search

apache / tvm

Open deep learning compiler stack for cpu, gpu and specialized accelerators
https://tvm.apache.org/
Apache License 2.0
11.69k stars 3.46k forks source link

The following operators are not implemented: ['aten::max_unpool2d'] in pytorch frontend #6761

Closed IversonTang closed 3 years ago

IversonTang commented 3 years ago

Hello: I encountered a problem the following operators are not implemented: ['aten::max_unpool2d'] when deploying segnet model using pytorch frontend. Then I observed the operator aten:: max pool2d with_ indicies seems return dummy indices in relay/frontend/pytorch.py.How can I modify operator max pool2d with_ indicies to save indicies and add new operator max_unpool2d?

masahi commented 3 years ago

Unfortunately, upsampling via unpooling is not possible with TVM, since TVM max pool op doesn't record indices (that's why we return dummy indices from max_ pool2d_ with_ indicies). If possible, you can modify your model to do upsampling via torch.interpolate(...) function.

For further questions, please use the forum instead.