change_ordering flag not working on a VGG19 pre-trained model

I am trying to convert a pre-trained VGG19 model from ONNX to Keras and then run it on my CPU (run = just predict, not train).

I managed to convert it with onnx2keras but then ran into issues with NCHW channel_first (3, 224, 224) v. NHWC channel_last (224, 224, 3). Here is the error I get when running the converted model with Keras (TensorFlow backend):

tensorflow.python.framework.errors_impl.InvalidArgumentError: Default MaxPoolingOp only supports NHWC on device type CPU

I went back to onnx2keras and realized there is an experimental flag change_ordering that seems to do what I needed!

Unfortunately, change_ordering does not seem to work with my model:

ValueError: Operands could not be broadcast together with shapes (224, 224, 3) (3, 224, 224)

Here is the full stack trace:

ValueError                                Traceback (most recent call last)
<ipython-input-1-3a0e0ae82f31> in <module>()
      8 
      9 
---> 10 keras_m = onnx_to_keras(onnx_m, ['input'], verbose=True, change_ordering=True)

~/Documents/dev/venv/lib/python3.7/site-packages/onnx2keras/converter.py in onnx_to_keras(onnx_model, input_names, input_shapes, name_policy, verbose, change_ordering)
    212
    213 
--> 214         model_tf_ordering = keras.models.Model.from_config(conf)
    215 
    216         for dst_layer, src_layer in zip(model_tf_ordering.layers,

~/Documents/dev/venv/lib/python3.7/site-packages/keras/engine/network.py in from_config(cls, config, custom_objects)
   1030                 if layer in unprocessed_nodes:
   1031                     for node_data in unprocessed_nodes.pop(layer):
-> 1032                         process_node(layer, node_data)
   1033 
   1034         name = config.get('name')

~/Documents/dev/venv/lib/python3.7/site-packages/keras/engine/network.py in process_node(layer, node_data)
    989             # and building the layer if needed.
    990             if input_tensors:
--> 991                 layer(unpack_singleton(input_tensors), **kwargs)
    992 
    993         def process_layer(layer_data):

~/Documents/dev/venv/lib/python3.7/site-packages/keras/engine/base_layer.py in __call__(self, inputs, **kwargs)
    429                                          'You can build it manually via: '
    430                                          '`layer.build(batch_input_shape)`')
--> 431                 self.build(unpack_singleton(input_shapes))
    432                 self.built = True
    433 

~/Documents/dev/venv/lib/python3.7/site-packages/keras/layers/merge.py in build(self, input_shape)
    254 
    255     def build(self, input_shape):
--> 256         super(Subtract, self).build(input_shape)
    257         if len(input_shape) != 2:
    258             raise ValueError('A `Subtract` layer should be called '

~/Documents/dev/venv/lib/python3.7/site-packages/keras/layers/merge.py in build(self, input_shape)
     89                 shape = input_shape[i][1:]
     90             output_shape = self._compute_elemwise_op_output_shape(output_shape,
---> 91                                                                   shape)
     92         # If the inputs have different ranks, we have to reshape them
     93         # to make them broadcastable.

~/Documents/dev/venv/lib/python3.7/site-packages/keras/layers/merge.py in _compute_elemwise_op_output_shape(self, shape1, shape2)
     59                     raise ValueError('Operands could not be broadcast '
     60                                      'together with shapes ' +
---> 61                                      str(shape1) + ' ' + str(shape2))
     62                 output_shape.append(i)
     63         return tuple(output_shape)

ValueError: Operands could not be broadcast together with shapes (224, 224, 3) (3, 224, 224)

I am using python3.7 and (all from pip): onnx==1.5.0, onnx2keras==0.0.4, Keras==2.2.4, and tensorflow==1.14.0.

Here is the code that gets me the stack trace above (on a Jupyter Notebook):

import onnx
from onnx2keras import onnx_to_keras
onnx_m = onnx.load('VGG19_PRETRAINED.onnx')
keras_m = onnx_to_keras(onnx_m, ['input'], verbose=True, change_ordering=True)

@nerox8664, thanks a lot for this (very useful!) library. I do appreciate any hints you might have about my issue.

[gmalivenko]

The model contains constants (e.g. for input normalization) which were not permuted well. Now the converter (0.0.9) is able to permute constant shapes, and finally I’m able to convert your ONNX model.

Snippet

[lucasrla]

Thank you so much, Grigory. It is now working perfectly.