I don't understand which command I have to write to convert a model about keras

[MarcBS]

As it explained in the README file, you can use the caffe2keras.py as a command line tool the following way:

python caffe2keras.py -load_path <your_path> -prototxt <your_prototxt> -caffemodel <your_caffemodel>

where you just have to include the path to your model files (your_path) and the names of the prototxt (your_prototxt) and caffemodel (your_caffemodel) files included in that path.

The converted model will be stored by default in your_path.

I have just updated the README file for making this more clear.

[cyrilgarcia009]

You mean I have to write this command line in the Terminal ?

[cyrilgarcia009]

When I use this line in the terminal there is this error :

google.protobuf.text_format.ParseError: 37:2 : Message type "caffe.LayerParameter" has no field named "scale_param".

Do you understand why ?

Thank you very much for your help !

[MarcBS]

This error is probably caused because the parameter 'scale_param' is relatively new in the .prototxt files. I have just updated the file caffe_pb2.py in the repo, which is in charge of controlling the valid .prototxt parameters.

Re-sync your cloned files with the new changes and try it again, it should work properly now.

[cyrilgarcia009]

Hi,

I am really sorry but it still doesn't work. I'm getting this error even after the sync :

Traceback (most recent call last):
File "/Users/cyrilgarcia/keras/keras/caffe/caffe2keras.py", line 45, in
main(args)
File "/Users/cyrilgarcia/keras/keras/caffe/caffe2keras.py", line 34, in main
model = convert.caffe_to_keras(args.load_path+'/'+args.prototxt, args.load_path+'/'+args.caffemodel, debug=args.debug)
File "/Users/cyrilgarcia/keras/keras/caffe/convert.py", line 43, in caffe_to_keras
debug)
File "/Users/cyrilgarcia/keras/keras/caffe/convert.py", line 255, in create_model
raise RuntimeError('layer type', type_of_layer, 'used in this model is not currently supported')
RuntimeError: ('layer type', 'batchnorm', 'used in this model is not currently supported')
MacBook-Air-de-Cyril:~ cyrilgarcia$

May be this is because I don't know how to use caffe_pb2. py ?

[MarcBS]

That error is thrown because the layer BatchNorm was not implemented in the converter yet.
I have just added it now. Try again and tell me if it worked.

[cyrilgarcia009]

Still the same type of error :

Traceback (most recent call last):
File "/Users/cyrilgarcia/keras/keras/caffe/caffe2keras.py", line 45, in
main(args)
File "/Users/cyrilgarcia/keras/keras/caffe/caffe2keras.py", line 34, in main
model = convert.caffe_to_keras(args.load_path+'/'+args.prototxt, args.load_path+'/'+args.caffemodel, debug=args.debug)
File "/Users/cyrilgarcia/keras/keras/caffe/convert.py", line 43, in caffe_to_keras
debug)
File "/Users/cyrilgarcia/keras/keras/caffe/convert.py", line 260, in create_model
raise RuntimeError('layer type', type_of_layer, 'used in this model is not currently supported')
RuntimeError: ('layer type', 'scale', 'used in this model is not currently supported')

[ternaus]

Same error as @cyrilgarcia009 .

To add more context: I am trying to convert Resnet50 from https://github.com/KaimingHe/deep-residual-networks. Link to proto and weights: https://onedrive.live.com/?authkey=%21AAFW2-FVoxeVRck&id=4006CBB8476FF777%2117887&cid=4006CBB8476FF777

[MarcBS]

Thanks for the context @ternaus.

I have applied some new changes to the converter and ResNets can be correctly transferred to Keras. As an example you can use the weights and prototxt provided by @ternaus.

Just as a reminder, don't forget to add the following layer as the first layer in your prototxt files:

layer {
name: "data"
type: "Data"
top: "data"
top: "label"
}

I have not checked the performance of the converted model, so please, if any of you apply some validation report the results comparison here for considering any possible modification. Thanks.

[EpochalEngineer]

I was able to convert ResNet50 with the current commit and adding the data layer, but wasn't able to get inference working.

Exception: You are attempting to share a same BatchNormalization layer across different data flows. This is not possible. You should use mode=2 in BatchNormalization, which has a similar behavior but is shareable (see docs for a description of the behavior).

https://github.com/MarcBS/keras/blob/master/keras/layers/normalization.py#L193

Some debugging info that doesn't quite make sense:

ipdb> self.name
u'bn_conv1'
ipdb> self.called_with
Elemwise{add,no_inplace}.0
ipdb> x
Elemwise{add,no_inplace}.0
ipdb> x.name
ipdb>
ipdb> x.get_parents()[0]
Elemwise{add,no_inplace}(AbstractConv2d{border_mode='valid', subsample=(2, 2), filter_flip=True, imshp=(None, None, None, None), kshp=(64, 3, 7, 7)}.0, Reshape{4}.0)
ipdb> self.called_with.get_parents()[0]
Elemwise{add,no_inplace}(AbstractConv2d{border_mode='valid', subsample=(2, 2), filter_flip=True, imshp=(None, None, None, None), kshp=(64, 3, 7, 7)}.0, Reshape{4}.0)

I'm not all that familiar with keras/theano, but bn_conv1 has one input and output node and it's conv1 and caffe's separate batchnorm scale layer. It's not anywhere near a skip connection or addition (I searched the json and yaml files). Not sure how to get more information as to how x and self.called_with differ.

Here is a graphviz of the converted model:
https://i.imgur.com/PfMXhDw.png

Assuming it was accidentally getting recompiled , I removed the explicit compile step and got this error (same as if replacing mode=0 with mode=2 in the model files):

ValueError: GpuReshape: cannot reshape input of shape (2048) to shape (1, 1, 1, 7).
Apply node that caused the error: GpuReshape{4}(scale5a_branch1_gamma, TensorConstant{[1 1 1 7]})
Toposort index: 447
Inputs types: [CudaNdarrayType(float32, vector), TensorType(int64, vector)]
Inputs shapes: [(2048,), (4,)]
Inputs strides: [(1,), (8,)]
Inputs values: ['not shown', array([1, 1, 1, 7])]
Outputs clients: [[GpuElemwise{Composite{(((i0 * (i1 + i2)) + i3) + ((i4 * i5) + i6))},no_inplace}(GpuReshape{4}.0, GpuElemwise{Composite{(i0 * (i1 / i2))},no_inplace}.0, GpuReshape{4}.0, GpuReshape{4}.0, GpuReshape{4}.0, GpuElemwise{Composite{((i0 * (i1 / i2)) + i3)},no_inplace}.0, GpuReshape{4}.0)]]

only change to test_converted.py was replacing the 3 losses with one when compiling.

model.summary() looks reasonable around that area.