fix bug in #128

shenweichen · Oct 23, 2020 · 61b5775 · 61b5775
1 parent 7996548
commit 61b5775
Show file tree

Hide file tree

Showing 3 changed files with 17 additions and 14 deletions.
diff --git a/deepctr_torch/inputs.py b/deepctr_torch/inputs.py
@@ -266,6 +266,6 @@ def get_dense_input(X, features, feature_columns):
 
 def maxlen_lookup(X, sparse_input_dict, maxlen_column):
     if maxlen_column is None or len(maxlen_column)==0:
-        raise ValueError('please add max length column for VarLenSparseFeat of DIEN input')
+        raise ValueError('please add max length column for VarLenSparseFeat of DIN/DIEN input')
     lookup_idx = np.array(sparse_input_dict[maxlen_column[0]])
     return X[:, lookup_idx[0]:lookup_idx[1]].long()
diff --git a/deepctr_torch/models/din.py b/deepctr_torch/models/din.py
@@ -79,16 +79,15 @@ def __init__(self, dnn_feature_columns, history_feature_list, dnn_use_bn=False,
 
 
     def forward(self, X):
-        sparse_embedding_list, dense_value_list = self.input_from_feature_columns(X, self.dnn_feature_columns,
-                                                                                  self.embedding_dict)
+        _, dense_value_list = self.input_from_feature_columns(X, self.dnn_feature_columns, self.embedding_dict)
 
         # sequence pooling part
         query_emb_list = embedding_lookup(X, self.embedding_dict, self.feature_index, self.sparse_feature_columns,
-                                          self.history_feature_list, self.history_feature_list, to_list=True)
+                                          return_feat_list=self.history_feature_list, to_list=True)
         keys_emb_list = embedding_lookup(X, self.embedding_dict, self.feature_index, self.history_feature_columns,
-                                         self.history_fc_names, self.history_fc_names, to_list=True)
-        dnn_input_emb_list = embedding_lookup(X, self.embedding_dict, self.feature_index, self.sparse_feature_columns, mask_feat_list=self.history_feature_list, to_list=True)
-
+                                         return_feat_list=self.history_fc_names, to_list=True)
+        dnn_input_emb_list = embedding_lookup(X, self.embedding_dict, self.feature_index, self.sparse_feature_columns,
+                                              to_list=True)
 
         sequence_embed_dict = varlen_embedding_lookup(X, self.embedding_dict, self.feature_index,
                                                       self.sparse_varlen_feature_columns)
@@ -97,12 +96,15 @@ def forward(self, X):
                                                       self.sparse_varlen_feature_columns, self.device)
 
         dnn_input_emb_list += sequence_embed_list
+        deep_input_emb = torch.cat(dnn_input_emb_list, dim=-1)
 
         # concatenate
         query_emb = torch.cat(query_emb_list, dim=-1)                     # [B, 1, E]
         keys_emb = torch.cat(keys_emb_list, dim=-1)                       # [B, T, E]
-        keys_length = torch.ones((query_emb.size(0), 1)).to(self.device)  # [B, 1]
-        deep_input_emb = torch.cat(dnn_input_emb_list, dim=-1)
+
+        keys_length_feature_name = [feat.length_name for feat in self.varlen_sparse_feature_columns if
+                                    feat.length_name is not None]
+        keys_length = torch.squeeze(maxlen_lookup(X, self.feature_index, keys_length_feature_name), 1)  # [B, 1]
 
         hist = self.attention(query_emb, keys_emb, keys_length)           # [B, 1, E]
 

diff --git a/examples/run_din.py b/examples/run_din.py
@@ -14,9 +14,8 @@ def get_xy_fd():
                        SparseFeat('item', 3 + 1, embedding_dim=8), SparseFeat('item_gender', 2 + 1, embedding_dim=8),
                        DenseFeat('score', 1)]
 
-    feature_columns += [VarLenSparseFeat(SparseFeat('hist_item', 3 + 1, embedding_dim=8), 4),
-                        VarLenSparseFeat(SparseFeat('hist_item_gender', 2 + 1, embedding_dim=8), 4)]
-
+    feature_columns += [VarLenSparseFeat(SparseFeat('hist_item', 3 + 1, embedding_dim=8), 4, length_name="seq_length"),
+                        VarLenSparseFeat(SparseFeat('hist_item_gender', 2 + 1, embedding_dim=8), 4, length_name="seq_length")]
     behavior_feature_list = ["item", "item_gender"]
     uid = np.array([0, 1, 2])
     ugender = np.array([0, 1, 0])
@@ -26,9 +25,11 @@ def get_xy_fd():
 
     hist_iid = np.array([[1, 2, 3, 0], [1, 2, 3, 0], [1, 2, 0, 0]])
     hist_igender = np.array([[1, 1, 2, 0], [2, 1, 1, 0], [2, 1, 0, 0]])
+    behavior_length = np.array([3, 3, 2])
 
     feature_dict = {'user': uid, 'gender': ugender, 'item': iid, 'item_gender': igender,
-                    'hist_item': hist_iid, 'hist_item_gender': hist_igender, 'score': score}
+                    'hist_item': hist_iid, 'hist_item_gender': hist_igender, 'score': score,
+                    "seq_length": behavior_length}
     x = {name: feature_dict[name] for name in get_feature_names(feature_columns)}
     y = np.array([1, 0, 1])
 
@@ -43,7 +44,7 @@ def get_xy_fd():
         print('cuda ready...')
         device = 'cuda:0'
 
-    model = DIN(feature_columns, behavior_feature_list, device=device)
+    model = DIN(feature_columns, behavior_feature_list, device=device, att_weight_normalization=True)
     model.compile('adagrad', 'binary_crossentropy',
                   metrics=['binary_crossentropy'])
     history = model.fit(x, y, batch_size=3, epochs=10, verbose=2, validation_split=0.0)