fix models because the gradient clip strategy has been upgraded (#4515)

* fix models because the gradient clip strategy has been upgraded,test=develop

* fix models because the gradient clip strategy has been upgraded,test=develop

PaddleCV/ocr_recognition/crnn_ctc_model.py

@@ -114,7 +114,6 @@ def encoder_net(images,
                 num_classes,
                 rnn_hidden_size=200,
                 regularizer=None,
-                gradient_clip=None,
                 is_test=False,
                 use_cudnn=False):
     conv_features = ocr_convs(
@@ -130,16 +129,13 @@ def encoder_net(images,
 
     para_attr = fluid.ParamAttr(
         regularizer=regularizer,
-        gradient_clip=gradient_clip,
         initializer=fluid.initializer.Normal(0.0, 0.02))
     bias_attr = fluid.ParamAttr(
         regularizer=regularizer,
-        gradient_clip=gradient_clip,
         initializer=fluid.initializer.Normal(0.0, 0.02),
         learning_rate=2.0)
     bias_attr_nobias = fluid.ParamAttr(
         regularizer=regularizer,
-        gradient_clip=gradient_clip,
         initializer=fluid.initializer.Normal(0.0, 0.02))
 
     fc_1 = fluid.layers.fc(input=sliced_feature,

PaddleNLP/dialogue_domain_classification/nets.py

@@ -33,26 +33,21 @@ def textcnn_net_multi_label(data,
     """
     init_bound = 0.1
     initializer = fluid.initializer.Uniform(low=-init_bound, high=init_bound)
-    #gradient_clip = fluid.clip.GradientClipByNorm(10.0)
-    gradient_clip = None
     regularizer = fluid.regularizer.L2DecayRegularizer(
                                         regularization_coeff=1e-4)
     seg_param_attrs = fluid.ParamAttr(name="seg_weight",
                                   learning_rate=640.0,
                                   initializer=initializer,
-                                  gradient_clip=gradient_clip,
                                   trainable=True)
     fc_param_attrs_1 = fluid.ParamAttr(name="fc_weight_1",
                                                learning_rate=1.0,
                                                regularizer=regularizer,
                                                initializer=initializer,
-                                               gradient_clip=gradient_clip,
                                                trainable=True)
     fc_param_attrs_2 = fluid.ParamAttr(name="fc_weight_2",
                                                learning_rate=1.0,
                                                regularizer=regularizer,
                                                initializer=initializer,
-                                               gradient_clip=gradient_clip,
                                                trainable=True)
 
     if win_sizes is None:

PaddleRec/gru4rec/dy_graph/gru4rec_dy.py

@@ -408,7 +408,7 @@ def train_ptb_lm():
             if args.ce:
                 print("kpis\ttest_ppl\t%0.3f" % ppl[0])
 
-        grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(max_grad_norm)
+        grad_clip = fluid.clip.GradientClipByGlobalNorm(max_grad_norm)
         for epoch_id in range(max_epoch):
             ptb_model.train()
             total_loss = 0.0

dygraph/bert/optimization.py

@@ -134,7 +134,7 @@ class Optimizer(object):
         param_list = dict()
 
         clip_norm_thres = 1.0
-        #grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(clip_norm_thres)
+        #grad_clip = fluid.clip.GradientClipByGlobalNorm(clip_norm_thres)
 
         if use_data_parallel:
             loss = model.scale_loss(loss)

dygraph/ocr_recognition/train.py

@@ -74,7 +74,7 @@ def train(args):
             learning_rate = LR
 
         optimizer = fluid.optimizer.Adam(learning_rate=learning_rate, parameter_list=ocr_attention.parameters())
-        grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(args.gradient_clip)
+        grad_clip = fluid.clip.GradientClipByGlobalNorm(args.gradient_clip)
 
         train_reader = data_reader.data_reader(
             args.batch_size,

dygraph/ptb_lm/ptb_dy.py

@@ -32,8 +32,8 @@ import time
 
 from args import *
 
-#import fluid.dygraph_grad_clip as dygraph_clip
-#from fluid.dygraph_grad_clip  import *
+#import fluid.clip as clip
+#from fluid.clip  import *
 
 import sys
 if sys.version[0] == '2':
@@ -371,7 +371,7 @@ def train_ptb_lm():
 
         ce_time = []
         ce_ppl = []
-        grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(max_grad_norm)
+        grad_clip = fluid.clip.GradientClipByGlobalNorm(max_grad_norm)
         for epoch_id in range(max_epoch):
             ptb_model.train()
             total_loss = 0.0

dygraph/seq2seq/train.py

@@ -27,7 +27,7 @@ import contextlib
 
 import paddle
 import paddle.fluid as fluid
-from paddle.fluid.dygraph_grad_clip import GradClipByGlobalNorm
+from paddle.fluid.clip import GradientClipByGlobalNorm
 
 import reader
 
@@ -84,7 +84,7 @@ def main():
                 num_layers=num_layers,
                 init_scale=init_scale,
                 dropout=dropout)
-        gloabl_norm_clip = GradClipByGlobalNorm(max_grad_norm)
+        gloabl_norm_clip = GradientClipByGlobalNorm(max_grad_norm)
         lr = args.learning_rate
         opt_type = args.optimizer
         if opt_type == "sgd":