GwiHwan-Go opened a new issue, #14927:
URL: https://github.com/apache/tvm/issues/14927
Description
TVM's AvgPool2d function exhibits a bug when the stride is larger than the
kernel size. The function produces incorrect values, which are not consistent
with the expected output as verified by PyTorch. This bug surfaces when certain
configurations of stride and kernel size are used.
Notably, the following argument settings trigger the issue:
{'kernel_size': 6, 'stride': 10, 'padding': 0, 'ceil_mode': True,
'count_include_pad': False, 'divisor_override': None}
{'kernel_size': 4, 'stride': 9, 'padding': 0, 'ceil_mode': True,
'count_include_pad': False, 'divisor_override': None}
{'kernel_size': 7, 'stride': 7, 'padding': 2, 'ceil_mode': True,
'count_include_pad': False, 'divisor_override': None}
Interestingly, when {'kernel_size': 7, 'stride': 7, 'padding': 1} is used,
the function does not produce incorrect results.
The input information for the AvgPool2d function across all scenarios is:
input_size = [2, 3, 18, 18], input_channels_first = True,
input_same_width_height = False, input_data_type = torch.float32.
### Expected behavior
When using the AvgPool2d function in TVM, the output should match the output
from PyTorch, regardless of the values of stride and kernel size.
### Actual behavior
When the stride is larger than the kernel size, the function should
correctly handle it and produce results consistent with PyTorch. Additionally,
the output tensor shapes also differ, pointing to an issue with the function's
implementation. Furthermore, the divisor_override attribute is reportedly
always causing incorrect results according to issue #14795.
### Environment
OS : Linux 5.4.0-148-generic
Python : 3.9.16
PyTorch : 2.0.0
TVM : 0.11.1
Any environment details, such as: Operating System, TVM version, etc
### Steps to reproduce
```python
import torch
from tvm import relay
import tvm
import numpy as np
m = torch.nn.AvgPool2d(kernel_size=6, stride=10, padding=0, ceil_mode=True,
count_include_pad=False)
input_data=[torch.randn([2, 3, 18, 18], dtype=torch.float32)]
torch_outputs = m(*[input.clone() for input in input_data])
trace = torch.jit.trace(m, input_data)
input_shapes = [('input0', torch.Size([2, 3, 18, 18]))]
mod, params = relay.frontend.from_pytorch(trace, input_shapes)
with tvm.transform.PassContext(opt_level=3):
exe = relay.create_executor('graph', mod=mod, params=params,
device=tvm.device('llvm', 0), target='llvm').evaluate()
input_tvm = dict(zip(['input0'], [inp.clone().cpu().numpy() for inp in
input_data]))
tvm_outputs = exe(**input_tvm).asnumpy()
np.testing.assert_allclose(torch_outputs, tvm_outputs, rtol=1e-3, atol=1e-3)
```
### Triage
Please refer to the list of label tags
[here](https://github.com/apache/tvm/wiki/Issue-Triage-Labels) to find the
relevant tags and add them below in a bullet format (example below).
* needs-triage
frontend:pytorch
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
