[Bug][relay][pytorch] Wrong implementation logic in Threshold operator

[jikechao]

Relay has a wrong implementation logic about the threshold operator, It was used the same as relu. You can see Line 1334-1336 in file frontend/pytorch.py

The implementation logic in PyTorch documentation is shown as follows:

This bug was detected by this test case:

import torch
from tvm import relay
import tvm
import numpy as np

m = torch.nn.Threshold(1.0, 3.0,)
input_data = torch.tensor([[-1.0, 2.0]], dtype=torch.float32)

torch_outputs = m(input_data)

trace = torch.jit.trace(m, input_data)
input_shapes = [('input0', torch.Size([1, 2]))]

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 = {'input0': np.array([[-1.,  2.]], dtype='float32')}
tvm_outputs = exe(**input_tvm).asnumpy()

np.testing.assert_allclose(torch_outputs, tvm_outputs, rtol=1e-3, atol=1e-3)

Expected behavior

TVM gives same result as Pytorch

Actual behavior

TVM gave different result with Pytorch, It seems TVM produce the wrong result.

Triage

• frontend:pytorch

cc @Hzfengsy @echuraev @shingjan @yelite

[jikechao]

Fixing this bug needs to master the complete code structure of Relay. I am not very familiar with this process.
@Hzfengsy @echuraev @shingjan
Are you willing to give me some advice and suggestion on how to learn it?
BTW, From my knowledge, the core of TVM is optimization, do you think detecting and fixing frontend bugs is meaningful work?

[echuraev]

@jikechao Thank you for your finds. IMHO, fixing frontend bugs is also very important work. TVM also cares about the accuracy because if we have a good performance but bad accuracy then it makes no sense because the final result is incorrect. By fixing such bugs and adding new operators support, you make a great contribution to model coverage by TVM. I think it is an important work.

Are you willing to give me some advice and suggestion on how to learn it?

Also, it is my personal opinion, but usually I use official documentation, you can find there many useful articles. E.g.:

• Python API of Relay: https://tvm.apache.org/docs/reference/api/python/relay/index.html
  If you type relay in the search field, then you see many different articles related to relay and you can find something useful for you.

Another trick which I commonly use, I try to find in the source code or in the git history something similar to my task and understand which components I have to modify. For example, if I need to add new operator support, I would find another commit which adds some another operator to TVM and learn how it was done.

[jikechao]

@echuraev Thank you for your valuable suggestion. I will continue to learn and try to detect and fix frontend bugs.

[jikechao]

This issue has been fixed by the PR-14820