failed to convert official ResNet models

[shuangwu]

when converting models downloaded from https://github.com/KaimingHe/deep-residual-networks#models, ended up with error:

Traceback (most recent call last):
  File "./convert.py", line 60, in <module>
    main()
  File "./convert.py", line 56, in main
    args.phase)
  File "./convert.py", line 27, in convert
    transformer = TensorFlowTransformer(def_path, caffemodel_path, phase=phase)
  File "/Users/shuang/github/caffe-tensorflow/kaffe/tensorflow/transformer.py", line 227, in __init__
    self.load(def_path, data_path, phase)
  File "/Users/shuang/github/caffe-tensorflow/kaffe/tensorflow/transformer.py", line 237, in load
    graph = DataInjector(def_path, data_path)(graph)
  File "/Users/shuang/github/caffe-tensorflow/kaffe/transformers.py", line 86, in __call__
    node.data = self.adjust_parameters(node, data)
  File "/Users/shuang/github/caffe-tensorflow/kaffe/transformers.py", line 79, in adjust_parameters
    data[idx] = np.squeeze(data[idx])
IndexError: list index out of range

If saving only source code, ended up with error:

Traceback (most recent call last):
  File "./convert.py", line 60, in <module>
    main()
  File "./convert.py", line 56, in main
    args.phase)
  File "./convert.py", line 37, in convert
    src_out.write(transformer.transform_source())
  File "/Users/shuang/github/caffe-tensorflow/kaffe/tensorflow/transformer.py", line 288, in transform_source
    chains = mapper.map()
  File "/Users/shuang/github/caffe-tensorflow/kaffe/graph.py", line 288, in map
    mapped_chains.append(self.map_chain(chain))
  File "/Users/shuang/github/caffe-tensorflow/kaffe/graph.py", line 292, in map_chain
    return [self.map_node(node) for node in chain]
  File "/Users/shuang/github/caffe-tensorflow/kaffe/graph.py", line 296, in map_node
    mapped_node = map_func(node)
  File "/Users/shuang/github/caffe-tensorflow/kaffe/tensorflow/transformer.py", line 155, in map_batch_norm
    scale_offset = len(node.data) == 4
TypeError: object of type 'NoneType' has no len()

My fix is around line 155 of transformer.py:

        if node.data:
            scale_offset = len(node.data) == 4
        else:
            scale_offset = False

but the first error remains.

[shuangwu]

this may be related to #63

[shuangwu]

FYI, this is using protobuf backend, not caffe backend.

[j50888]

I have the same problem How to modify the code to solve it?

[Max-Jesch]

Same issues here. Anybody found a solution?

[Max-Jesch]

I got it to work by replacing scale_offset= len(node.data) == 4 with scale_offset = len(node.output_shape) == 4 in transformer.py I don't know if this breaks something else, but it definitely makes the ResNet50 and ResNet101 work.

[zhyfzy]

I solved this problem by installing pycaffe.

[eylon-stroh]

To address the first issue, try changing

        squeeze_indices = [1]  # Squeeze biases.
        if node.kind == NodeKind.InnerProduct:
            squeeze_indices.append(0)  # Squeeze FC.

to

        squeeze_indices = []
        if node.kind == NodeKind.Convolution or node.kind == NodeKind.InnerProduct:
            if node.parameters.bias_term:
                squeeze_indices.append(1)
            if node.kind == NodeKind.InnerProduct:
                squeeze_indices.append(0)  # Squeeze FC.

in transformers.py. This fixes the "index out of range" issue for convolutions without a bias term, as well as an issue where BatchNorm variance was getting squeezed incorrectly.

[xuanyuansen]

scale_offset = len(node.output_shape) == 4

this

I got it to work by replacing scale_offset= len(node.data) == 4 with scale_offset = len(node.output_shape) == 4 in transformer.py I don't know if this breaks something else, but it definitely makes the ResNet50 and ResNet101 work.

thanks, it fix the error.