Unsupported ops: argmax, expand, gather about tinyneuralnetwork

[peterjc123]

@lee-man Thanks for taking it a shot. We've pushed fixes for argmax and expand. gather is taking us a bit longer, but we will make it available soon.

[lee-man]

@peterjc123 Got it. Thanks a lot!

[peterjc123]

@lee-man gather is supported now. You may have a try.

[lee-man]

Got an Error (tinynn/converter/operators/torch/aten.py, line 1568):
TypeError: create_attr_tensor() got an unexpected keyword argument 'dtype'.

[peterjc123]

@lee-man Should be fixed by 10da5c6. Please try again.

[lee-man]

Some problems with the data type for index:
File "/Users/mli/prj/pose/movenet-pytorch/TinyNeuralNetwork/tinynn/converter/base.py", line 256, in convert
self.init_operations()
File "/Users/mli/prj/pose/movenet-pytorch/TinyNeuralNetwork/tinynn/converter/base.py", line 226, in init_operations
converter.parse(node, attrs, args, self.common_graph)
File "/Users/mli/prj/pose/movenet-pytorch/TinyNeuralNetwork/tinynn/converter/operators/torch/aten.py", line 1661, in parse
self.run(node)
File "/Users/mli/prj/pose/movenet-pytorch/TinyNeuralNetwork/tinynn/converter/operators/torch/base.py", line 135, in run
raise e
File "/Users/mli/prj/pose/movenet-pytorch/TinyNeuralNetwork/tinynn/converter/operators/torch/base.py", line 129, in run
o = func(**kwargs)
RuntimeError: gather_out_cpu(): Expected dtype int64 for index

I check the current type of index, and it's int32.

[lee-man]

I tried to convert the type manually, and other errors were raised. I think it's due to the complex decoding part of the model I use. I will join the DingTalk group and discuss it with you guys directly. Thanks! @peterjc123

[peterjc123]

@lee-man Okay. I managed to fixed the type-related issues.

[peterjc123]

@lee-man I've looked into your model. It seems that you tries to call gather with the index tensor as a variable, which is unsupported at this time.

[lee-man]

@peterjc123 Got it. Really appreciate your help.

[peterjc123]

@lee-man I mean this can be fixed. We will do that later.

[lee-man]

@peterjc123 Great! Looking forward to your update!

[peterjc123]

@lee-man BTW, could you please share the TorchScript model and the shapes of the inputs so that we can look into it easier? Or, maybe we should communicate with each other via DingTalk?

[peterjc123]

Now that gather should be fully supported. Let's close this issue.