pytorch/test/test_stateless.py
Emilio Castillo fa38e93fe9 Add lightweight reparametrization for _stateless calls (#68969)
Summary:
https://github.com/pytorch/pytorch/issues/61447 introduced a mechanism for performing functional calls in a model using the reparametrization API. However, the overhead introduced in a single call was too large.
I tried to address this by modifying the reparametrization code to support spare tensors, but the changes needed were too large due to type checking and several parts of the code expecting actual `nn.Module` objects so this option was not feasible.

resnet50 and call functional with a parameters dict covering the 0, 25, 50, and 100% of the model total parameters.

Used script:
https://gist.github.com/emcastillo/f344a58638bd71d130c71c45f86f0c3a

| % of parameters passed | CPU Time (us) | GPU Time (us) |
|------------------------|---------------|---------------|
| regular call           | 5539          | 184909        |
| 0                      | 5561          | 184843        |
| 25                     | 11363         | 189236        |
| 50                     | 18716         | 195378        |
| 75                     | 22851         | 198641        |
| 100                    | 27441         | 202281        |

This PR just swaps the `__getattr__` of the submodules to look into a dict holding only the parameters when called, greatly reducing the burden of having to instantiate custom modules and calling forward to just retrieve a tensor.

The execution times now are as follows:

| % of parameters passed | CPU Time (us) | GPU Time (us) |
|------------------------|---------------|---------------|
| regular call           | 5939          | 187533        |
| 0                      | 5899          | 187570        |
| 25                     | 8541         | 188953        |
| 50                     | 10045         | 189826        |
| 75                     | 11049         | 190344        |
| 100                    | 11911         | 190800        |
| functorch with 100% params | 14014 | 191727

Now we see that the CPU time overhead is greatly reduced and the GPU time barely increases due to the effective overlap.

cc albanD zou3519

Pull Request resolved: https://github.com/pytorch/pytorch/pull/68969

Reviewed By: george-qi

Differential Revision: D33836360

Pulled By: albanD

fbshipit-source-id: 532561f64b18ca14c6ae2d77dcacb339397a589d
(cherry picked from commit fd4b6bdfbf)
2022-01-28 14:38:45 +00:00

