Add gradient accumulation, fix lr scheduler

colossalai/builder/builder.py

@@ -235,7 +235,7 @@ def build_optimizer_wrapper(config, optimizer, model=None):
         return OPTIMIZER_WRAPPERS.get_module(mod_type)(optimizer, **config_)
 
 
-def build_lr_scheduler(config, optimizer, total_steps, num_steps_per_epoch):
+def build_lr_scheduler(config, optimizer):
     """Returns a learning rate scheduler object of :class:`torch.optim.lr_scheduler` 
     constructed from `config`, `optimizer`, `total_steps` and `num_steps_per_epoch`.
 
@@ -255,8 +255,7 @@ def build_lr_scheduler(config, optimizer, total_steps, num_steps_per_epoch):
     config_ = config.copy()
     mod_type = config_.pop('type')
     # warmup epochs will overwrite warmup steps
-    if 'warmup_epochs' in config_:
+    # if 'warmup_epochs' in config_:
-        warmup_epochs = config_.pop('warmup_epochs')
+    #     warmup_epochs = config_.pop('warmup_epochs')
-        config_['warmup_steps'] = int(num_steps_per_epoch * warmup_epochs)
+    #     config_['warmup_steps'] = int(num_steps_per_epoch * warmup_epochs)
-    return LR_SCHEDULERS.get_module(mod_type)(optimizer, total_steps, num_steps_per_epoch=num_steps_per_epoch,
+    return LR_SCHEDULERS.get_module(mod_type)(optimizer, **config_)
-                                              **config_)

colossalai/engine/_base_engine.py

@@ -44,7 +44,9 @@ class Engine:
                  criterion: _Loss = None,
                  optimizer: Optimizer = None,
                  lr_scheduler: Optional[_LRScheduler] = None,
-                 schedule: BaseSchedule = None):
+                 schedule: BaseSchedule = None,
+                 gradient_accumulation: int = 1,
+                 lr_scheduler_step: str = 'epoch'):
         self.train_dataloader = train_dataloader
         self.test_dataloader = test_dataloader
         assert model is not None, "Engine requires a model"
@@ -54,6 +56,11 @@ class Engine:
         self.lr_scheduler = lr_scheduler
         self.schedule = schedule if schedule is not None \
             else NoPipelineSchedule()
+        self.grad_accum_size = gradient_accumulation
+        self.grad_accum_step = 0
+        self.lr_step = 0  # for epoch updating
+        if lr_scheduler_step != 'epoch':
+            self.lr_step = 1
         self._logger = get_global_dist_logger()
 
         # build gradient handler
@@ -89,8 +96,12 @@ class Engine:
             self._gradient_handlers.append(handler)
 
         self.schedule.initialize(self.train_dataloader, self.model,
-                                 self.criterion, self.optimizer,
+                                 self.criterion, self.optimizer)
-                                 self.lr_scheduler)
+        self.schedule.grad_accum = self.grad_accum_size
+        # add for robustness
+        if self.optimizer is None:
+            self.forward_only = True
+        else:
             self.forward_only = False
 
     def handle_gradient(self):
@@ -116,6 +127,7 @@ class Engine:
         """Returns the neural network model in the engine.
         """
         return self.model
+
     def get_optimizer(self):
         """Returns optimizier in the engine.
         """
@@ -146,7 +158,10 @@ class Engine:
     def get_lr(self):
         """Gets current learning rate.
         """
-        return self.schedule.get_lr()
+        if self.lr_scheduler is not None:
+            return self.lr_scheduler.get_lr()[0]
+        else:
+            return self.optimizer.param_groups[0]['lr']
 
     def step(self, return_loss=True):
         """A running step based on the schedule. Usually, it runs a training or
@@ -156,15 +171,27 @@ class Engine:
         :type return_loss: bool
         :return: (output, lablel, loss)
         """
-        self.schedule.zero_grad(forward_only=self.forward_only)
+        if not self.forward_only and self.grad_accum_step == 0:
+            self.schedule.zero_grad()
 
         output, label, loss = self.schedule.forward_backward_step(
             forward_only=self.forward_only, return_loss=return_loss)
 
         if not self.forward_only:
+            self.grad_accum_step += 1
+            if self.grad_accum_step == self.grad_accum_size:
                 # all reduce gradients
                 self.handle_gradient()
-
                 self.schedule.step()
+                if self.lr_scheduler is not None and self.lr_step:
+                    self.lr_scheduler.step()
+                self.grad_accum_step = 0
 
         return output, label, loss
+
+    def complete(self):
+        """Updating after a epoch.
+        """
+        self.schedule.consume_batch()
+        if self.lr_scheduler is not None and self.lr_step == 0:
+            self.lr_scheduler.step()

colossalai/engine/schedule/_base_schedule.py

@@ -15,6 +15,8 @@ class BaseSchedule(ABC):
     def __init__(self):
         self.initialized = False
         self.logger = get_global_dist_logger()
+        self.grad_accum = 1
+        self.training = False
 
     @property
     @abstractmethod
@@ -27,15 +29,13 @@ class BaseSchedule(ABC):
                    dataloader=None,
                    model=None,
                    criterion=None,
-                   optimizer=None,
+                   optimizer=None):
-                   lr_scheduler=None):
         """Initializes the schedule and set parameters before running.
 
