fix double softmax ,fix test function and change default config of pairwise (#2303)

PaddleNLP/models/matching/losses/softmax_cross_entropy_loss.py

@@ -3,6 +3,7 @@ softmax loss
 """
 
 import sys
+import paddle.fluid as fluid
 
 sys.path.append("../../../")
 import models.matching.paddle_layers as layers
@@ -23,8 +24,7 @@ class SoftmaxCrossEntropyLoss(object):
         """
         compute loss
         """
-        softmax_with_cross_entropy = layers.SoftmaxWithCrossEntropyLayer()
         reduce_mean = layers.ReduceMeanLayer()
-        cost = softmax_with_cross_entropy.ops(input, label)
+        cost = fluid.layers.cross_entropy(input=input, label=label)
         avg_cost = reduce_mean.ops(cost)
         return avg_cost

PaddleNLP/models/matching/mm_dnn.py

@@ -49,10 +49,10 @@ class MMDNN(object):
             input=input,
             size=[self.vocab_size, self.emb_size],
             padding_idx=(0 if zero_pad else None),
-            param_attr=fluid.ParamAttr(name="word_embedding",
-                                       initializer=fluid.initializer.Xavier()))
+            param_attr=fluid.ParamAttr(
+                name="word_embedding", initializer=fluid.initializer.Xavier()))
         if scale:
-            emb = emb * (self.emb_size ** 0.5)
+            emb = emb * (self.emb_size**0.5)
         return emb
 
     def bi_dynamic_lstm(self, input, hidden_size):
@@ -64,7 +64,9 @@ class MMDNN(object):
                                      param_attr=fluid.ParamAttr(name="fw_fc.w"),
                                      bias_attr=False)
         forward, _ = fluid.layers.dynamic_lstm(
-            input=fw_in_proj, size=4 * hidden_size, is_reverse=False,
+            input=fw_in_proj,
+            size=4 * hidden_size,
+            is_reverse=False,
             param_attr=fluid.ParamAttr(name="forward_lstm.w"),
             bias_attr=fluid.ParamAttr(name="forward_lstm.b"))
 
@@ -73,7 +75,9 @@ class MMDNN(object):
                                      param_attr=fluid.ParamAttr(name="rv_fc.w"),
                                      bias_attr=False)
         reverse, _ = fluid.layers.dynamic_lstm(
-            input=rv_in_proj, size=4 * hidden_size, is_reverse=True,
+            input=rv_in_proj,
+            size=4 * hidden_size,
+            is_reverse=True,
             param_attr=fluid.ParamAttr(name="reverse_lstm.w"),
             bias_attr=fluid.ParamAttr(name="reverse_lstm.b"))
         return [forward, reverse]
@@ -96,7 +100,7 @@ class MMDNN(object):
 
         if mask is not None:
             cross_mask = fluid.layers.stack(x=[mask] * self.kernel_size, axis=1)
-            conv = cross_mask * conv + (1 - cross_mask) * (-2 ** 32 + 1)
+            conv = cross_mask * conv + (1 - cross_mask) * (-2**32 + 1)
         # valid padding
         pool = fluid.layers.pool2d(
             input=conv,
@@ -157,6 +161,8 @@ class MMDNN(object):
                                     act="tanh",
                                     size=self.hidden_size)
 
-        pred = fluid.layers.fc(input=relu_hid1, size=self.out_size)
+        pred = fluid.layers.fc(input=relu_hid1,
+                               size=self.out_size,
+                               act="softmax")
 
         return left_seq_encoder, pred

PaddleNLP/similarity_net/README.md

@@ -3,13 +3,13 @@
 ### 任务说明
 短文本语义匹配(SimilarityNet, SimNet)是一个计算短文本相似度的框架，可以根据用户输入的两个文本，计算出相似度得分。SimNet框架在百度各产品上广泛应用，主要包括BOW、CNN、RNN、MMDNN等核心网络结构形式，提供语义相似度计算训练和预测框架，适用于信息检索、新闻推荐、智能客服等多个应用场景，帮助企业解决语义匹配问题。可通过[AI开放平台-短文本相似度](https://ai.baidu.com/tech/nlp_basic/simnet)线上体验。
 ### 效果说明
-基于百度海量搜索数据，我们训练了一个SimNet-BOW-Pairwise语义匹配模型，在一些真实的FAQ问答场景中，该模型效果比基于字面的相似度方法AUC提升5%以上，我们基于百度自建测试集（包含聊天、客服等数据集）和语义匹配数据集(LCQMC)进行评测，效果如下表所示。LCQMC数据集以Accuracy为评测指标，而pairwise模型的输出为相似度，因此我们采用0.91作为分类阈值，相比于基线模型中网络结构同等复杂的CBOW模型（准确率为0.737），我们模型的准确率为0.7517。
+基于百度海量搜索数据，我们训练了一个SimNet-BOW-Pairwise语义匹配模型，在一些真实的FAQ问答场景中，该模型效果比基于字面的相似度方法AUC提升5%以上，我们基于百度自建测试集（包含聊天、客服等数据集）和语义匹配数据集(LCQMC)进行评测，效果如下表所示。LCQMC数据集以Accuracy为评测指标，而pairwise模型的输出为相似度，因此我们采用0.958作为分类阈值，相比于基线模型中网络结构同等复杂的CBOW模型（准确率为0.737），我们模型的准确率为0.7532。
 
 
 | 模型       | 百度知道         | ECOM        |QQSIM | UNICOM | LCQMC |
 |:-----------:|:-------------:|:-------------:|:-------------:|:-------------:|:-------------:|
 |   | AUC  | AUC | AUC|正逆序比|Accuracy|
-|BOW_Pairwise|0.6766|0.7308|0.7643|1.5630|0.7517|
+|BOW_Pairwise|0.6767|0.7329|0.7650|1.5630|0.7532|
 ## 快速开始
 #### 版本依赖
 本项目依赖于 Paddlepaddle Fluid 1.3.1，请参考[安装指南](http://www.paddlepaddle.org/#quick-start)进行安装。

