[WIP] continue PR #784 (#1221)

* Add gradient cumulative optimizer

fixes #190

* Update optimizer.py

* Update optimizer.py

* fix loss scale improperly in last equivalent_iter

* Add `GradientCumulativeOptimizerHook` in `__init__.py`.

* Add docstring of `GradientCumulativeOptimizerHook`.

* Add type check, BN warning and resume warning. And fix typo, lint the
code.

* Add unit test

* Update docstring example.

* Change GradientCumulativeOptimizerHook `__init__` arguments.

* Add GradientCumulativeOptimzierHook unit tests with IterBasedRunner.

* Add GradientCumulativeFp16OptimizerHook.

* Add unit tests of GradientCumulativeFp16OptimizerHook

* Use '!=' instead of '>' to determine resume

Co-authored-by: Zhiyuan Chen <this@zyc.ai>

mmcv/runner/__init__.py

@@ -11,10 +11,12 @@ from .epoch_based_runner import EpochBasedRunner, Runner
 from .fp16_utils import LossScaler, auto_fp16, force_fp32, wrap_fp16_model
 from .hooks import (HOOKS, CheckpointHook, ClosureHook, DistEvalHook,
                     DistSamplerSeedHook, DvcliveLoggerHook, EMAHook, EvalHook,
-                    Fp16OptimizerHook, Hook, IterTimerHook, LoggerHook,
-                    LrUpdaterHook, MlflowLoggerHook, NeptuneLoggerHook,
-                    OptimizerHook, PaviLoggerHook, SyncBuffersHook,
-                    TensorboardLoggerHook, TextLoggerHook, WandbLoggerHook)
+                    Fp16OptimizerHook, GradientCumulativeFp16OptimizerHook,
+                    GradientCumulativeOptimizerHook, Hook, IterTimerHook,
+                    LoggerHook, LrUpdaterHook, MlflowLoggerHook,
+                    NeptuneLoggerHook, OptimizerHook, PaviLoggerHook,
+                    SyncBuffersHook, TensorboardLoggerHook, TextLoggerHook,
+                    WandbLoggerHook)
 from .iter_based_runner import IterBasedRunner, IterLoader
 from .log_buffer import LogBuffer
 from .optimizer import (OPTIMIZER_BUILDERS, OPTIMIZERS,
@@ -39,5 +41,6 @@ __all__ = [
     'SyncBuffersHook', 'EMAHook', 'build_runner', 'RUNNERS', 'allreduce_grads',
     'allreduce_params', 'LossScaler', 'CheckpointLoader', 'BaseModule',
     '_load_checkpoint_with_prefix', 'EvalHook', 'DistEvalHook', 'Sequential',
-    'ModuleList'
+    'ModuleList', 'GradientCumulativeOptimizerHook',
+    'GradientCumulativeFp16OptimizerHook'
 ]

mmcv/runner/hooks/__init__.py

@@ -11,7 +11,8 @@ from .logger import (DvcliveLoggerHook, LoggerHook, MlflowLoggerHook,
 from .lr_updater import LrUpdaterHook
 from .memory import EmptyCacheHook
 from .momentum_updater import MomentumUpdaterHook
-from .optimizer import Fp16OptimizerHook, OptimizerHook
+from .optimizer import (Fp16OptimizerHook, GradientCumulativeFp16OptimizerHook,
+                        GradientCumulativeOptimizerHook, OptimizerHook)
 from .profiler import ProfilerHook
 from .sampler_seed import DistSamplerSeedHook
 from .sync_buffer import SyncBuffersHook
@@ -23,5 +24,6 @@ __all__ = [
     'PaviLoggerHook', 'TextLoggerHook', 'TensorboardLoggerHook',
     'NeptuneLoggerHook', 'WandbLoggerHook', 'DvcliveLoggerHook',
     'MomentumUpdaterHook', 'SyncBuffersHook', 'EMAHook', 'EvalHook',
-    'DistEvalHook', 'ProfilerHook'
+    'DistEvalHook', 'ProfilerHook', 'GradientCumulativeOptimizerHook',
+    'GradientCumulativeFp16OptimizerHook'
 ]

