🔗 Helpful links

• 🧪  See artifacts and rendered test results at hud.pytorch.org/pr/72962
• 📄  Preview Python docs built from this PR
• 📄  Preview C++ docs built from this PR
• ❓Need help or want to give feedback on the CI? Visit our office hours

💊 CI failures summary and remediations

As of commit d35ce16 (more details on the Dr. CI page):

@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

The changes in the PR look good to me. But I have a more high-level question. What is the usecase we are trying to target here?

The reason here is to remove assumptions of output memory format in our lowerings which is based on knowledge of internal implementation of the ops.
E.g. for
add(lhs, rhs):
if (isNHWC(lhs) || is NHWC(rhs)) -> output memory format NHWC else NCHW

But take the output format from the outputStrides of result of tracing graph for shapes (which also traces for strides). In that case for all the lowerings we need to add outputStrides where we had outputShapes.

@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

@ZolotukhinM @navahgar this PR is a part of NHWC support.

@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

Hey @IvanKobzarev.
You've committed this PR, but it does not have both a 'release notes: ...' and 'topics: ...' label. Please add one of each to the PR. The 'release notes: ...' label should represent the part of PyTorch that this PR changes (fx, autograd, distributed, etc) and the 'topics: ...' label should represent the kind of PR it is (not user facing, new feature, bug fix, perf improvement, etc). The list of valid labels can be found here for the 'release notes: ...' and here for the 'topics: ...'.
For changes that are 'topic: not user facing' there is no need for a release notes label.

This PR is likely to have broken TorchVision. See pytorch/vision#5873 for details.

Minimal code to reproduce the issue:

import torch
from torchvision.transforms import ColorJitter


meth_kwargs = {"brightness": 0.2, "contrast": 0.2, "saturation": 0.2, "hue": 0.2}

print("-")
channels = 3
device = "cpu"

torch.random.manual_seed(0)
t = ColorJitter(**meth_kwargs)
scripted_fn = torch.jit.script(t)
batch_tensor = torch.randint(0, 256, (4, channels, 23, 34), dtype=torch.uint8, device=device)
torch.manual_seed(12)
scripted_fn(batch_tensor)

print("--")
channels = 1
device = "cpu"

torch.random.manual_seed(0)
t = ColorJitter(**meth_kwargs)
scripted_fn = torch.jit.script(t)
batch_tensor = torch.randint(0, 256, (4, channels, 23, 34), dtype=torch.uint8, device=device)
torch.manual_seed(12)
scripted_fn(batch_tensor)


print("---")
channels = 3
device = "cuda"

torch.random.manual_seed(0)
t = ColorJitter(**meth_kwargs)
scripted_fn = torch.jit.script(t)
batch_tensor = torch.randint(0, 256, (4, channels, 23, 34), dtype=torch.uint8, device=device)
torch.manual_seed(12)
scripted_fn(batch_tensor)


print("----")
channels = 1
device = "cuda"

torch.random.manual_seed(0)
t = ColorJitter(**meth_kwargs)
scripted_fn = torch.jit.script(t)
batch_tensor = torch.randint(0, 256, (4, channels, 23, 34), dtype=torch.uint8, device=device)
torch.manual_seed(12)
scripted_fn(batch_tensor)

Output:

Traceback (most recent call last):
  File "temp/check_torchscript_colorjitter_all_flakiness.py", line 51, in <module>
    scripted_fn(batch_tensor)
  File "/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py", line 1129, in _call_impl
    return forward_call(*input, **kwargs)
RuntimeError: The following operation failed in the TorchScript interpreter.
Traceback of TorchScript (most recent call last):
RuntimeError: is_tensor_creation || ((is_contiguous ^ is_channels_last_contiguous) && (is_contiguous || is_channels_last_contiguous)) INTERNAL ASSERT FAILED at "../torch/csrc/jit/tensorexpr/kernel.cpp":519, please report a bug to PyTorch. 

Surprisingly, changing order of test channels : 3,1,3,1 -> 1,3,1,3 and there is no error...

More precisely, it seems to fail on these lines :

https://github.com/pytorch/vision/blob/6d85d74be6ac72b2ac3057d85d8ae0004fa0de3f/torchvision/transforms/functional_tensor.py#L190

https://github.com/pytorch/vision/blob/6d85d74be6ac72b2ac3057d85d8ae0004fa0de3f/torchvision/transforms/functional_tensor.py#L259

@pytorchbot revert this

@IvanKobzarev @datumbox we believe this is causing errors in. torch audio so going to try to revert. If this is wrongfully reverted, please just re open and re land.

Reverting PR 72962 failed due to Can't revert PR that was landed via phabricator as D34589306
Raised by https://github.com/pytorch/pytorch/actions/runs/2221879588

@zengk95 It is causing errors on Torch Video as well, so I'm fine to have it reverted. I don't think that worked though judging from the bot's message.

@vfdev-5
Thanks for repro steps.