Functorch gradients: investigation and fix #510
Conversation
facebook-github-bot
added
the
CLA Signed
|
@ffuuugor has updated the pull request. You must reimport the pull request before landing. |
|
@facebook-github-bot has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@ffuuugor has updated the pull request. You must reimport the pull request before landing. |
|
@facebook-github-bot has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@ffuuugor has updated the pull request. You must reimport the pull request before landing. |
|
@facebook-github-bot has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
ffuuugor
requested review from
karthikprasad,
alexandresablayrolles and
romovpa
and removed request for
alexandresablayrolles
|
@ffuuugor has updated the pull request. You must reimport the pull request before landing. |
|
@facebook-github-bot has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
There was a problem hiding this comment.
Thanks a lot for implementing this solution and doing a proper benchmark. I like this latter option as I think it's more conceptually elegant and more performant. Also, the code for iterate_submodules is short and pretty self-explanatory so it matches the conceptual elegance of the idea!
| if has_trainable_params(module): | ||
| yield module | ||
|
|
||
| # we'll apply functorch for the entire substree |
There was a problem hiding this comment.
Nit: can we replace by
# Don't recurse if module is handled by functorch
|
@ffuuugor has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
1 similar comment
|
@ffuuugor has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@ffuuugor has updated the pull request. You must reimport the pull request before landing. |
|
@ffuuugor has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Summary:
*The investigation part for this PR was done by alexandresablayrolles, thanks for figuring out the reason the tests were failing*
## Background
Current implementation of functorch-based per sample gradients fails on modules which have both trainable non-recursive parameters and standard submodules, e.g. below
```
class LinearWithExtraParam(nn.Module):
def __init__(self, in_features: int, out_features: int, hidden_dim: int = 8):
super().__init__()
self.fc = nn.Linear(in_features, hidden_dim)
self.extra_param = nn.Parameter(torch.randn(hidden_dim, out_features))
def forward(self, x):
x = self.fc(x)
x = x.matmul(self.extra_param)
return x
```
The reason is - functorch hook actually computes gradients for recursive submodules too. The problem is, normal hooks are also attached to these submodules. GradSampleModule then sees two grad_sample tensors, thinks it needs to accumulate and adds them up together
## Solution(s)
There are essentially two ways we can fix this: either make functorch compute per sample gradients for non-recursive parameters only or don't attach normal hooks to submodules where the parent module is handled by functorch.
This diff implements the latter option (reasoning below), for demo purposes the former option can be seen in pytorch#531
For the pure code perspective the former option (let's call it "non-recursive functorch") is more appealing to me. It better fits the existing paradigm and matches normal hooks behaviour - all of the existing code only deals with the immediate non-recursive parameters.
However, it doesn't make much sense from the efficiency perspective. "non-recursive functorch" would do all the work to compute per-sample gradients for its submodules, only for them to be filtered out at the very last stage.
Alternative option (a.k.a. "functorch for subtrees") does involve a bit more convoluted
This has a noticeable effect on performance.
Below is the results of MNIST benchmarks with different configurations. I've tested this with different configurations, because at the end of the day, the impact on performance depends on how deep are subtrees
* Standard model- our model from MNIST example, standard layers only (2 conv + 2 linear). No overhead expected, functorch doesn't kick in
* Mid-level model - leaf nodes (two linear layers) have one extra param and are computed with functorch. Overhead: 2x Linear hook
* Extreme model - root model have one extra param and needs to be handled by functorch. Overhead: 2x linear hook + 2x conv hook
| Mode | non-recursive functorch | functorch for subtrees |
|:-----------------------:|:------------------------:|:-----------------------:|
| Standard model (CPU) | 138s | 136s |
| Standard model (GPU) | 149s | 150s |
| Mid-level model (CPU) | 157s | 150s |
| Mid-level model (GPU) | 100s | 97s |
| Extreme model (CPU) | 207s | 172s |
| Extreme model (GPU) | 101s | 94s |
Pull Request resolved: pytorch#510
Reviewed By: alexandresablayrolles
Differential Revision: D39579487
Pulled By: ffuuugor
fbshipit-source-id: 1b089bd04ab110174a1f2ebb371380eb2ce76054
The investigation part for this PR was done by @alexandresablayrolles, thanks for figuring out the reason the tests were failing
Background
Current implementation of functorch-based per sample gradients fails on modules which have both trainable non-recursive parameters and standard submodules, e.g. below
The reason is - functorch hook actually computes gradients for recursive submodules too. The problem is, normal hooks are also attached to these submodules. GradSampleModule then sees two grad_sample tensors, thinks it needs to accumulate and adds them up together
Solution(s)
There are essentially two ways we can fix this: either make functorch compute per sample gradients for non-recursive parameters only or don't attach normal hooks to submodules where the parent module is handled by functorch.
This diff implements the latter option (reasoning below), for demo purposes the former option can be seen in #531
For the pure code perspective the former option (let's call it "non-recursive functorch") is more appealing to me. It better fits the existing paradigm and matches normal hooks behaviour - all of the existing code only deals with the immediate non-recursive parameters.
However, it doesn't make much sense from the efficiency perspective. "non-recursive functorch" would do all the work to compute per-sample gradients for its submodules, only for them to be filtered out at the very last stage.
Alternative option (a.k.a. "functorch for subtrees") does involve a bit more convoluted
This has a noticeable effect on performance.
Below is the results of MNIST benchmarks with different configurations. I've tested this with different configurations, because at the end of the day, the impact on performance depends on how deep are subtrees