Update matching task (#814)

demo/pairwise_text_matching/ernie_pairwise_matching.py

@@ -52,9 +52,9 @@ if __name__ == '__main__':
 
     # Construct transfer learning network
     # Use sequence-level output.
-    query = outputs["pooled_output"]
-    left = outputs['pooled_output_2']
-    right = outputs['pooled_output_3']
+    query = outputs["sequence_output"]
+    left = outputs['sequence_output_2']
+    right = outputs['sequence_output_3']
 
     # Select fine-tune strategy
     strategy = hub.AdamWeightDecayStrategy(

demo/pairwise_text_matching/ernie_pairwise_matching_predict.py

@@ -48,9 +48,9 @@ if __name__ == '__main__':
 
     # Construct transfer learning network.
     # Use sequence-level output.
-    query = outputs["pooled_output"]
-    left = outputs['pooled_output_2']
-    right = outputs['pooled_output_3']
+    query = outputs["sequence_output"]
+    left = outputs['sequence_output_2']
+    right = outputs['sequence_output_3']
 
     # Select fine-tune strategy.
     strategy = hub.AdamWeightDecayStrategy()
@@ -91,6 +91,6 @@ if __name__ == '__main__':
         max_seq_len=args.max_seq_len,
         label_list=dataset.get_labels(),
         return_result=True,
-        accelerate_mode=False)
+        accelerate_mode=True)
     for index, text in enumerate(text_pairs):
         print("data: %s, prediction_label: %s" % (text, results[index]))

demo/pairwise_text_matching/run_embedding_pairwise_matching.sh

@@ -3,7 +3,7 @@ CKPT_DIR="./ckpt_embedding_pairwise_matching"
 python -u embedding_pairwise_matching.py \
             --batch_size=128 \
             --checkpoint_dir=${CKPT_DIR} \
-            --learning_rate=5e-2 \
+            --learning_rate=5e-3 \
             --max_seq_len=128 \
-            --num_epoch=300 \
+            --num_epoch=100 \
             --network=bow

demo/pairwise_text_matching/run_embedding_pairwise_matching_predict.sh

@@ -4,4 +4,4 @@ python -u embedding_pairwise_matching_predict.py \
             --batch_size=1 \
             --checkpoint_dir=${CKPT_DIR} \
             --max_seq_len=128 \
-            --network=lstm
+            --network=bow

demo/pointwise_text_matching/embedding_pointwise_matching_predict.py

@@ -88,12 +88,12 @@ if __name__ == '__main__':
     # Prediction data sample.
     text_pairs = [
         [
-            "请问不是您的账户吗？",  # query
-            "您好，请问您使用的邮箱类型是？"  # title
+            "淘宝上怎么用信用卡分期付款",  # query
+            "淘宝上怎么分期付款，没有信用卡",  # title
         ],
         [
-            "推荐个手机游戏",  # query
-            "手机游戏推荐"  # title
+            "山楂干怎么吃好吃？",  # query
+            "山楂怎么做好吃",  # title
         ]
     ]
 

demo/pointwise_text_matching/ernie_pointwise_matching.py

@@ -52,8 +52,8 @@ if __name__ == '__main__':
 
     # Construct transfer learning network
     # Use token-level output.
-    query = outputs["pooled_output"]
-    left = outputs['pooled_output_2']
+    query = outputs["sequence_output"]
+    left = outputs['sequence_output_2']
 
     # Select fine-tune strategy
     strategy = hub.AdamWeightDecayStrategy(

demo/pointwise_text_matching/ernie_pointwise_matching_predict.py

@@ -48,8 +48,8 @@ if __name__ == '__main__':
 
     # Construct transfer learning network
     # Use token-level output.
-    query = outputs["pooled_output"]
-    left = outputs['pooled_output_2']
+    query = outputs["sequence_output"]
+    left = outputs['sequence_output_2']
 
