fix gru_dygraph (#4248)

* fix gru_dygraph

* fix mmoe doc

PaddleRec/gru4rec/dy_graph/README.md

 # gru4rec 动态图实现
 
+# 环境配置
+paddle 1.7
+
+
+
 # 下载数据
 ```
 wget https://paddlerec.bj.bcebos.com/gru4rec/dy_graph/data_rsc15.tar
 tar xvf data_rsc15.tar
 ```
 
+# 数据格式
+数据格式及预处理处理同静态图相同。
+
 # 训练及预测
 
 ```

PaddleRec/gru4rec/dy_graph/gru4rec_dy.py

@@ -21,6 +21,7 @@ import paddle.fluid.core as core
 from paddle.fluid.dygraph.nn import Embedding
 import paddle.fluid.framework as framework
 from paddle.fluid.optimizer import SGDOptimizer
+from paddle.fluid.optimizer import AdagradOptimizer
 from paddle.fluid.dygraph.base import to_variable
 import numpy as np
 import six
@@ -67,7 +68,7 @@ class SimpleGRURNN(fluid.Layer):
                 dtype="float32",
                 default_initializer=fluid.initializer.UniformInitializer(
                     low=-self._init_scale, high=self._init_scale))
-            self.weight_1_arr.append(self.add_parameter('w_%d' % i, weight_1))
+            self.weight_1_arr.append(self.add_parameter('w1_%d' % i, weight_1))
             weight_2 = self.create_parameter(
                 attr=fluid.ParamAttr(
                     initializer=fluid.initializer.UniformInitializer(
@@ -76,7 +77,7 @@ class SimpleGRURNN(fluid.Layer):
                 dtype="float32",
                 default_initializer=fluid.initializer.UniformInitializer(
                     low=-self._init_scale, high=self._init_scale))
-            self.weight_2_arr.append(self.add_parameter('w_%d' % i, weight_2))
+            self.weight_2_arr.append(self.add_parameter('w2_%d' % i, weight_2))
             weight_3 = self.create_parameter(
                 attr=fluid.ParamAttr(
                     initializer=fluid.initializer.UniformInitializer(
@@ -85,7 +86,7 @@ class SimpleGRURNN(fluid.Layer):
                 dtype="float32",
                 default_initializer=fluid.initializer.UniformInitializer(
                     low=-self._init_scale, high=self._init_scale))
-            self.weight_3_arr.append(self.add_parameter('w_%d' % i, weight_3))
+            self.weight_3_arr.append(self.add_parameter('w3_%d' % i, weight_3))
             bias_1 = self.create_parameter(
                 attr=fluid.ParamAttr(
                     initializer=fluid.initializer.UniformInitializer(
@@ -93,7 +94,7 @@ class SimpleGRURNN(fluid.Layer):
                 shape=[self._hidden_size * 2],
                 dtype="float32",
                 default_initializer=fluid.initializer.Constant(0.0))
-            self.bias_1_arr.append(self.add_parameter('b_%d' % i, bias_1))
+            self.bias_1_arr.append(self.add_parameter('b1_%d' % i, bias_1))
             bias_2 = self.create_parameter(
                 attr=fluid.ParamAttr(
                     initializer=fluid.initializer.UniformInitializer(
@@ -101,7 +102,7 @@ class SimpleGRURNN(fluid.Layer):
                 shape=[self._hidden_size * 1],
                 dtype="float32",
                 default_initializer=fluid.initializer.Constant(0.0))
-            self.bias_2_arr.append(self.add_parameter('b_%d' % i, bias_2))
+            self.bias_2_arr.append(self.add_parameter('b2_%d' % i, bias_2))
 
     def forward(self, input_embedding, init_hidden=None):
         hidden_array = []
@@ -278,10 +279,10 @@ def train_ptb_lm():
         init_scale = 0.1
         max_grad_norm = 5.0
         epoch_start_decay = 10
-        max_epoch = 3
+        max_epoch = 5
         dropout = 0.0
         lr_decay = 0.5