mmcv/runner/hooks/optimizer.py

@@ -5,7 +5,7 @@ from itertools import chain
 
 from torch.nn.utils import clip_grad
 
-from mmcv.utils import TORCH_VERSION, digit_version
+from mmcv.utils import TORCH_VERSION, _BatchNorm, digit_version
 from ..dist_utils import allreduce_grads
 from ..fp16_utils import LossScaler, wrap_fp16_model
 from .hook import HOOKS, Hook
@@ -42,6 +42,91 @@ class OptimizerHook(Hook):
         runner.optimizer.step()
 
 
+@HOOKS.register_module()
+class GradientCumulativeOptimizerHook(OptimizerHook):
+    """Optimizer Hook implements multi-iters gradient cumulating.
+
+    Args:
+        cumulative_iters (int, optional): Num of gradient cumulative iters.
+            The optimizer will step every `cumulative_iters` iters.
+            Defaults to 1.
+
+    Examples:
+        >>> # Use cumulative_iters to simulate a large batch size
+        >>> # It is helpful when the hardware cannot handle a large batch size.
+        >>> loader = DataLoader(data, batch_size=64)
+        >>> optim_hook = GradientCumulativeOptimizerHook(cumulative_iters=4)
+        >>> # almost equals to
+        >>> loader = DataLoader(data, batch_size=256)
+        >>> optim_hook = OptimizerHook()
+    """
+
+    def __init__(self, cumulative_iters=1, **kwargs):
+        super(GradientCumulativeOptimizerHook, self).__init__(**kwargs)
+
+        assert isinstance(cumulative_iters, int) and cumulative_iters > 0, \
+            f'cumulative_iters only accepts positive int, but got ' \
+            f'{type(cumulative_iters)} instead.'
+
+        self.cumulative_iters = cumulative_iters
+        self.divisible_iters = 0
+        self.remainder_iters = 0
+        self.initialized = False
+
+    def has_batch_norm(self, module):
+        if isinstance(module, _BatchNorm):
+            return True
+        for m in module.children():
+            if self.has_batch_norm(m):
+                return True
+        return False
+
+    def _init(self, runner):
+        if runner.iter % self.cumulative_iters != 0:
+            runner.logger.warning(
+                'Resume iter number is not divisible by cumulative_iters in '
+                'GradientCumulativeOptimizerHook, which means the gradient of '
+                'some iters is lost and the result may be influenced slightly.'
+            )
+
+        if self.has_batch_norm(runner.model) and self.cumulative_iters > 1:
+            runner.logger.warning(
+                'GradientCumulativeOptimizerHook may slightly decrease '
+                'performance if the model has BatchNorm layers.')
+
+        residual_iters = runner.max_iters - runner.iter
+
+        self.divisible_iters = (
+            residual_iters // self.cumulative_iters * self.cumulative_iters)
+        self.remainder_iters = residual_iters - self.divisible_iters
+
+        self.initialized = True
+
+    def after_train_iter(self, runner):
+        if not self.initialized:
+            self._init(runner)
+
+        if runner.iter < self.divisible_iters:
+            loss_factor = self.cumulative_iters
+        else:
+            loss_factor = self.remainder_iters
+        loss = runner.outputs['loss']
+        loss = loss / loss_factor
+        loss.backward()
+
+        if (self.every_n_iters(runner, self.cumulative_iters)
+                or self.is_last_iter(runner)):
+
+            if self.grad_clip is not None:
+                grad_norm = self.clip_grads(runner.model.parameters())
+                if grad_norm is not None:
+                    # Add grad norm to the logger
+                    runner.log_buffer.update({'grad_norm': float(grad_norm)},
+                                             runner.outputs['num_samples'])
+            runner.optimizer.step()
+            runner.optimizer.zero_grad()
+
+
 if (TORCH_VERSION != 'parrots'
         and digit_version(TORCH_VERSION) >= digit_version('1.6.0')):
 