PaddleNLP/similarity_net/config/bow_pointwise.json

@@ -10,8 +10,11 @@
         "class_name": "SoftmaxCrossEntropyLoss"
     },
     "optimizer": {
-        "class_name": "SGDOptimizer",
-        "learning_rate" : 0.001
+        "class_name": "AdamOptimizer",
+        "learning_rate": 0.001,
+        "beta1": 0.9,
+        "beta2": 0.999,
+        "epsilon": 1e-08
     },
     "task_mode": "pointwise",
     "model_path": "bow_pointwise"

PaddleNLP/similarity_net/config/cnn_pointwise.json

@@ -12,8 +12,11 @@
         "class_name": "SoftmaxCrossEntropyLoss"
     },
     "optimizer": {
-        "class_name": "SGDOptimizer",
-        "learning_rate" : 0.001
+        "class_name": "AdamOptimizer",
+        "learning_rate": 0.001,
+        "beta1": 0.9,
+        "beta2": 0.999,
+        "epsilon": 1e-08
     },
     "task_mode": "pointwise",
     "model_path": "cnn_pointwise"

PaddleNLP/similarity_net/config/gru_pointwise.json

@@ -11,8 +11,11 @@
         "class_name": "SoftmaxCrossEntropyLoss"
     },
     "optimizer": {
-        "class_name": "SGDOptimizer",
-        "learning_rate" : 0.001
+        "class_name": "AdamOptimizer",
+        "learning_rate" : 0.001,
+        "beta1": 0.9,
+        "beta2": 0.999,
+        "epsilon": 1e-08
     },
     "task_mode": "pointwise",
     "model_path": "gru_pointwise"

PaddleNLP/similarity_net/config/lstm_pointwise.json

@@ -11,8 +11,11 @@
         "class_name": "SoftmaxCrossEntropyLoss"
     },
     "optimizer": {
-        "class_name": "SGDOptimizer",
-        "learning_rate" : 0.001
+        "class_name": "AdamOptimizer",
+        "learning_rate": 0.001,
+        "beta1": 0.9,
+        "beta2": 0.999,
+        "epsilon": 1e-08
     },
     "task_mode": "pointwise",
     "model_path": "lstm_pointwise"

PaddleNLP/similarity_net/run.sh

@@ -38,7 +38,7 @@ train() {
 		--save_steps 1000 \
 		--validation_steps 100 \
 		--compute_accuracy False \
-		--lamda 0.91 \
+		--lamda 0.958 \
 		--task_mode ${TASK_MODE}
 }
 #run_evaluate
@@ -55,7 +55,7 @@ evaluate() {
 		--vocab_path ${VOCAB_PATH} \
 		--task_mode ${TASK_MODE} \
 		--compute_accuracy False \
-		--lamda 0.91 \
+		--lamda 0.958 \
 		--init_checkpoint ${INIT_CHECKPOINT}
 }
 # run_infer

PaddleNLP/similarity_net/utils.py

@@ -11,7 +11,6 @@ import six
 import numpy as np
 import logging
 import logging.handlers
-
 """
 ******functions for file processing******
 """
@@ -165,7 +164,11 @@ def print_arguments(args):
     print('------------------------------------------------')
 
 
-def init_log(log_path, level=logging.INFO, when="D", backup=7,
+def init_log(
+        log_path,
+        level=logging.INFO,
+        when="D",
+        backup=7,
         format="%(levelname)s: %(asctime)s - %(filename)s:%(lineno)d * %(thread)d %(message)s",
         datefmt=None):
     """
@@ -209,16 +212,14 @@ def init_log(log_path, level=logging.INFO, when="D", backup=7,
     if not os.path.isdir(dir):
         os.makedirs(dir)
 
-    handler = logging.handlers.TimedRotatingFileHandler(log_path + ".log",
-                                                        when=when,
-                                                        backupCount=backup)
+    handler = logging.handlers.TimedRotatingFileHandler(
+        log_path + ".log", when=when, backupCount=backup)
     handler.setLevel(level)
     handler.setFormatter(formatter)
     logger.addHandler(handler)
 
-    handler = logging.handlers.TimedRotatingFileHandler(log_path + ".log.wf",
-                                                        when=when,
-                                                        backupCount=backup)
+    handler = logging.handlers.TimedRotatingFileHandler(
+        log_path + ".log.wf", when=when, backupCount=backup)
     handler.setLevel(logging.WARNING)
     handler.setFormatter(formatter)
     logger.addHandler(handler)
@@ -241,7 +242,7 @@ def get_level():
     return logger.level
 
 
-def get_accuracy(preds, labels, mode, lamda=0.91):
+def get_accuracy(preds, labels, mode, lamda=0.958):
     """
     compute accuracy
     """