     # Select fine-tune strategy
     strategy = hub.AdamWeightDecayStrategy()
@@ -73,12 +73,12 @@ if __name__ == '__main__':
     # Prediction data sample.
     text_pairs = [
         [
-            "小品《战狼故事》中，吴京突破重重障碍解救爱人，深情告白太感人;爱人",  # query
-            "外文名:愛人;摘要:爱人，意思是：情人。;义项描述:日本语词汇;语言:日文;中文名:爱人;标签:文化;",  # title
+            "淘宝上怎么用信用卡分期付款",  # query
+            "淘宝上怎么分期付款，没有信用卡",  # title
         ],
         [
-            "儿子祝融被杀害，西天王大发雷霆，立即下令捉拿天庭三公主;儿子",  # query
-            "摘要:《儿子》是曹国昌1983年创作的木雕，收藏于中国美术馆。;材质：:木雕;作者：:曹国昌;中文名:儿子;创作年代：:1983年;义项描述:曹国昌木雕;标签:文化;",  # title
+            "山楂干怎么吃好吃？",  # query
+            "山楂怎么做好吃",  # title
         ]
     ]
 
@@ -88,6 +88,6 @@ if __name__ == '__main__':
         max_seq_len=args.max_seq_len,
         label_list=dataset.get_labels(),
         return_result=True,
-        accelerate_mode=False)
+        accelerate_mode=True)
     for index, text in enumerate(text_pairs):
         print("data: %s, prediction_label: %s" % (text, results[index]))

demo/pointwise_text_matching/run_embedding_pointwise_matching.sh

@@ -5,5 +5,5 @@ python -u embedding_pointwise_matching.py \
             --checkpoint_dir=${CKPT_DIR} \
             --learning_rate=5e-3 \
             --max_seq_len=128 \
-            --num_epoch=300 \
+            --num_epoch=100 \
             --network=bow

paddlehub/finetune/task/matching_task.py

@@ -106,32 +106,81 @@ class PairwiseTextMatchingTask(BaseTask):
                             % self.network)
 
             query_feats, left_feats = net_func(query_unpad, left_unpad)
+            query_feats = fluid.layers.fc(
+                input=query_feats,
+                size=300,
+                param_attr=fluid.ParamAttr(
+                    name="query_fc_w",
+                    initializer=fluid.initializer.TruncatedNormal(scale=0.02),
+                ),
+                bias_attr=fluid.ParamAttr(
+                    name="query_fc_b",
+                    initializer=fluid.initializer.Constant(0.),
+                ),
+                act="tanh")
+            left_feats = fluid.layers.fc(
+                input=left_feats,
+                size=300,
+                param_attr=fluid.ParamAttr(
+                    name="title_fc_w",
+                    initializer=fluid.initializer.TruncatedNormal(scale=0.02),
+                ),
+                bias_attr=fluid.ParamAttr(
+                    name="title_fc_b",
+                    initializer=fluid.initializer.Constant(0.),
+                ),
+                act="tanh")
+
             left_concat = fluid.layers.concat(
                 input=[query_feats, left_feats], axis=1)
 
-            query_feats, right_feats = net_func(query_unpad, right_unpad)
+            _, right_feats = net_func(query_unpad, right_unpad)
+            right_feats = fluid.layers.fc(
+                input=right_feats,
+                size=300,
+                param_attr=fluid.ParamAttr(
+                    name="title_fc_w",
+                    initializer=fluid.initializer.TruncatedNormal(scale=0.02),
+                ),
+                bias_attr=fluid.ParamAttr(
+                    name="title_fc_b",
+                    initializer=fluid.initializer.Constant(0.),
+                ),
+                act="tanh")
+
             right_concat = fluid.layers.concat(
-                input=[query_feats, right_feats],
-                axis=1,
-            )
+                input=[query_feats, right_feats], axis=1)
+
         else:
             query_feats = fluid.layers.dropout(
                 x=self.query_feature,
                 dropout_prob=0.1,
                 dropout_implementation="upscale_in_train")
+            query_mean = fluid.layers.reduce_mean(query_feats, dim=1)
+
             left_feats = fluid.layers.dropout(
                 x=self.left_feature,
                 dropout_prob=0.1,
                 dropout_implementation="upscale_in_train")
+            left_mean = fluid.layers.reduce_mean(left_feats, dim=1)
+
+            left_sub = fluid.layers.elementwise_sub(query_mean, left_mean)
+            left_sub = fluid.layers.abs(left_sub)
+
             left_concat = fluid.layers.concat(
-                input=[query_feats, left_feats], axis=-1)
+                input=[query_mean, left_mean, left_sub], axis=-1)
 
             right_feats = fluid.layers.dropout(
                 x=self.right_feature,
                 dropout_prob=0.1,
                 dropout_implementation="upscale_in_train")
+            right_mean = fluid.layers.reduce_mean(right_feats, dim=1)
+
+            right_sub = fluid.layers.elementwise_sub(query_mean, right_mean)
+            right_sub = fluid.layers.abs(right_sub)
+
             right_concat = fluid.layers.concat(
-                input=[query_feats, right_feats],
+                input=[query_mean, right_mean, right_sub],
                 axis=-1,
             )
 