@@ -152,6 +237,60 @@ if (TORCH_VERSION != 'parrots'
             # save state_dict of loss_scaler
             runner.meta.setdefault(
                 'fp16', {})['loss_scaler'] = self.loss_scaler.state_dict()
+
+    @HOOKS.register_module()
+    class GradientCumulativeFp16OptimizerHook(GradientCumulativeOptimizerHook,
+                                              Fp16OptimizerHook):
+        """Fp16 optimizer Hook (using PyTorch's implementation) implements
+        multi-iters gradient cumulating.
+
+        If you are using PyTorch >= 1.6, torch.cuda.amp is used as the backend,
+        to take care of the optimization procedure.
+        """
+
+        def __init__(self, *args, **kwargs):
+            super(GradientCumulativeFp16OptimizerHook,
+                  self).__init__(*args, **kwargs)
+
+        def after_train_iter(self, runner):
+            if not self.initialized:
+                self._init(runner)
+
+            if runner.iter < self.divisible_iters:
+                loss_factor = self.cumulative_iters
+            else:
+                loss_factor = self.remainder_iters
+            loss = runner.outputs['loss']
+            loss = loss / loss_factor
+
+            self.loss_scaler.scale(loss).backward()
+
+            if (self.every_n_iters(runner, self.cumulative_iters)
+                    or self.is_last_iter(runner)):
+
+                # copy fp16 grads in the model to fp32 params in the optimizer
+                self.loss_scaler.unscale_(runner.optimizer)
+
+                if self.grad_clip is not None:
+                    grad_norm = self.clip_grads(runner.model.parameters())
+                    if grad_norm is not None:
+                        # Add grad norm to the logger
+                        runner.log_buffer.update(
+                            {'grad_norm': float(grad_norm)},
+                            runner.outputs['num_samples'])
+
+                # backward and update scaler
+                self.loss_scaler.step(runner.optimizer)
+                self.loss_scaler.update(self._scale_update_param)
+
+                # save state_dict of loss_scaler
+                runner.meta.setdefault(
+                    'fp16', {})['loss_scaler'] = self.loss_scaler.state_dict()
+
+                # clear grads
+                runner.model.zero_grad()
+                runner.optimizer.zero_grad()
+
 else:
 
     @HOOKS.register_module()
@@ -295,3 +434,75 @@ else:
             # save state_dict of loss_scaler
             runner.meta.setdefault(
                 'fp16', {})['loss_scaler'] = self.loss_scaler.state_dict()
+
+    @HOOKS.register_module()
+    class GradientCumulativeFp16OptimizerHook(GradientCumulativeOptimizerHook,
+                                              Fp16OptimizerHook):
+        """Fp16 optimizer Hook (using mmcv implementation) implements multi-
+        iters gradient cumulating."""
+
+        def __init__(self, *args, **kwargs):
+            super(GradientCumulativeFp16OptimizerHook,
+                  self).__init__(*args, **kwargs)
+
+        def after_train_iter(self, runner):
+            if not self.initialized:
+                self._init(runner)
+
+            if runner.iter < self.divisible_iters:
+                loss_factor = self.cumulative_iters
+            else:
+                loss_factor = self.remainder_iters
+
+            loss = runner.outputs['loss']
+            loss = loss / loss_factor
+
+            # scale the loss value
+            scaled_loss = loss * self.loss_scaler.loss_scale
+            scaled_loss.backward()
+
+            if (self.every_n_iters(runner, self.cumulative_iters)
+                    or self.is_last_iter(runner)):
+
+                # copy fp16 grads in the model to fp32 params in the optimizer
+                fp32_weights = []
+                for param_group in runner.optimizer.param_groups:
+                    fp32_weights += param_group['params']
+                self.copy_grads_to_fp32(runner.model, fp32_weights)
+                # allreduce grads
+                if self.distributed:
+                    allreduce_grads(fp32_weights, self.coalesce,
+                                    self.bucket_size_mb)
+
+                has_overflow = self.loss_scaler.has_overflow(fp32_weights)
+                # if has overflow, skip this iteration
+                if not has_overflow:
+                    # scale the gradients back
+                    for param in fp32_weights:
+                        if param.grad is not None:
+                            param.grad.div_(self.loss_scaler.loss_scale)
+                    if self.grad_clip is not None:
+                        grad_norm = self.clip_grads(fp32_weights)
+                        if grad_norm is not None:
+                            # Add grad norm to the logger
+                            runner.log_buffer.update(
+                                {'grad_norm': float(grad_norm)},
+                                runner.outputs['num_samples'])
+                    # update fp32 params
+                    runner.optimizer.step()
+                    # copy fp32 params to the fp16 model
+                    self.copy_params_to_fp16(runner.model, fp32_weights)
+                else:
+                    runner.logger.warning(
+                        'Check overflow, downscale loss scale '
+                        f'to {self.loss_scaler.cur_scale}')
+
+                self.loss_scaler.update_scale(has_overflow)
+
+                # save state_dict of loss_scaler
+                runner.meta.setdefault(
+                    'fp16', {})['loss_scaler'] = self.loss_scaler.state_dict()
+
+                # clear grads
+                runner.model.zero_grad()
+                runner.optimizer.zero_grad()