157 lines
6.4 KiB
Python

# Owner(s): ["module: nn"]
import unittest
import torch
import torch.nn.utils._stateless as _stateless
from torch.testing._internal.common_cuda import TEST_MULTIGPU
from torch.testing._internal.common_utils import run_tests, TestCase
class MockModule(torch.nn.Module):
def __init__(self):
super().__init__()
self.l1 = torch.nn.Linear(1, 1)
self.register_buffer('buffer', torch.ones(1))
def forward(self, x):
return self.l1(x) + self.buffer
class TestStatelessFunctionalAPI(TestCase):
def _run_call_with_mock_module(self, module, device='cpu', prefix=''):
x = torch.rand((1, 1)).to(device)
weight = torch.tensor([[1.0]], device=device)
bias = torch.tensor([0.0], device=device)
buffer = torch.tensor([0.0], device=device)
if prefix != '':
parameters = {f'{prefix}.l1.weight': weight,
f'{prefix}.l1.bias': bias,
f'{prefix}.buffer': buffer}
else:
parameters = {'l1.weight': weight,
'l1.bias': bias,
'buffer': buffer}
to_check = module
if prefix != '':
to_check = getattr(module, prefix)
prev_weight = to_check.l1.weight.clone()
prev_buffer = to_check.buffer.clone()
# the parameters represent an identity function contrary to the
# existing params in module. So here we expect the result to be the
# same as the input if the weight swapping went well.
res = _stateless.functional_call(module, parameters, x)
self.assertEqual(x, res)
# check that the weight remain unmodified
cur_weight = to_check.l1.weight
cur_buffer = to_check.buffer
self.assertEqual(cur_weight, prev_weight)
self.assertEqual(cur_buffer, prev_buffer)
def test_functional_call(self):
module = MockModule()
self._run_call_with_mock_module(module)
def test_functional_call_with_jit(self):
module = MockModule()
jit_module = torch.jit.script(module)
with self.assertRaisesRegex(
RuntimeError,
r'used with Jitted modules'
):
self._run_call_with_mock_module(jit_module)
x = torch.rand((1, 1))
traced_module = torch.jit.trace(module, x)
with self.assertRaisesRegex(
RuntimeError,
r'used with Jitted modules'
):
self._run_call_with_mock_module(traced_module)
@unittest.skipIf(not TEST_MULTIGPU, 'multi-GPU not supported')
def test_functional_call_with_data_parallel(self):
module = MockModule()
module.cuda()
dp_module = torch.nn.DataParallel(module, [0, 1])
self._run_call_with_mock_module(dp_module, device='cuda', prefix='module')
def test_functional_call_with_gradient(self):
module = MockModule()
x = torch.rand((1, 1))
weight = torch.tensor([[1.0]], requires_grad=True)
bias = torch.tensor([0.0], requires_grad=True)
buffer = torch.tensor([0.0])
parameters = {'l1.weight': weight,
'l1.bias': bias,
'buffer': buffer}
res = _stateless.functional_call(module, parameters, x)
# Check that a backward step calculates the gradient of the supplied parameters
res.backward()
self.assertIsNotNone(weight.grad)
self.assertIsNotNone(bias.grad)
self.assertIsNone(buffer.grad)
# Gradient was not calculated for the module stated and buffers
self.assertIsNone(module.l1.weight.grad)
self.assertIsNone(module.l1.bias.grad)
self.assertIsNone(module.buffer.grad)
def test_functional_batch_norm(self):
module = torch.nn.BatchNorm1d(10)
module.train() # Allow stats update
# lets replace the running_mean buffer and check if its correctly updated
x = torch.full((20, 10), 128.0)
rm = torch.zeros(10)
parameters = {'running_mean': rm}
prev_rm = module.running_mean.clone()
res = _stateless.functional_call(module, parameters, x)
cur_rm = module.running_mean
self.assertEqual(cur_rm, prev_rm)
self.assertEqual(rm, torch.full((10,), 12.8))
# Now run functional without reparametrization and check that the module has
# been updated
res = _stateless.functional_call(module, {}, x)
self.assertEqual(module.running_mean, torch.full((10,), 12.8))
def test_circular_references(self):
module = MockModule()
# Add a circular reference
module.l1.m = module
x = torch.rand((1, 1))
weight = torch.tensor([[1.0]])
bias = torch.tensor([0.0])
buffer = torch.tensor([0.0])
parameters = {'l1.m.l1.weight': weight,
'l1.bias': bias,
'l1.m.buffer': buffer}
prev_weight = module.l1.weight.clone()
prev_buffer = module.buffer.clone()
res = _stateless.functional_call(module, parameters, x)
self.assertEqual(x, res)
# check that the weights remain unmodified and were correctly accesed
cur_weight = module.l1.weight
cur_buffer = module.buffer
self.assertEqual(cur_weight, prev_weight)
self.assertEqual(cur_buffer, prev_buffer)
def test_reparametrized_module_change_parametrization_original(self):
module = MockModule()
torch.nn.utils.parametrizations.spectral_norm(module.l1)
self.assertTrue('l1.parametrizations.weight.original' in dict(module.named_parameters()))
orig_sn_weight = module.l1.weight.clone()
x = torch.rand((1, 1))
# We substitute the parameter inside the parametrization
# the parametrization itself is not overwritten so it will be applied with a different
# value for the original tensor
parameters = {'l1.parametrizations.weight.original': torch.nn.Parameter(torch.tensor([[1.0]])),
'l1.bias': torch.tensor([0.0]),
'buffer': torch.tensor([0.0])}
res = torch.nn.utils._stateless.functional_call(module, parameters, x)
self.assertEqual(x, res)
# verify that the spectral normalization is still applied
self.assertTrue('l1.parametrizations.weight.original' in dict(module.named_parameters()))
self.assertEqual(orig_sn_weight, module.l1.weight)
if __name__ == '__main__':
run_tests()