Export the model through Onnx

[TommasoBendinelli]

Hello,
I have successfully run the network getting the results in the result_mvae folder. I would like to export the model into other deep learning frameworks (i.e. TensorFlow). Particularly I tried to do it through onnx_chainer.export with the following:

x = np.zeros(1,7,128,128, type=np.float32)
onnx_chainer.export(model, x, filename='export.onnx')

where model is from net.Conv_MVAE

Unfortunately I get the following error:
incompatible array mixed in the forward'sd input (Concat)
Actual: <class 'numpy.ndarray'>, <class 'copy.core.core.ndarray>

I tried also to change x as cupy.zeros((1,7,128,128), type=cupy.float32) but I still get similar error

[yordanh]

Hi,

Unfortunately I don't have much experience with onnx and exporting models like this. From what I can see in your message above, some of the concats we are doing in the forward pass fail, potentially because we are concatenating cupy with bumpy arrays. Notice that in the spatial decoding step, the xy tiled 1s are defined as a cupy array.

Is the message above the full error trace and is the message different when you initialise X as a copy array or exactly the same?

[TommasoBendinelli]

Hi,
I managed to the issue by passing the following:
import onnx_chainer onnx_chainer.export(model, cp.array(train[0:1]), filename='test.onnx')

where train is the output from the DataGenerator.

Unfortunately It still raises a (new) exception:
ValueError: The all 'outputs' must be Chainer Variable