tests/test_runner/test_hooks.py

@@ -20,8 +20,12 @@ from torch.nn.init import constant_
 from torch.utils.data import DataLoader
 
 from mmcv.runner import (CheckpointHook, DvcliveLoggerHook, EMAHook,
-                         IterTimerHook, MlflowLoggerHook, NeptuneLoggerHook,
+                         Fp16OptimizerHook,
+                         GradientCumulativeFp16OptimizerHook,
+                         GradientCumulativeOptimizerHook, IterTimerHook,
+                         MlflowLoggerHook, NeptuneLoggerHook, OptimizerHook,
                          PaviLoggerHook, WandbLoggerHook, build_runner)
+from mmcv.runner.fp16_utils import auto_fp16
 from mmcv.runner.hooks.hook import HOOKS, Hook
 from mmcv.runner.hooks.lr_updater import (CosineRestartLrUpdaterHook,
                                           CyclicLrUpdaterHook,
@@ -1229,3 +1233,207 @@ def test_get_triggered_stages():
     # stages output have order, so here is list instead of set.
     expected_stages = ['before_run', 'after_train_epoch', 'after_val_epoch']
     assert hook.get_triggered_stages() == expected_stages
+
+
+def test_gradient_cumulative_optimizer_hook():
+
+    class ToyModel(nn.Module):
+
+        def __init__(self, with_norm=False):
+            super().__init__()
+            self.fp16_enabled = False
+            self.fc = nn.Linear(3, 2)
+            nn.init.constant_(self.fc.weight, 1.)
+            nn.init.constant_(self.fc.bias, 1.)
+            self.with_norm = with_norm
+            if with_norm:
+                self.norm = nn.BatchNorm1d(2)
+
+        def forward(self, x):
+            x = self.fc(x)
+            if self.with_norm:
+                x = self.norm(x)
+            return x
+
+        def train_step(self, x, optimizer, **kwargs):
+            return dict(loss=self(x).mean(), num_samples=x.shape[0])
+
+        def val_step(self, x, optimizer, **kwargs):
+            return dict(loss=self(x).mean(), num_samples=x.shape[0])
+
+    def build_toy_runner(config=dict(type='EpochBasedRunner', max_epochs=3)):
+        model = ToyModel()
+        optimizer = torch.optim.SGD(model.parameters(), lr=0.02)
+        tmp_dir = tempfile.mkdtemp()
+
+        runner = build_runner(
+            config,
+            default_args=dict(
+                model=model,
+                work_dir=tmp_dir,
+                optimizer=optimizer,
+                logger=logging.getLogger(),
+                meta=dict()))
+        return runner
+
+    with pytest.raises(AssertionError):
+        # cumulative_iters only accepts int
+        GradientCumulativeOptimizerHook(cumulative_iters='str')
+
+    with pytest.raises(AssertionError):
+        # cumulative_iters only accepts positive number
+        GradientCumulativeOptimizerHook(cumulative_iters=-1)
+
+    # test epoch based runner
+    data = torch.rand((6, 3))
+    # optimize with cumulative_iters
+    loader_1 = DataLoader(data, batch_size=1)
+    runner_1 = build_toy_runner()
+    optimizer_hook = GradientCumulativeOptimizerHook(
+        grad_clip=dict(max_norm=0.2), cumulative_iters=3)
