Functorch gradients: investigation and fix

ffuuugor commented 1 year ago

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 #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