GlobalAveragePooling is converted into "Lambda"

[talcs]

Hi,

Thank you for this very friendly and beautiful tool.

I was trying to convert a Pytorch model that runs global-average-pooling after the last conv layer and before the first fc layer (In Pytorch's forward method, it is implemented by running the line x = x.mean([2,3])). Note that it is the same as done in Pytorch's official MobileNet-V2 implementation.

The problem is that after converting the Pytorch model into ONNX, keras2onnx converts the average operation into a Lambda layer. The problem with the lambda layer is that I cannot export it as a Tensorflow frozen graph.

The ONNX input file visualized in Netron - the operation is ReduceMean:

The Keras HDF5 model after importing it from ONNX using onnx2keras, visualized in Netron - the operation is Lambda with something that looks like base64 encoded content:

1. Is there a way to make onnx2keras convert this operation into Keras' GlobalAveragePooling2D layer?
2. Is it planned to implement this conversion?
3. Will onnx2keras work differently if I replace the Pytroch x = x.mean([2,3]) command by a Pytorch AdaptiveAvgPool2d layer with a 1x1 output size?

Thanks!

[talcs]

Hi,

Update: When I changed the x = x.mean([2,3]) Pytorch operation in the network's forward() method into x = self.apool(x) (where self.apool has been set in the __init__ method to be nn.AdaptiveAvgPool2d((1,1))), then:

1. When I converted it into ONNX, it has been represented as a GlobalAveragePool operation, rather than ReduceMean

2. onnx2keras has converted it to GlobalAveragePooling2D and I could export the model as a TF frozen graph!

The HDF5 Keras model visualized in Netron after importing it from Keras (with change_ordering = True):

[bourcierj]

@talcs could you close this ? Seems it was not an issue.

[pedrofrodenas]

I think I found a solution:

GlobalAveragePooling has an option to keep dims so that you can avoid using lambdas, here is the edited convert_global_avg_pool function,

def convert_global_avg_pool(node, params, layers, lambda_func, node_name, keras_name):
    """
    Convert GlobalAvgPool layer
    :param node: current operation node
    :param params: operation attributes
    :param layers: available keras layers
    :param lambda_func: function for keras Lambda layer
    :param node_name: internal converter name
    :param keras_name: resulting layer name
    :return: None
    """
    logger = logging.getLogger('onnx2keras.global_avg_pool')

    input_0 = ensure_tf_type(layers[node.input[0]], layers[list(layers)[0]], name="%s_const" % keras_name)

    global_pool = keras.layers.GlobalAveragePooling2D(data_format='channels_first', name=keras_name, keepdims=True)
    layers[node_name] = global_pool(input_0)