@@ -435,22 +484,91 @@ class PointwiseTextMatchingTask(BaseTask):
                     self.network)
 
             query_feats, title_feats = net_func(query_unpad, title_unpad)
+
+            query_fc = fluid.layers.fc(
+                input=query_feats,
+                size=300,
+                param_attr=fluid.ParamAttr(
+                    name="query_fc_w",
+                    initializer=fluid.initializer.TruncatedNormal(scale=0.02),
+                ),
+                bias_attr=fluid.ParamAttr(
+                    name="query_fc_b",
+                    initializer=fluid.initializer.Constant(0.),
+                ),
+                act="tanh")
+            title_fc = fluid.layers.fc(
+                input=title_feats,
+                size=300,
+                param_attr=fluid.ParamAttr(
+                    name="title_fc_w",
+                    initializer=fluid.initializer.TruncatedNormal(scale=0.02),
+                ),
+                bias_attr=fluid.ParamAttr(
+                    name="title_fc_b",
+                    initializer=fluid.initializer.Constant(0.),
+                ),
+                act="tanh")
             title_concat = fluid.layers.concat(
-                input=[query_feats, title_feats], axis=1)
+                input=[query_fc, title_fc], axis=1)
+
+            fc1 = fluid.layers.fc(
+                input=title_concat,
+                size=256,
+                param_attr=fluid.ParamAttr(
+                    name="matching_fc1_w",
+                    initializer=fluid.initializer.TruncatedNormal(scale=0.02),
+                ),
+                bias_attr=fluid.ParamAttr(
+                    name="matching_fc1_b",
+                    initializer=fluid.initializer.Constant(0.),
+                ),
+                act="tanh")
+            fc2 = fluid.layers.fc(
+                input=fc1,
+                size=128,
+                param_attr=fluid.ParamAttr(
+                    name="matching_fc2_w",
+                    initializer=fluid.initializer.TruncatedNormal(scale=0.02),
+                ),
+                bias_attr=fluid.ParamAttr(
+                    name="matching_fc2_b",
+                    initializer=fluid.initializer.Constant(0.),
+                ),
+                act="tanh")
+            projection = fluid.layers.fc(
+                input=fc2,
+                size=96,
+                param_attr=fluid.ParamAttr(
+                    name="matching_fc3_w",
+                    initializer=fluid.initializer.TruncatedNormal(scale=0.02),
+                ),
+                bias_attr=fluid.ParamAttr(
+                    name="matching_fc3_b",
+                    initializer=fluid.initializer.Constant(0.),
+                ),
+                act="tanh")
+
         else:
             query_feats = fluid.layers.dropout(
                 x=self.query_feature,
                 dropout_prob=0.1,
                 dropout_implementation="upscale_in_train")
+            query_mean = fluid.layers.reduce_mean(query_feats, dim=1)
             title_feats = fluid.layers.dropout(
                 x=self.title_feature,
                 dropout_prob=0.1,
                 dropout_implementation="upscale_in_train")
-            title_concat = fluid.layers.concat(
-                input=[query_feats, title_feats], axis=-1)
+            title_mean = fluid.layers.reduce_mean(title_feats, dim=1)
+
+            sub = fluid.layers.elementwise_sub(query_mean, title_mean)
+            sub = fluid.layers.abs(sub)
+
+            projection = fluid.layers.concat(
+                input=[query_mean, title_mean, sub], axis=-1)
 
         score = fluid.layers.fc(
-            input=title_concat,
+            input=projection,
             size=2,
             param_attr=fluid.ParamAttr(
                 name="matching_out_w",

paddlehub/version.py

@@ -13,5 +13,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """ PaddleHub version string """
-hub_version = "1.8.0"
+hub_version = "1.8.1"
 module_proto_version = "1.0.0"