!4799 Improve performance of bert by mimic higher batch size

Merge pull request !4799 from shibeiji/bert_accu_steps

mindspore/ops/functional.py

@@ -63,6 +63,7 @@ check_bprop = P.CheckBprop()
 equal = P.Equal()
 not_equal = P.NotEqual()
 assign_sub = P.AssignSub()
+assign_add = P.AssignAdd()
 assign = P.Assign()
 square = P.Square()
 sqrt = P.Sqrt()

model_zoo/official/nlp/bert/README.md

@@ -123,6 +123,7 @@ usage: run_pretrain.py  [--distribute DISTRIBUTE] [--epoch_size N] [----device_n
                         [--enable_save_ckpt ENABLE_SAVE_CKPT] [--device_target DEVICE_TARGET]
                         [--enable_lossscale ENABLE_LOSSSCALE] [--do_shuffle DO_SHUFFLE]
                         [--enable_data_sink ENABLE_DATA_SINK] [--data_sink_steps N] 
+                        [--accumulation_steps N]
                         [--save_checkpoint_path SAVE_CHECKPOINT_PATH]
                         [--load_checkpoint_path LOAD_CHECKPOINT_PATH]
                         [--save_checkpoint_steps N] [--save_checkpoint_num N] 
@@ -139,6 +140,7 @@ options:
     --do_shuffle               enable shuffle: "true" | "false", default is "true"
     --enable_data_sink         enable data sink: "true" | "false", default is "true"
     --data_sink_steps          set data sink steps: N, default is 1
+    --accumulation_steps       accumulate gradients N times before weight update: N, default is 1
     --save_checkpoint_path     path to save checkpoint files: PATH, default is ""
     --load_checkpoint_path     path to load checkpoint files: PATH, default is ""
     --save_checkpoint_steps    steps for saving checkpoint files: N, default is 1000

model_zoo/official/nlp/bert/run_pretrain.py

@@ -30,7 +30,8 @@ from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, TimeMoni
 from mindspore.train.serialization import load_checkpoint, load_param_into_net
 from mindspore.nn.optim import Lamb, Momentum, AdamWeightDecay
 from mindspore import log as logger
-from src import BertNetworkWithLoss, BertTrainOneStepCell, BertTrainOneStepWithLossScaleCell
+from src import BertNetworkWithLoss, BertTrainOneStepCell, BertTrainOneStepWithLossScaleCell, \
+                BertTrainAccumulateStepsWithLossScaleCell
 from src.dataset import create_bert_dataset
 from src.config import cfg, bert_net_cfg
 from src.utils import LossCallBack, BertLearningRate
@@ -51,6 +52,8 @@ def run_pretrain():
     parser.add_argument("--do_shuffle", type=str, default="true", help="Enable shuffle for dataset, default is true.")
     parser.add_argument("--enable_data_sink", type=str, default="true", help="Enable data sink, default is true.")
     parser.add_argument("--data_sink_steps", type=int, default="1", help="Sink steps for each epoch, default is 1.")
+    parser.add_argument("--accumulation_steps", type=int, default="1",
+                        help="Accumulating gradients N times before weight update, default is 1.")
     parser.add_argument("--save_checkpoint_path", type=str, default="", help="Save checkpoint path")
     parser.add_argument("--load_checkpoint_path", type=str, default="", help="Load checkpoint file path")
     parser.add_argument("--save_checkpoint_steps", type=int, default=1000, help="Save checkpoint steps, "
@@ -98,6 +101,16 @@ def run_pretrain():
         logger.warning('Gpu only support fp32 temporarily, run with fp32.')
         bert_net_cfg.compute_type = mstype.float32
 