+    runner_1.register_hook(optimizer_hook)
+    runner_1.run([loader_1], [('train', 1)])
+
+    # optimize without cumulative_iters
+    loader_2 = DataLoader(data, batch_size=3)
+    runner_2 = build_toy_runner()
+    optimizer_hook = OptimizerHook(grad_clip=dict(max_norm=0.2))
+    runner_2.register_hook(optimizer_hook)
+    runner_2.run([loader_2], [('train', 1)])
+
+    # test optimizer works well
+    assert (runner_1.model.fc.weight < 1).all()
+    assert (runner_1.model.fc.bias < 1).all()
+    # test optimizer with cumulative_iters gets the same results
+    assert torch.allclose(runner_1.model.fc.weight, runner_2.model.fc.weight)
+    assert torch.allclose(runner_1.model.fc.bias, runner_2.model.fc.bias)
+    shutil.rmtree(runner_1.work_dir)
+    shutil.rmtree(runner_2.work_dir)
+
+    # test iter based runner
+    data = torch.rand((8, 3))
+    # optimize with cumulative_iters
+    loader_1 = DataLoader(data, batch_size=1)
+    runner_1 = build_toy_runner(dict(type='IterBasedRunner', max_iters=8))
+    optimizer_hook = GradientCumulativeOptimizerHook(
+        grad_clip=dict(max_norm=0.2), cumulative_iters=3)
+    runner_1.register_hook(optimizer_hook)
+    runner_1.run([loader_1], [('train', 1)])
+
+    # optimize without cumulative_iters
+    loader_2_divisible = DataLoader(data[:6], batch_size=3)
+    loader_2_remainder = DataLoader(data[6:], batch_size=2)
+    runner_2 = build_toy_runner(dict(type='IterBasedRunner', max_iters=3))
+    optimizer_hook = OptimizerHook(grad_clip=dict(max_norm=0.2))
+    runner_2.register_hook(optimizer_hook)
+    runner_2.run([loader_2_divisible, loader_2_remainder], [('train', 2),
+                                                            ('train', 1)])
+
+    # test optimizer works well
+    assert (runner_1.model.fc.weight < 1).all()
+    assert (runner_1.model.fc.bias < 1).all()
+    # test optimizer with cumulative_iters gets the same results
+    assert torch.allclose(runner_1.model.fc.weight, runner_2.model.fc.weight)
+    assert torch.allclose(runner_1.model.fc.bias, runner_2.model.fc.bias)
+    shutil.rmtree(runner_1.work_dir)
+    shutil.rmtree(runner_2.work_dir)
+
+    # test has_batch_norm
+    model = ToyModel(with_norm=True)
+    optimizer_hook = GradientCumulativeOptimizerHook(
+        grad_clip=dict(max_norm=0.2), cumulative_iters=3)
+    assert optimizer_hook.has_batch_norm(model)
+
+
+@pytest.mark.skipif(
+    not torch.cuda.is_available(), reason='requires CUDA support')
+def test_gradient_cumulative_fp16_optimizer_hook():
+
+    class ToyModel(nn.Module):
+
+        def __init__(self):
+            super().__init__()
+            self.fp16_enabled = False
+            self.fc = nn.Linear(3, 2)
+            nn.init.constant_(self.fc.weight, 1.)
+            nn.init.constant_(self.fc.bias, 1.)
+
+        @auto_fp16(apply_to=('x', ))
+        def forward(self, x):
+            x = self.fc(x)
+            return x
+
+        def train_step(self, x, optimizer, **kwargs):
+            return dict(loss=self(x).mean(), num_samples=x.shape[0])
+
+        def val_step(self, x, optimizer, **kwargs):
