _type : "torch.CudaTensor"
              shared : true
              dH : 1
              dW : 1
              nOutputPlane : 64
              padH : 1
              kH : 3
              finput : CudaTensor - empty
              weight : CudaTensor - size: 64x64x3x3
              tmp_out : CudaTensor - empty
              kW : 3
              bias : CudaTensor - size: 64
              groups : 1
              fgradInput : CudaTensor - empty
              tmp_in : CudaTensor - empty
            }
          8 : 
            {
              tmp_in : CudaTensor - empty
              tmp_out : CudaTensor - empty
              norm : CudaTensor - empty
              output : CudaTensor - empty
              fgradInput : CudaTensor - empty
              finput : CudaTensor - empty
              _type : "torch.CudaTensor"
            }
          9 : 
            {
              tmp_in : CudaTensor - empty
              fgradInput : CudaTensor - empty
              output_L : CudaTensor - empty
              finput : CudaTensor - empty
              direction : -1
              disp_max : 1
              output : CudaTensor - empty
              _type : "torch.CudaTensor"
              tmp_out : CudaTensor - empty
            }
        }
      finput : CudaTensor - empty
      _type : "torch.CudaTensor"
    }
  2 : 
    {
      tiny : false
      m : 0.2
      d_contrast : 1
      hflip : 0
      d_hscale : 1
      subset : 1
      a : "train_all"
      sm_terminate : ""
      tau1 : 0
      net_fname : ""
      sgm_q2 : 2.5
      l1 : 4
      gpu : 1
      cbca_i2 : 0
      lr : 0.0002
      d_hshear : 0
      d_vtrans : 0
      L1 : 0
      pi1 : 4
      blur_sigma : 7.74
      print_args : false
      left : ""
      disp_max : ""
      false1 : 4
      contrast : 1.3
      use_cache : false
      mom : 0.9
      tau_so : 0.02
      hshear : 0.1
      pi2 : 55.72
      make_cache : false
      rotate : 7
      d_rotate : 0
      false2 : 10
      right : ""
      d : "kitti | mb"
      sgm_i : 1
      sm_skip : ""
      ks : 3
      rect : "imperfect"
      true1 : 1
      debug : false
      blur_t : 5
      fm : 64
      sgm_q1 : 3
      pow : 1
      scale : 1
      alpha1 : 1.5
      hscale : 0.9
      cbca_i1 : 0
      at : 0
      trans : 0
      bs : 128
      color : "gray"
      seed : 42
      vflip : 0
      brightness : 0.7
      d_brightness : 0.3
    }
}

[yuvalaviel]

No, sorry.

[vrbala]

@yuvalaviel @ptriantd I encountered the same issue yesterday. My observation is, when torch.load('somemodel.t7') is called, all that happens is the model parameters are getting loaded and not the actual torch definition of the model (so all you get is a lua table with the parameters). What you should actually do is to create a model definition (equivalent to caffe prototxt) for the model, retrieve the state of it using the loaded snapshot (from torch.load invocation) and then use it in torch2caffe.

Eg:

       print('Loading model...')
        local modelDir = '../model/crnn_demo/'
        paths.dofile(paths.concat(modelDir, 'config.lua'))
        local modelLoadPath = paths.concat(modelDir, 'crnn_demo_model.t7')
        gConfig = getConfig()
        gConfig.modelDir = modelDir
        gConfig.maxT = 0
        model, criterion = createModel(gConfig)
        local snapshot = torch.load(modelLoadPath)
        loadModelState(model, snapshot)
        model:evaluate()
        print(string.format('Model loaded from %s', modelLoadPath))
        assert(model)

I took the above code from https://github.com/bgshih/crnn/blob/master/src/demo.lua#L23 as I am trying to convert the crnn torch model to caffe. The above block should replace https://github.com/facebook/fb-caffe-exts/blob/master/torch2caffe/lib.lua#L185.

[yuvalaviel]

thanks, @vrbala .
Please keep us posted if you succeed.

[dabenzhong-wang]

@vrbala Hello! I met with this error too when I convert CRNN model to caffe. If you've solved it?

[vrbala]

@dabenzhong-wang Hello, I didn't spend too much time resolving this to the fullest because CRNN had too many custom layers that didn't have straight forward caffe equivalents. I didn't want to invest time and effort writing those caffe equivalents because my need was very preliminary and ended up using torch implementation of CRNN itself.

[dabenzhong-wang]

@vrbala I see，thanks~

[stalagmite7]

@dabenzhong-wang , have you found any success in converting the CRNN to Caffe? I am working on it right now! Would love some help!