Fix stride default value None in torch.nn.functional.avg_pool #4984
Conversation
python/tvm/relay/frontend/pytorch.py
Outdated
| @@ -918,7 +921,7 @@ def _get_constant(node): | |||
|
|
|||
| def _get_operator_nodes(nodes): | |||
| """ Returns torch IR nodes that need conversion to Relay """ | |||
| ops = {} | |||
| ops = [] | |||
There was a problem hiding this comment.
You should not have this change in this pr. Fetch and rebase against master.
|
Please add a test case. Add your test below. |
| if inputs[2]: | ||
| strides = _infer_shape(inputs[2]) | ||
| else: | ||
| strides = pool_size |
There was a problem hiding this comment.
is this correct? we should use the default strides
There was a problem hiding this comment.
https://github.com/pytorch/pytorch/blob/17a5c677963dc3ecb7ff505585ed15eadaaf74ef/torch/nn/functional.py#L269-L270
According to the description of strides above, for avg_pool2d. Stride's default value kernel_size, which should have the same size as pool_size here
|
Let's wait until the commit author issue is resolved https://discuss.tvm.ai/t/github-issue-the-commit-author-is-wrong-since-today/5880/ |
|
It seems the github bug was resolved, I'll try merge it. |
|
Looks like it is ok. |
…#4984) * fix unordered dictionary problem for python version 3.5 * modify style * default value of stride in torch.nn.functional.avg_pool is None * delete prev modifications * add testcase for nn.functional.avg_pool2d
…#4984) * fix unordered dictionary problem for python version 3.5 * modify style * default value of stride in torch.nn.functional.avg_pool is None * delete prev modifications * add testcase for nn.functional.avg_pool2d
The default value of stride in nn.functional.avg_pool is set as None,
https://github.com/apache/incubator-tvm/blob/5a0f39b5481a30a2eec49e27cbc17a722bd6ee6a/python/tvm/relay/frontend/pytorch.py#L461
It will cause
TVMError: Check failed: ObjectTypeChecker<TObjectRef>: :Check(ptr): Expect relay.Expr but get Arraysinceinputs[2]is empty.Fix it by detecting None type of stride in tvm frontend pytorch.