+    if args_opt.accumulation_steps > 1:
+        logger.info("accumulation steps: {}".format(args_opt.accumulation_steps))
+        logger.info("global batch size: {}".format(bert_net_cfg.batch_size * args_opt.accumulation_steps))
+        if args_opt.enable_data_sink == "true":
+            args_opt.data_sink_steps *= args_opt.accumulation_steps
+            logger.info("data sink steps: {}".format(args_opt.data_sink_steps))
+        if args_opt.enable_save_ckpt == "true":
+            args_opt.save_checkpoint_steps *= args_opt.accumulation_steps
+            logger.info("save checkpoint steps: {}".format(args_opt.save_checkpoint_steps))
+
     ds = create_bert_dataset(device_num, rank, args_opt.do_shuffle, args_opt.data_dir, args_opt.schema_dir)
     net_with_loss = BertNetworkWithLoss(bert_net_cfg, True)
 
@@ -157,8 +170,15 @@ def run_pretrain():
         update_cell = DynamicLossScaleUpdateCell(loss_scale_value=cfg.loss_scale_value,
                                                  scale_factor=cfg.scale_factor,
                                                  scale_window=cfg.scale_window)
-        net_with_grads = BertTrainOneStepWithLossScaleCell(net_with_loss, optimizer=optimizer,
-                                                           scale_update_cell=update_cell)
+
+        if args_opt.accumulation_steps <= 1:
+            net_with_grads = BertTrainOneStepWithLossScaleCell(net_with_loss, optimizer=optimizer,
+                                                               scale_update_cell=update_cell)
+        else:
+            accumulation_steps = args_opt.accumulation_steps
+            net_with_grads = BertTrainAccumulateStepsWithLossScaleCell(net_with_loss, optimizer=optimizer,
+                                                                       scale_update_cell=update_cell,
+                                                                       accumulation_steps=accumulation_steps)
     else:
         net_with_grads = BertTrainOneStepCell(net_with_loss, optimizer=optimizer)
 

model_zoo/official/nlp/bert/scripts/ascend_distributed_launcher/hyper_parameter_config.ini

@@ -6,6 +6,7 @@ enable_lossscale=true
 do_shuffle=true
 enable_data_sink=true
 data_sink_steps=100
+accumulation_steps=1
 save_checkpoint_path=./checkpoint/
 save_checkpoint_steps=10000
-save_checkpoint_num=1
+save_checkpoint_num=1

model_zoo/official/nlp/bert/scripts/run_standalone_pretrain_ascend.sh

@@ -39,6 +39,7 @@ python ${PROJECT_DIR}/../run_pretrain.py  \
     --do_shuffle="true" \
     --enable_data_sink="true" \
     --data_sink_steps=1 \
+    --accumulation_steps=1 \
     --load_checkpoint_path="" \
     --save_checkpoint_steps=10000 \
     --save_checkpoint_num=1 \

model_zoo/official/nlp/bert/src/__init__.py

@@ -15,7 +15,8 @@
 """Bert Init."""
 from .bert_for_pre_training import BertNetworkWithLoss, BertPreTraining, \
     BertPretrainingLoss, GetMaskedLMOutput, GetNextSentenceOutput, \
-    BertTrainOneStepCell, BertTrainOneStepWithLossScaleCell
+    BertTrainOneStepCell, BertTrainOneStepWithLossScaleCell, \
+    BertTrainAccumulateStepsWithLossScaleCell
 from .bert_model import BertAttention, BertConfig, BertEncoderCell, BertModel, \
     BertOutput, BertSelfAttention, BertTransformer, EmbeddingLookup, \
     EmbeddingPostprocessor, RelaPosEmbeddingsGenerator, RelaPosMatrixGenerator, \
@@ -23,7 +24,8 @@ from .bert_model import BertAttention, BertConfig, BertEncoderCell, BertModel, \
 
 __all__ = [
     "BertNetworkWithLoss", "BertPreTraining", "BertPretrainingLoss",
-    "GetMaskedLMOutput", "GetNextSentenceOutput", "BertTrainOneStepCell", "BertTrainOneStepWithLossScaleCell",
+    "GetMaskedLMOutput", "GetNextSentenceOutput", "BertTrainOneStepCell",
+    "BertTrainOneStepWithLossScaleCell", "BertTrainAccumulateStepsWithLossScaleCell",
     "BertAttention", "BertConfig", "BertEncoderCell", "BertModel", "BertOutput",
     "BertSelfAttention", "BertTransformer", "EmbeddingLookup",
     "EmbeddingPostprocessor", "RelaPosEmbeddingsGenerator",

