!5632 Add clip_by_global_nrom in bert

Merge pull request !5632 from chenhaozhe/add-global-norm-to-bert

model_zoo/official/nlp/bert/run_pretrain.py

@@ -179,12 +179,14 @@ def run_pretrain():
 
         if args_opt.accumulation_steps <= 1:
             net_with_grads = BertTrainOneStepWithLossScaleCell(net_with_loss, optimizer=optimizer,
-                                                               scale_update_cell=update_cell)
+                                                               scale_update_cell=update_cell,
+                                                               enable_global_norm=cfg.enable_global_norm)
         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)
+                                                                       accumulation_steps=accumulation_steps,
+                                                                       enable_global_norm=cfg.enable_global_norm)
     else:
         net_with_grads = BertTrainOneStepCell(net_with_loss, optimizer=optimizer)
 

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

@@ -29,6 +29,7 @@ from mindspore.communication.management import get_group_size
 from mindspore import context
 from mindspore.ops import _selected_ops
 from .bert_model import BertModel
+from .utils import ClipByGlobalNorm
 
 GRADIENT_CLIP_TYPE = 1
 GRADIENT_CLIP_VALUE = 1.0
@@ -348,11 +349,12 @@ class BertTrainOneStepWithLossScaleCell(nn.Cell):
         optimizer (Optimizer): Optimizer for updating the weights.
         scale_update_cell (Cell): Cell to do the loss scale. Default: None.
     """
-    def __init__(self, network, optimizer, scale_update_cell=None):
+    def __init__(self, network, optimizer, scale_update_cell=None, enable_global_norm=False):
         super(BertTrainOneStepWithLossScaleCell, self).__init__(auto_prefix=False)
         self.network = network
         self.weights = optimizer.parameters
         self.optimizer = optimizer
+        self.enable_global_norm = enable_global_norm
         self.grad = C.GradOperation(get_by_list=True,
                                     sens_param=True)
         self.reducer_flag = False
@@ -419,7 +421,10 @@ class BertTrainOneStepWithLossScaleCell(nn.Cell):
         # apply grad reducer on grads
         grads = self.grad_reducer(grads)
         grads = self.hyper_map(F.partial(grad_scale, scaling_sens * self.degree), grads)
-        grads = self.hyper_map(F.partial(clip_grad, GRADIENT_CLIP_TYPE, GRADIENT_CLIP_VALUE), grads)
+        if self.enable_global_norm:
+            grads = ClipByGlobalNorm()(grads)
+        else:
+            grads = self.hyper_map(F.partial(clip_grad, GRADIENT_CLIP_TYPE, GRADIENT_CLIP_VALUE), grads)
         self.get_status(init)
         flag_sum = self.reduce_sum(init, (0,))
         if self.is_distributed:
@@ -474,12 +479,13 @@ class BertTrainAccumulateStepsWithLossScaleCell(nn.Cell):
         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):
+    def __init__(self, network, optimizer, scale_update_cell=None, accumulation_steps=1, enable_global_norm=False):
         super(BertTrainAccumulateStepsWithLossScaleCell, self).__init__(auto_prefix=False)
         self.network = network
         self.weights = optimizer.parameters
         self.optimizer = optimizer
         self.accumulation_steps = accumulation_steps
+        self.enable_global_norm = enable_global_norm
         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")
@@ -580,7 +586,10 @@ class BertTrainAccumulateStepsWithLossScaleCell(nn.Cell):
             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)
+            if self.enable_global_norm:
+                grads = ClipByGlobalNorm()(grad)
+            else:
+                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)

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

@@ -24,6 +24,7 @@ cfg = edict({
     'scale_factor': 2,
     'scale_window': 1000,
     'optimizer': 'Lamb',
+    'enable_global_norm': False,
     'AdamWeightDecay': edict({
         'learning_rate': 3e-5,
         'end_learning_rate': 0.0,
@@ -115,6 +116,5 @@ if cfg.bert_network == 'large':
         input_mask_from_dataset=True,
         token_type_ids_from_dataset=True,
         dtype=mstype.float32,
-        compute_type=mstype.float16,
-        enable_fused_layernorm=True
+        compute_type=mstype.float16
     )

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

@@ -23,12 +23,62 @@ import numpy as np
 import mindspore.nn as nn
 from mindspore import log as logger
 from mindspore.ops import operations as P
+from mindspore.ops import functional as F
+from mindspore.ops import composite as C
 from mindspore.common.tensor import Tensor
 from mindspore.common import dtype as mstype
 from mindspore.train.callback import Callback
 from mindspore.nn.learning_rate_schedule import LearningRateSchedule, PolynomialDecayLR, WarmUpLR
 
 
+get_square_sum = C.MultitypeFuncGraph("get_square_sum")
+@get_square_sum.register("Tensor")
+def _get_square_sum(grad):
+    norm = P.ReduceSum(False)(F.square(grad), ())
+    norm = F.expand_dims(F.cast(norm, mstype.float32), 0)
+    return norm
+
+
+apply_global_norm = C.MultitypeFuncGraph("apply_global_norm")
+@apply_global_norm.register("Tensor", "Tensor", "Tensor")
+def _apply_global_norm(clip_norm, global_norm, grad):
+    grad = grad * clip_norm / global_norm
+    return grad
+
+
+class GlobalNorm(nn.Cell):
+    """
+    Calculate the global norm value of given tensors
+    """
+    def __init__(self):
+        super(GlobalNorm, self).__init__()
+        self.norm = nn.Norm()
+        self.hyper_map = C.HyperMap()
+
+    def construct(self, grads):
+        square_sum = self.hyper_map(get_square_sum, grads)
+        global_norms = F.sqrt(F.addn(square_sum) / F.scalar_to_array(len(square_sum)))
+        return global_norms
+
+
+class ClipByGlobalNorm(nn.Cell):
+    """
+    Clip grads by global norm
+    """
+    def __init__(self, clip_norm=1.0):
+        super(ClipByGlobalNorm, self).__init__()
+        self.global_norm = GlobalNorm()
+        self.clip_norm = Tensor([clip_norm], mstype.float32)
+        self.hyper_map = C.HyperMap()
+
+    def construct(self, grads):
+        global_norm = self.global_norm(grads)
+        cond = P.GreaterEqual()(global_norm, self.clip_norm)
+        global_norm = F.select(cond, global_norm, self.clip_norm)
+        grads = self.hyper_map(F.partial(apply_global_norm, self.clip_norm, global_norm), grads)
+        return grads
+
+
 class CrossEntropyCalculation(nn.Cell):
     """
     Cross Entropy loss