This updates ruff to 0.285 which is faster, better, and have fixes a bunch of false negatives with regards to fstrings.
I also enabled RUF017 which looks for accidental quadratic list summation. Luckily, seems like there are no instances of it in our codebase, so enabling it so that it stays like that. :)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107519
Approved by: https://github.com/ezyang
This updates ruff to 0.285 which is faster, better, and have fixes a bunch of false negatives with regards to fstrings.
I also enabled RUF017 which looks for accidental quadratic list summation. Luckily, seems like there are no instances of it in our codebase, so enabling it so that it stays like that. :)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107519
Approved by: https://github.com/ezyang
Summary:
Adds a new file under `torch/nn/utils/parametrizations.py` which should contain all the parametrization implementations
For spectral_norm we add the `SpectralNorm` module which can be registered using `torch.nn.utils.parametrize.register_parametrization` or using a wrapper: `spectral_norm`, the same API the old implementation provided.
Most of the logic is borrowed from the old implementation:
- Just like the old implementation, there should be cases when retrieving the weight should perform another power iteration (thus updating the weight) and cases where it shouldn't. For example in eval mode `self.training=True`, we do not perform power iteration.
There are also some differences/difficulties with the new implementation:
- Using new parametrization functionality as-is there doesn't seem to be a good way to tell whether a 'forward' call was the result of parametrizations are unregistered (and leave_parametrizations=True) or when the injected property's getter was invoked. The issue is that we want perform power iteration in the latter case but not the former, but we don't have this control as-is. So, in this PR I modified the parametrization functionality to change the module to eval mode before triggering their forward call
- Updates the vectors based on weight on initialization to fix https://github.com/pytorch/pytorch/issues/51800 (this avoids silently update weights in eval mode). This also means that we perform twice any many power iterations by the first forward.
- right_inverse is just the identity for now, but maybe it should assert that the passed value already satisfies the constraints
- So far, all the old spectral_norm tests have been cloned, but maybe we don't need so much testing now that the core functionality is already well tested
Pull Request resolved: https://github.com/pytorch/pytorch/pull/57784
Reviewed By: ejguan
Differential Revision: D28413201
Pulled By: soulitzer
fbshipit-source-id: e8f1140f7924ca43ae4244c98b152c3c554668f2
Summary:
Retake on https://github.com/pytorch/pytorch/issues/40493 after all the feedback from albanD
This PR implements the generic Lazy mechanism and a sample `LazyLinear` layer with the `UninitializedParameter`.
The main differences with the previous PR are two;
Now `torch.nn.Module` remains untouched.
We don't require an explicit initialization or a dummy forward pass before starting the training or inference of the actual module. Making this much simpler to use from the user side.
As we discussed offline, there was the suggestion of not using a mixin, but changing the `__class__` attribute of `LazyLinear` to become `Linear` once it's completely initialized. While this can be useful, by the time being we need `LazyLinear` to be a `torch.nn.Module` subclass since there are many checks that rely on the modules being instances of `torch.nn.Module`.
This can cause problems when we create complex modules such as
```
class MyNetwork(torch.nn.Module):
def __init__(self):
super(MyNetwork, self).__init__()
self.conv = torch.nn.Conv2d(20, 4, 2)
self.linear = torch.nn.LazyLinear(10)
def forward(self, x):
y = self.conv(x).clamp(min=0)
return self.linear(y)
```
Here, when the __setattr__ function is called at the time LazyLinear is registered, it won't be added to the child modules of `MyNetwork`, so we have to manually do it later, but currently there is no way to do such thing as we can't access the parent module from LazyLinear once it becomes the Linear module. (We can add a workaround to this if needed).
TODO:
Add convolutions once the design is OK
Fix docstrings
Pull Request resolved: https://github.com/pytorch/pytorch/pull/44538
Reviewed By: ngimel
Differential Revision: D24162854
Pulled By: albanD
fbshipit-source-id: 6d58dfe5d43bfb05b6ee506e266db3cf4b885f0c
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18598
ghimport-source-id: c74597e5e7437e94a43c163cee0639b20d0d0c6a
Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18598 Turn on F401: Unused import warning.**
This was requested by someone at Facebook; this lint is turned
on for Facebook by default. "Sure, why not."