model_zoo/official/nlp/bert/src/bert_for_pre_training.py

@@ -438,3 +438,164 @@ class BertTrainOneStepWithLossScaleCell(nn.Cell):
             succ = self.optimizer(grads)
         ret = (loss, cond, scaling_sens)
         return F.depend(ret, succ)
+
+
+cast = P.Cast()
+update_accu_grads = C.MultitypeFuncGraph("update_accu_grads")
+
+
+@update_accu_grads.register("Tensor", "Tensor")
+def _update_accu_grads(accu_grad, grad):
+    succ = True
+    return F.depend(succ, F.assign_add(accu_grad, cast(grad, mstype.float32)))
+
+
+zeroslike = P.ZerosLike()
+reset_accu_grads = C.MultitypeFuncGraph("reset_accu_grads")
+
+
+@reset_accu_grads.register("Tensor")
+def _reset_accu_grads(accu_grad):
+    succ = True
+    return F.depend(succ, F.assign(accu_grad, zeroslike(accu_grad)))
+
+
+class BertTrainAccumulateStepsWithLossScaleCell(nn.Cell):
+    """
+    Encapsulation class of bert network training.
+
+    Append an optimizer to the training network after that the construct
+    function can be called to create the backward graph. To mimic higher batch size, gradients are
+    accumulated N times before weight update.
+
+    Args:
+        network (Cell): The training network. Note that loss function should have been added.
+        optimizer (Optimizer): Optimizer for updating the weights.
+        scale_update_cell (Cell): Cell to do the loss scale. Default: None.
+        accumulation_steps (int): Number of accumulation steps before gradient update. The global batch size =
+                                batch_size * accumulation_steps. Default: 1.
+    """
+    def __init__(self, network, optimizer, scale_update_cell=None, accumulation_steps=1):
+        super(BertTrainAccumulateStepsWithLossScaleCell, self).__init__(auto_prefix=False)
+        self.network = network
+        self.weights = optimizer.parameters
+        self.optimizer = optimizer
+        self.accumulation_steps = accumulation_steps
+        self.one = Tensor(np.array([1]).astype(np.int32))
+        self.zero = Tensor(np.array([0]).astype(np.int32))
+        self.local_step = Parameter(initializer(0, [1], mstype.int32), name="local_step")
+        self.accu_grads = self.weights.clone(prefix="accu_grads", init='zeros')
+        self.accu_overflow = Parameter(initializer(0, [1], mstype.int32), name="accu_overflow")
+        self.loss = Parameter(initializer(0, [1], mstype.float32), name="accu_loss")
+
+        self.grad = C.GradOperation('grad',
+                                    get_by_list=True,
+                                    sens_param=True)
+        self.reducer_flag = False
+        self.parallel_mode = context.get_auto_parallel_context("parallel_mode")
+        if self.parallel_mode in [ParallelMode.DATA_PARALLEL, ParallelMode.HYBRID_PARALLEL]:
+            self.reducer_flag = True
+        self.grad_reducer = F.identity
+        self.degree = 1
+        if self.reducer_flag:
+            self.degree = get_group_size()
+            self.grad_reducer = DistributedGradReducer(optimizer.parameters, False, self.degree)
+        self.is_distributed = (self.parallel_mode != ParallelMode.STAND_ALONE)
+        self.overflow_reducer = F.identity
+        if self.is_distributed:
+            self.overflow_reducer = P.AllReduce()
+        self.cast = P.Cast()
+        self.alloc_status = P.NPUAllocFloatStatus()
+        self.get_status = P.NPUGetFloatStatus()
+        self.clear_before_grad = P.NPUClearFloatStatus()
+        self.reduce_sum = P.ReduceSum(keep_dims=False)
+        self.base = Tensor(1, mstype.float32)
+        self.less_equal = P.LessEqual()
+        self.logical_or = P.LogicalOr()
+        self.not_equal = P.NotEqual()
+        self.select = P.Select()
+        self.reshape = P.Reshape()
+        self.hyper_map = C.HyperMap()
+        self.loss_scale = None
+        self.loss_scaling_manager = scale_update_cell
+        if scale_update_cell:
+            self.loss_scale = Parameter(Tensor(scale_update_cell.get_loss_scale(), dtype=mstype.float32),
+                                        name="loss_scale")
+
+    @C.add_flags(has_effect=True)
