Add Python/C++ torch.nn API parity test harness #23852
Conversation
pytorchbot
added
module: nn
test/common_nn.py
Outdated
| dict( | ||
| module_name='Linear', | ||
| constructor_args=(10, 8), | ||
| cpp_constructor_args='(10, 8)', |
There was a problem hiding this comment.
You don't need to do this, but we should really turn this into a namedtuple at some point
zou3519
left a comment
zou3519
left a comment
There was a problem hiding this comment.
Not a full review yet, but some initial comments/questions.
yf225
changed the title
zou3519
left a comment
zou3519
left a comment
There was a problem hiding this comment.
Approach looks great, some comments
test/test_cpp_api_parity.py
Outdated
| cpp_args = [] | ||
| for module_file_name in module_file_names: | ||
| cpp_args.append(module_file_name) | ||
| for arg in args[1:]: |
There was a problem hiding this comment.
How dynamic args is makes me a little uncomfortable, but... I guess it is okay
There was a problem hiding this comment.
Yeah it's a bit unfortunate because the C++ test function has to take an unpacked list of input arguments.. (I briefly thought about accepting std::vector<IValue> in all torch::nn modules' forward, but it seems to need a bigger use case than just making the test nicer :/)
yf225
force-pushed
the
cpp_parity_test
branch
3 times, most recently
from
496826c to
7e833f3
Compare
|
@pytorchbot rebase this please |
yf225
force-pushed
the
cpp_parity_test
branch
4 times, most recently
from
be6c62f to
d03cad2
Compare
facebook-github-bot
left a comment
facebook-github-bot
left a comment
There was a problem hiding this comment.
@yf225 has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
facebook-github-bot
left a comment
facebook-github-bot
left a comment
There was a problem hiding this comment.
@yf225 is landing this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
5 participants
This PR adds test harness for checking Python / C++ API parity for
torch.nn.Modulesubclasses. Under the hood, we use JIT tracing to transfernn.Modulestate from Python to C++, so that we can test initialization / forward / backward on Python / C++ modules with the same parameters and buffers.