I had to noqa a number of imports in __init__. Hypothetically
we're supposed to use __all__ in this case, but I was too lazy
to fix it. Left for future work.
Be careful! flake8-2 and flake8-3 behave differently with
respect to import resolution for # type: comments. flake8-3 will
report an import unused; flake8-2 will not. For now, I just
noqa'd all these sites.
All the changes were done by hand.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Differential Revision: D14687478
fbshipit-source-id: 30d532381e914091aadfa0d2a5a89404819663e3
Summary:
Problems with SN and DP after #12671 :
1. in eval mode, `weight_orig` is not getting correct gradient #12737 .
Fix: keep `v` vector around as a buffer and always calculate `W = W_orig / (u @ W_orig @ v)` even in eval.
2. in training mode, the `weight` buffer of the parallelized module is never updated, if someone touches `weight_orig` and/or `weight` and makes them not sharing storage. So in `eval` the weight used is wrong.
Fix: Make `weight` not a buffer anymore and always calculate it as above.
3. #12671 changed SN to update `u` in-place to make DP work correctly, but then it breaks backward through two forwards (e.g., the common GAN loss `D(real) - D(fake)`) because the vectors needed to backprop the 1st forward is changed in the 2nd forward.
Fix: This PR clones `u` and `v` before using them.
To maintain BC, I added a hook interface for producing and loading state_dict. This is ugly and we should really have better interface for spectral_norm. But for the purpose to fix this issue, I make this patch. Even if we have a better interface, BC mechanism for legacy loading legacy state_dict still needs to be done.
cc The controller you requested could not be found. crcrpar
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13350
Differential Revision: D12931044
Pulled By: SsnL
fbshipit-source-id: 8be6f934eaa62414d76d2c644dedd7e1b7eb31ef
Summary:
There were two problems with SN + DP:
1. In SN, the updated _u vector is saved back to module via a `setattr`. However, in DP, everything is run on a replica, so those updates are lost.
2. In DP, the buffers are broadcast via a `broadcast_coalesced`, so on replicas they are all views. Therefore, the `detach_` call won't work.
Fixes are:
1. Update _u vector in-place so, by the shared storage between 1st replica and the parallelized module, the update is retained
2. Do not call `detach_`.
3. Added comments in SN about the subtlety.
4. Added a note to the DP doc on this particular behavior of DP.
cc crcrpar taesung89 The controller you requested could not be found. yaoshengfu
Fixes https://github.com/pytorch/pytorch/issues/11476
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12671
Differential Revision: D10410232
Pulled By: SsnL
fbshipit-source-id: c447951844a30366d8c196bf9436340e88f3b6d9
Summary:
As we left weight to be the last calculated weight in eval mode, we need to detach it from the computation in order to facilitate using backward.
The typical use case is in GANs when the discriminator has spectral norm, is in eval mode and we want to backprop through the discriminator to get weight gradients for the generator.
Closes https://github.com/pytorch/pytorch/pull/9020
Reviewed By: ezyang
Differential Revision: D8694054
Pulled By: SsnL
fbshipit-source-id: 09ee5843687cac3ed4c40759ac577a14c5371730
* Spectral norm improvements
- Don't do iterations on weight in eval mode
To facilitate this, register weight as buffer in order to be able
to use module with spectral norm in eval mode after immediately
after loading state dict (#8208)
- Use weight instead of weight_orig as weight when removing
spectral norm
- Add dim parameter in case the normalization should occur w.r.t.
a dimension other than 0 (#7865)
* add and update spectral norm tests
* More spectral norm tests
Thank you, Simon, for the suggestions.
* initial commit for spectral norm
* fix comment
* edit rst
* fix doc
* remove redundant empty line
* fix nit mistakes in doc
* replace l2normalize with F.normalize
* fix chained `by`
* fix docs
fix typos
add comments related to power iteration and epsilon
update link to the paper
make some comments specific
* fix typo