+    def construct(self,
+                  input_ids,
+                  input_mask,
+                  token_type_id,
+                  next_sentence_labels,
+                  masked_lm_positions,
+                  masked_lm_ids,
+                  masked_lm_weights,
+                  sens=None):
+        """Defines the computation performed."""
+        weights = self.weights
+        loss = self.network(input_ids,
+                            input_mask,
+                            token_type_id,
+                            next_sentence_labels,
+                            masked_lm_positions,
+                            masked_lm_ids,
+                            masked_lm_weights)
+        if sens is None:
+            scaling_sens = self.loss_scale
+        else:
+            scaling_sens = sens
+
+        # update accumulation parameters
+        is_accu_step = self.not_equal(self.local_step, self.accumulation_steps)
+        self.local_step = self.select(is_accu_step, self.local_step + self.one, self.one)
+        self.loss = self.select(is_accu_step, self.loss + loss, loss)
+        mean_loss = self.loss / self.local_step
+        is_accu_step = self.not_equal(self.local_step, self.accumulation_steps)
+
+        # alloc status and clear should be right before gradoperation
+        init = self.alloc_status()
+        self.clear_before_grad(init)
+        grads = self.grad(self.network, weights)(input_ids,
+                                                 input_mask,
+                                                 token_type_id,
+                                                 next_sentence_labels,
+                                                 masked_lm_positions,
+                                                 masked_lm_ids,
+                                                 masked_lm_weights,
+                                                 self.cast(scaling_sens,
+                                                           mstype.float32))
+
+        accu_succ = self.hyper_map(update_accu_grads, self.accu_grads, grads)
+        mean_loss = F.depend(mean_loss, accu_succ)
+
+        self.get_status(init)
+        flag_sum = self.reduce_sum(init, (0,))
+        overflow = self.less_equal(self.base, flag_sum)
+        overflow = self.logical_or(self.not_equal(self.accu_overflow, self.zero), overflow)
+        accu_overflow = self.select(overflow, self.one, self.zero)
+        self.accu_overflow = self.select(is_accu_step, accu_overflow, self.zero)
+
+        if is_accu_step:
+            succ = False
+        else:
+            # apply grad reducer on grads
+            grads = self.grad_reducer(self.accu_grads)
+            scaling = scaling_sens * self.degree * self.accumulation_steps
+            grads = self.hyper_map(F.partial(grad_scale, scaling), grads)
+            grads = self.hyper_map(F.partial(clip_grad, GRADIENT_CLIP_TYPE, GRADIENT_CLIP_VALUE), grads)
+            accu_overflow = self.overflow_reducer(accu_overflow)
+            F.control_depend(grads, accu_overflow)
+            overflow = self.less_equal(self.base, accu_overflow)
+            accu_succ = self.hyper_map(reset_accu_grads, self.accu_grads)
+            overflow = F.depend(overflow, accu_succ)
+            overflow = self.reshape(overflow, (()))
+            if sens is None:
+                overflow = self.loss_scaling_manager(self.loss_scale, overflow)
+            if overflow:
+                succ = False
+            else:
+                succ = self.optimizer(grads)
+
+        ret = (mean_loss, overflow, scaling_sens)
+        return F.depend(ret, succ)

model_zoo/official/nlp/bert/src/config.py

@@ -50,7 +50,7 @@ cfg = edict({
 
 '''
 Including two kinds of network: \
-base: Goole BERT-base(the base version of BERT model).
+base: Google BERT-base(the base version of BERT model).
 large: BERT-NEZHA(a Chinese pretrained language model developed by Huawei, which introduced a improvement of \
        Functional Relative Posetional Encoding as an effective positional encoding scheme).
 '''