Fail when converting the ReduceMax op type. Perhaps not considering the torch.max's return value

[maybeLee]

I was trying to convert a Keras model to PyTorch through ONNX but failed.
The information of my targeted model is as follows

_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
input_1 (InputLayer)         [(None, 32, 3)]           0
_________________________________________________________________
global_max_pooling1d         (None, 3)                 0
_________________________________________________________________
dense (Dense)                (None, 96)                384
_________________________________________________________________
reshape (Reshape)            (None, 32, 3)             0
=================================================================
Total params: 384
Trainable params: 384
Non-trainable params: 0
_________________________________________________________________

The script to reproduce is as follows:

# Build the model
import tensorflow as tf
import tf2onnx
x = tf.keras.layers.Input((32,3))
in1 = tf.keras.layers.GlobalMaxPooling1D()(x)
in2 = tf.keras.layers.Dense(96)(in1)
y = tf.keras.layers.Reshape((32,3))(in2)
model = tf.keras.Model(x, y)
model.summary()

# Convert the model
input_shape = model.layers[0].input_shape[0]
spec = (tf.TensorSpec(input_shape, tf.float32, name="input"),)
_, _ = tf2onnx.convert.from_keras(model, input_signature=spec, \
        opset=15, output_path="temp.onnx")
from onnx2pytorch import ConvertModel
import onnx
onnx_model = onnx.load("temp.onnx")
torch_model = ConvertModel(onnx_model, experimental=True)

# Predict
import torch
import numpy as np
input = np.random.rand(10, *input_shape[1:])
input = torch.from_numpy(input)
torch_model.double()
pred = torch_model(input)
pred = pred.detach().numpy()
print("The prediction is: ", pred.shape)

You may access the code here:
https://colab.research.google.com/drive/1EtuxhHjy3QdmCf4v6DSpN9jsde2SeNpW?usp=sharing

The crash information is as follows:

/usr/local/lib/python3.7/dist-packages/torch/nn/functional.py in linear(input, weight, bias)
   1846     if has_torch_function_variadic(input, weight, bias):
   1847         return handle_torch_function(linear, (input, weight, bias), input, weight, bias=bias)
-> 1848     return torch._C._nn.linear(input, weight, bias)
   1849 
   1850 

TypeError: linear(): argument 'input' (position 1) must be Tensor, not torch.return_types.max

Without very deep investigation, I assume this problem is caused by torch.max()'s output, which is torch.return_types.max instead of a torch tensor, while the linear layer expect the input to be a tensor.

I guess the fix of this bug would be change torch.max(**kwargs) to torch.max(**kwargs)[0] to make the output data type to be tensor. But I am new to this project and don't know how to write some fixable codes. Can you check if this is actually a bug and how can we fix it?

[calvinmccarter-at-lightmatter]

@maybeLee -- PR #37 should fix this for you.

[maybeLee]

Thanks for your contribution