+            return dict(loss=self(x).mean(), num_samples=x.shape[0])
+
+    def build_toy_runner(config=dict(type='EpochBasedRunner', max_epochs=3)):
+        model = ToyModel().cuda()
+        optimizer = torch.optim.SGD(model.parameters(), lr=0.02)
+        tmp_dir = tempfile.mkdtemp()
+
+        runner = build_runner(
+            config,
+            default_args=dict(
+                model=model,
+                work_dir=tmp_dir,
+                optimizer=optimizer,
+                logger=logging.getLogger(),
+                meta=dict()))
+        return runner
+
+    # test epoch based runner
+    data = torch.rand((6, 3)).cuda()
+    # optimize with cumulative_iters
+    loader_1 = DataLoader(data, batch_size=1)
+    runner_1 = build_toy_runner()
+    optimizer_hook = GradientCumulativeFp16OptimizerHook(
+        grad_clip=dict(max_norm=0.2), cumulative_iters=3)
+    runner_1.register_hook(optimizer_hook)
+    runner_1.run([loader_1], [('train', 1)])
+
+    # optimize without cumulative_iters
+    loader_2 = DataLoader(data, batch_size=3)
+    runner_2 = build_toy_runner()
+    optimizer_hook = Fp16OptimizerHook(grad_clip=dict(max_norm=0.2))
+    runner_2.register_hook(optimizer_hook)
+    runner_2.run([loader_2], [('train', 1)])
+
+    # test optimizer works well
+    assert (runner_1.model.fc.weight < 1).all()
+    assert (runner_1.model.fc.bias < 1).all()
+    # test optimizer with cumulative_iters gets the same results
+    assert torch.allclose(runner_1.model.fc.weight, runner_2.model.fc.weight)
+    assert torch.allclose(runner_1.model.fc.bias, runner_2.model.fc.bias)
+    shutil.rmtree(runner_1.work_dir)
+    shutil.rmtree(runner_2.work_dir)
+
+    # test iter based runner
+    data = torch.rand((8, 3)).cuda()
+    # optimize with cumulative_iters
+    loader_1 = DataLoader(data, batch_size=1)
+    runner_1 = build_toy_runner(dict(type='IterBasedRunner', max_iters=8))
+    optimizer_hook = GradientCumulativeFp16OptimizerHook(
+        grad_clip=dict(max_norm=0.2), cumulative_iters=3)
+    runner_1.register_hook(optimizer_hook)
+    runner_1.run([loader_1], [('train', 1)])
+
+    # optimize without cumulative_iters
+    loader_2_divisible = DataLoader(data[:6], batch_size=3)
+    loader_2_remainder = DataLoader(data[6:], batch_size=2)
+    runner_2 = build_toy_runner(dict(type='IterBasedRunner', max_iters=3))
+    optimizer_hook = Fp16OptimizerHook(grad_clip=dict(max_norm=0.2))
+    runner_2.register_hook(optimizer_hook)
+    runner_2.run([loader_2_divisible, loader_2_remainder], [('train', 2),
+                                                            ('train', 1)])
+
+    # test optimizer works well
+    assert (runner_1.model.fc.weight < 1).all()
+    assert (runner_1.model.fc.bias < 1).all()
+    # test optimizer with cumulative_iters gets the same results
+    assert torch.allclose(runner_1.model.fc.weight, runner_2.model.fc.weight)
+    assert torch.allclose(runner_1.model.fc.bias, runner_2.model.fc.bias)
+    shutil.rmtree(runner_1.work_dir)
+    shutil.rmtree(runner_2.work_dir)