         :param dataloader: DataLoader in training or evaluation
         :param model: The neural network model
         :param criterion: Criterion for calculating loss
         :param optimizer: Optimizer for updating the parameters
-        :param lr_scheduler: Learning rate scheduler in the process
         """
         self.dataloader = dataloader
         assert model is not None, "Schedule requires a model"
@@ -44,7 +44,6 @@ class BaseSchedule(ABC):
         self.criterion = criterion
         assert optimizer is not None, "Schedule requires an optimizer"
         self.optimizer = optimizer
-        self.lr_scheduler = lr_scheduler
         self.initialized = True
 
     def check_initialized(self):
@@ -66,6 +65,13 @@ class BaseSchedule(ABC):
         data, label = next(self.data_iter)
         return self._move_to_device(data), self._move_to_device(label)
 
+    def consume_batch(self):
+        while True:
+            try:
+                self.load_batch()
+            except StopIteration:
+                break
+
     def _move_to_device(self, data):
         if isinstance(data, (
                 tuple,
@@ -87,6 +93,7 @@ class BaseSchedule(ABC):
         :param mode: If True, the model will set as training mode. Otherwise, evaluation mode.
         """
         self.check_initialized()
+        self.training = mode
         if mode:
             self.model.train()
         else:
@@ -102,22 +109,11 @@ class BaseSchedule(ABC):
             self.check_initialized()
             self.optimizer.zero_grad()
 
-    def get_lr(self):
-        """Returns the current learning rate.
-        """
-        if self.lr_scheduler is not None:
-            return self.lr_scheduler.get_lr()[0]
-        else:
-            return self.optimizer.param_groups[0]['lr']
-
     def step(self):
         """Updates the parameters and learning rate with the optimizer.
         """
         self.check_initialized()
         self.optimizer.step()
-        # update lr scheduler
-        if self.lr_scheduler is not None:
-            self.lr_scheduler.step()
 
     @abstractmethod
     def forward_backward_step(self, forward_only=False, return_loss=True):

colossalai/engine/schedule/_no_pipeline.py

@@ -81,19 +81,20 @@ class NoPipelineSchedule(BaseSchedule):
 
     @property
     def num_steps(self):
-        return len(self.dataloader)
+        length = len(self.dataloader)
+        if self.training:
+            length -= length % self.grad_accum
+        return length
 
     def initialize(self,
-                   dataloader,
+                   dataloader=None,
-                   model,
+                   model=None,
-                   criterion,
+                   criterion=None,
-                   optimizer,
+                   optimizer=None):
-                   lr_scheduler=None):
         super().initialize(dataloader,
                            model,
                            criterion,
-                           optimizer,
+                           optimizer)
-                           lr_scheduler=lr_scheduler)
         if isinstance(self.optimizer, (ZeroRedundancyOptimizer_Level_2,
                                        ZeroRedundancyOptimizer_Level_3)):
             self.use_zero_level_2_3 = True
@@ -147,6 +148,7 @@ class NoPipelineSchedule(BaseSchedule):
                 output = (output,)
             if return_loss:
                 loss = self.criterion(*output, *label)
+        loss /= self.grad_accum
 
         if not forward_only:
             # backward
@@ -168,7 +170,7 @@ class NoPipelineSchedule(BaseSchedule):
                 loss.backward()
 
         if return_loss:
-            return output, label, loss
+            return output, label, loss * self.grad_accum
         else:
             return output, None, None
 
@@ -179,7 +181,3 @@ class NoPipelineSchedule(BaseSchedule):
             self._torch_amp_scaler.update()
         else:
             self.optimizer.step()
-
-        # update lr scheduler
-        if self.lr_scheduler is not None:
-            self.lr_scheduler.step()

colossalai/engine/schedule/_pipeline.py

@@ -119,19 +119,20 @@ class PipelineSchedule(BaseSchedule):
 
