OFA / fairseq /tests /test_amp_optimizer.py
root
init
93b9482
raw
history blame
2.47 kB
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import argparse
import copy
import unittest
import torch
from torch.cuda.amp import autocast, GradScaler
from fairseq.optim import build_optimizer
@unittest.skipIf(not torch.cuda.is_available(), "test requires a GPU")
class TestGradientScalingAMP(unittest.TestCase):
def setUp(self):
self.x = torch.tensor([2.0]).cuda().half()
weight = 3.0
bias = 5.0
self.error = 1.0
self.target = torch.tensor([self.x * weight + bias + self.error]).cuda()
self.loss_fn = torch.nn.L1Loss()
self.model = torch.nn.Linear(1, 1)
self.model.weight.data = torch.tensor([[weight]])
self.model.bias.data = torch.tensor([bias])
self.model.cuda()
self.params = list(self.model.parameters())
self.namespace_dls = argparse.Namespace(
optimizer="adam",
lr=[0.1],
adam_betas="(0.9, 0.999)",
adam_eps=1e-8,
weight_decay=0.0,
threshold_loss_scale=1,
min_loss_scale=1e-4,
)
self.scaler = GradScaler(
init_scale=1,
growth_interval=1,
)
def run_iter(self, model, params, optimizer):
optimizer.zero_grad()
with autocast():
y = model(self.x)
loss = self.loss_fn(y, self.target)
self.scaler.scale(loss).backward()
self.assertEqual(loss, torch.tensor(1.0, device="cuda:0", dtype=torch.float16))
self.scaler.unscale_(optimizer)
grad_norm = optimizer.clip_grad_norm(0)
self.assertAlmostEqual(grad_norm.item(), 2.2361, 4)
self.scaler.step(optimizer)
self.scaler.update()
self.assertEqual(
model.weight,
torch.tensor(
[[3.1]], device="cuda:0", requires_grad=True
),
)
self.assertEqual(
model.bias,
torch.tensor(
[5.1], device="cuda:0", requires_grad=True
),
)
self.assertEqual(self.scaler.get_scale(), 2.0)
def test_automatic_mixed_precision(self):
model = copy.deepcopy(self.model)
params = list(model.parameters())
optimizer = build_optimizer(self.namespace_dls, params)
self.run_iter(model, params, optimizer)