[WIP] autocolorize model

[kashif]

I am slowly converting the caffe autocolorize model to keras in the PR.

Some references:

• caffe model viz
• Project page and weights
• Caffe model and code
• Paper

[fchollet]

I think it great, do you have any update on this? What's the status?

[kashif]

thanks @fchollet currently I am stuck at translating the reshape layer from caffe to keras. I also cannot read in the original layer into caffe, due to some dim mis-match... which making it hard for me to debug and check things...

I would appreciate some help here, I was going to ask on gitter...

[fchollet]

Do post on Gitter, someone who is familiar with Caffe and Caffe conversion may be able to help out.

[fchollet]

@kashif please let me know when you think you are done.

[kashif]

@fchollet will do. Current issues with theano version:

• create weights from caffe weights
• TimeDistributed deconvolution also needs a padding (only for GPU device)
• Figure out a "better" way of doing deconvolution with padding?
• A layer giving error:
  ValueError: GpuReshape: cannot reshape input of shape (4096, 1, 2, 2) to shape (0, 4096, 120, 120). Apply node that caused the error: GpuReshape{4}(GpuElemwise{Add}[(0, 0)].0, TensorConstant{[ -1 4096.. 120 120]}) Toposort index: 516 Inputs types: [CudaNdarrayType(float32, 4D), TensorType(int64, vector)] Inputs shapes: [(4096, 1, 2, 2), (4,)] Inputs strides: [(14400, 0, 120, 1), (8,)] Inputs values: ['not shown', array([ -1, 4096, 120, 120])] Outputs clients: [[GpuIncSubtensor{InplaceSet;::, ::, int64:int64:, int64:int64:}(GpuAlloc{memset_0=True}.0, GpuReshape{4}.0, Constant{4}, Constant{124}, Constant{4}, Constant{124})]]

TensorFlow version:

• recreate weights for TF
• test on TF