     @property
     def num_steps(self):
-        return len(self.dataloader)
+        length = len(self.dataloader)
+        if self.training:
+            length -= length % self.grad_accum
+        return length
 
     def initialize(self,
-                   dataloader,
+                   dataloader=None,
-                   model,
+                   model=None,
-                   criterion,
+                   criterion=None,
-                   optimizer,
+                   optimizer=None):
-                   lr_scheduler=None):
         super().initialize(dataloader,
                            model,
                            criterion,
-                           optimizer,
+                           optimizer)
-                           lr_scheduler=lr_scheduler)
         if isinstance(self.optimizer, (ZeroRedundancyOptimizer_Level_2,
                                        ZeroRedundancyOptimizer_Level_3)):
             raise TypeError(
@@ -163,7 +164,7 @@ class PipelineSchedule(BaseSchedule):
             if return_loss:
                 input_tensor, label = self.load_micro_batch()
                 loss_reduced = self.criterion(output_tensor, *
-                label) / self.num_microbatches
+                label) / (self.num_microbatches * self.grad_accum)
                 return_tensors.append(
                     tuple((output_tensor, label[0], loss_reduced)))
                 return loss_reduced
@@ -309,7 +310,7 @@ class PipelineSchedule(BaseSchedule):
                 output, label, loss = tuple(map(list, zip(*return_tensors)))
                 return (torch.cat(output, dim=0),
                         torch.cat(label, dim=0),
-                        sum(loss))
+                        sum(loss) * self.grad_accum)
             else:
                 return tuple((torch.cat(return_tensors, dim=0), None, None))
         else:

colossalai/initialize.py

@@ -339,16 +339,15 @@ def initialize(config: Union[str, dict] = None,
 
     lr_scheduler = None
     if hasattr(gpc.config, 'lr_scheduler'):
-        if hasattr(gpc.config, 'num_steps'):
+        # if hasattr(gpc.config, 'num_steps'):
-            total_steps = gpc.config.num_steps
+        #     total_steps = gpc.config.num_steps
-        elif hasattr(gpc.config, 'num_epochs'):
+        # elif hasattr(gpc.config, 'num_epochs'):
-            total_steps = int(gpc.config.num_epochs * len(train_dataloader))
+        #     total_steps = int(gpc.config.num_epochs * len(train_dataloader))
-        else:
+        # else:
-            raise Exception(
+        #     raise Exception(
-                'Please specify training stopping criterion num_steps or num_epochs in your configuration.'
+        #         'Please specify training stopping criterion num_steps or num_epochs in your configuration.'
-            )
+        #     )
-        lr_scheduler = build_lr_scheduler(gpc.config.lr_scheduler, optimizer,
+        lr_scheduler = build_lr_scheduler(gpc.config.lr_scheduler, optimizer)
-                                          total_steps, len(train_dataloader))
         logger.info('Learning rate scheduler is created', ranks=[0])
 
     # pipeline or no pipeline schedule

colossalai/trainer/_trainer.py

@@ -147,6 +147,7 @@ class Trainer:
             if self.exceed_max_step():
                 # stop when max iter is reached
                 break
+        self._engine.complete()
         self._timer.stop('train-epoch', keep_in_history=True)
         self.call_hooks('after_train_epoch')
         self._timer.reset('train-step')

configs/vit/vit_2d.py

@@ -8,10 +8,10 @@ BATCH_SIZE = 512
 IMG_SIZE = 32
 PATCH_SIZE = 4
 DIM = 512
-NUM_ATTENTION_HEADS = 8
+NUM_ATTENTION_HEADS = 2
 SUMMA_DIM = 2
 NUM_CLASSES = 10
-DEPTH = 6
+DEPTH = 1
 
 train_data = dict(
     dataset=dict(
@@ -127,14 +127,14 @@ hooks = [
     dict(type='LogMetricByEpochHook'),
     dict(type='Accuracy2DHook'),
     dict(type='LossHook'),
-    dict(type='TensorboardHook', log_dir='./tfb_logs'),
+    # dict(type='TensorboardHook', log_dir='./tfb_logs'),
     # dict(type='SaveCheckpointHook', interval=5, checkpoint_dir='./ckpt'),
     # dict(type='LoadCheckpointHook', epoch=20, checkpoint_dir='./ckpt')
 ]
 
 parallel = dict(
     pipeline=dict(size=1),
-    tensor=dict(size=4, mode='2d'),
+    tensor=dict(size=1, mode='2d'),
 )
 
 # for fp16 training
@@ -146,7 +146,8 @@ parallel = dict(
 
 lr_scheduler = dict(
     type='LinearWarmupLR',
-    warmup_epochs=5
+    total_steps=60,
+    warmup_steps=5
 )
 
 # only needed when pipeline parallel is used

examples/run_trainer.py

@@ -17,7 +17,8 @@ def run_trainer():
         criterion=criterion,
         optimizer=optimizer,
         lr_scheduler=lr_scheduler,
-        schedule=schedule
+        schedule=schedule,
+        gradient_accumulation=5,
     )
     logger.info("engine is built", ranks=[0])