Merge pull request NTMC-Community#180 from faneshion/feature/dynamic_…

…pool_div_0

Feature/dynamic pool div 0

examples/QuoraQP/config/matchpyramid_quoraqp.config

@@ -18,7 +18,7 @@
         "embed_size": 300,
         "train_embed": true,
         "embed_path": "./data/QuoraQP/embed_glove_d300_norm",
-        "vocab_size": 28159,
+        "vocab_size": 28211,
         "target_mode": "classification",
         "class_num": 2,
         "text1_maxlen": 20,

matchzoo/layers/DynamicMaxPooling.py

@@ -46,14 +46,26 @@ def compute_output_shape(self, input_shape):
     def dynamic_pooling_index(len1, len2, max_len1, max_len2, 
                               compress_ratio1 = 1, compress_ratio2 = 1):
         def dpool_index_(batch_idx, len1_one, len2_one, max_len1, max_len2):
+            '''
+            TODO: Here is the check of sentences length to be positive.
+            To make sure that the lenght of the input sentences are positive. 
             if len1_one == 0:
-                print("len1 = 0 at batch_idx = {}".format(batch_idx))
+                print("[Error:DynamicPooling] len1 = 0 at batch_idx = {}".format(batch_idx))
                 exit()
             if len2_one == 0:
-                print("len2 = 0 at batch_idx = {}".format(batch_idx))
+                print("[Error:DynamicPooling] len2 = 0 at batch_idx = {}".format(batch_idx))
                 exit()
-            stride1 = 1.0 * max_len1 / len1_one
-            stride2 = 1.0 * max_len2 / len2_one
+            '''
+            if len1_one == 0:
+                stride1 = max_len1
+            else:
+                stride1 = 1.0 * max_len1 / len1_one
+
+            if len2_one == 0:
+                stride2 = max_len2
+            else:
+                stride2 = 1.0 * max_len2 / len2_one
+
             idx1_one = [int(i / stride1) for i in range(max_len1)]
             idx2_one = [int(i / stride2) for i in range(max_len2)]
             mesh1, mesh2 = np.meshgrid(idx1_one, idx2_one)

matchzoo/models/arci.py

@@ -5,7 +5,7 @@
 import keras.backend as K
 from keras.models import Sequential, Model
 from keras.layers import *
-from keras.layers import Reshape, Embedding,Merge, Dot
+from keras.layers import Reshape, Embedding, Dot
 from keras.optimizers import Adam
 from model import BasicModel
 

matchzoo/models/arcii.py

@@ -5,7 +5,7 @@
 import keras.backend as K
 from keras.models import Sequential, Model
 from keras.layers import *
-from keras.layers import Reshape, Embedding,Merge, Dot
+from keras.layers import Reshape, Embedding, Dot
 from keras.optimizers import Adam
 from model import BasicModel
 

matchzoo/models/bimpm.py

@@ -5,7 +5,7 @@
 import keras.backend as K
 from keras.models import Sequential, Model
 from keras.layers import *
-from keras.layers import Reshape, Embedding, Merge, Dot
+from keras.layers import Reshape, Embedding, Dot
 from keras.optimizers import Adam
 from model import BasicModel
 from layers.DynamicMaxPooling import *

matchzoo/models/duet.py

@@ -6,7 +6,6 @@
 from keras import regularizers
 from keras.models import Sequential, Model
 from keras.layers import *
-#from keras.layers import Reshape, Embedding,Merge, Dot, Lambda
 from keras.optimizers import Adam
 from model import BasicModel
 import tensorflow as tf

matchzoo/models/matchpyramid.py

@@ -5,7 +5,7 @@
 import keras.backend as K
 from keras.models import Sequential, Model
 from keras.layers import *
-from keras.layers import Reshape, Embedding, Merge, Dot
+from keras.layers import Reshape, Embedding, Dot
 from keras.optimizers import Adam
 from model import BasicModel
 from layers.DynamicMaxPooling import *

matchzoo/models/mvlstm.py

@@ -5,7 +5,7 @@
 import keras.backend as K
 from keras.models import Sequential, Model
 from keras.layers import *
-from keras.layers import Reshape, Embedding,Merge, Dot
+from keras.layers import Reshape, Embedding, Dot
 from keras.optimizers import Adam
 from model import BasicModel
 from utils.utility import *

matchzoo/utils/rank_io.py

@@ -98,12 +98,12 @@ def convert_embed_2_numpy(embed_dict, max_size=0, embed=None):
     if embed is None:
         embed = np.zeros((max_size, feat_size), dtype=np.float32)
 
-    if len(embed_dict) != len(embed):
-        raise Exception("vocab_size %d is not equal to embed_size %d, change the vocab_size in the config!"
+    if len(embed_dict) > len(embed):
+        raise Exception("vocab_size %d is larger than embed_size %d, change the vocab_size in the config!"
                         % (len(embed_dict), len(embed)))
 
     for k in embed_dict:
         embed[k] = np.array(embed_dict[k])
-    print('Generate numpy embed: %s', str(embed.shape), end='\n')
+    print('Generate numpy embed:', str(embed.shape), end='\n')
     return embed