-        base_learning_rate = 1.0
+        base_learning_rate = 0.05
     elif model_type == "medium":
         num_layers = 2
         batch_size = 20
@@ -353,15 +354,22 @@ def train_ptb_lm():
         log_interval = total_batch_size // 20
 
         bd = []
-        lr_arr = [1.0]
+        lr_arr = [base_learning_rate]
         for i in range(1, max_epoch):
             bd.append(total_batch_size * i)
             new_lr = base_learning_rate * (lr_decay**
                                            max(i + 1 - epoch_start_decay, 0.0))
             lr_arr.append(new_lr)
 
-        sgd = SGDOptimizer(learning_rate=fluid.layers.piecewise_decay(
-            boundaries=bd, values=lr_arr))
+        sgd = AdagradOptimizer(
+            parameter_list=ptb_model.parameters(),
+            learning_rate=fluid.layers.piecewise_decay(
+                boundaries=bd, values=lr_arr))
+
+        print("parameters:--------------------------------")
+        for para in ptb_model.parameters():
+            print(para.name)
+        print("parameters:--------------------------------")
 
         def eval(model, data):
             print("begion to eval")
@@ -450,7 +458,7 @@ def train_ptb_lm():
             print("Saved model to: %s.\n" % save_model_dir)
             eval(ptb_model, test_data)
 
-        eval(ptb_model, test_data)
+        #eval(ptb_model, test_data)
 
 
 train_ptb_lm()

PaddleRec/multi-task/MMoE/README.md

@@ -8,6 +8,9 @@ MMoE是经典的多任务（multi-task）模型，原论文[Modeling Task Relati
 
 我们基于实际工业界场景实现了MMoE的核心思想。
 
+## 配置
+1.6 及以上
+
 ## 数据
 
 我们采用了随机数据作为训练数据，可以根据自己的数据调整data部分。

PaddleRec/multi-task/MMoE/mmoe_train.py

 import paddle.fluid as fluid
 import numpy as np
-
-dict_dim = 1000
-emb_dim = 64
+import time
+from args import *
 
 
 def fc_layers(input, layers, acts, prefix):
@@ -59,7 +58,7 @@ def mmoe_layer(inputs, expert_num=8, gate_num=3):
     return outs
 
 
-def model():
+def model(dict_dim, emb_dim):
     label_like = fluid.layers.data(
         name="label_like",
         shape=[-1, 1],
@@ -116,13 +115,18 @@ def model():
     return cost, [a_data, label_like, label_comment, label_share]
 
 
-batch_size = 5
+args = parse_args()
+batch_size = args.batch_size
+dict_dim = args.dict_dim
+emb_dim = args.emb_dim
+
+print("batch_size:[%d], dict_dim:[%d], emb_dim:[%d], learning_rate:[%.4f]" %
+      (batch_size, dict_dim, emb_dim, args.base_lr))
 
-loss, data_list = model()
-sgd = fluid.optimizer.SGD(learning_rate=0.001)
+loss, data_list = model(dict_dim, emb_dim)
+sgd = fluid.optimizer.SGD(learning_rate=args.base_lr)
 sgd.minimize(loss)
-use_cuda = True
-place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
 feeder = fluid.DataFeeder(feed_list=data_list, place=place)
 exe = fluid.Executor(place)
 exe.run(fluid.default_startup_program())
@@ -131,6 +135,7 @@ for batch_id in range(100):
         np.random.randint(
             2, size=(batch_size, 1)).astype('int64') for i in range(4)
     ]
+    begin = time.time()
     loss_data, = exe.run(fluid.default_main_program(),
                          feed={
                              "a": data[0],
@@ -139,4 +144,6 @@ for batch_id in range(100):
                              "label_share": data[3]
                          },
                          fetch_list=[loss.name])
-    print(batch_id, " loss:", float(np.array(loss_data)))
+    end = time.time()
+    print("batch_id:[%d], loss:[%.5f], batch_time:[%.5f s]" %
+          (batch_id, float(np.array(loss_data)), end - begin))