pad mask issue and embedding issue

models/blocks/decoder_layer.py

@@ -29,7 +29,7 @@ def __init__(self, d_model, ffn_hidden, n_head, drop_prob):
     def forward(self, dec, enc, t_mask, s_mask):
         # 1. compute self attention
         _x = dec
-        x = self.self_attention(q=dec, k=dec, v=dec, mask=trg_mask)
+        x = self.self_attention(q=dec, k=dec, v=dec, mask=t_mask)
 
         # 2. add and norm
         x = self.norm1(x + _x)
@@ -38,7 +38,7 @@ def forward(self, dec, enc, t_mask, s_mask):
         if enc is not None:
             # 3. compute encoder - decoder attention
             _x = x
-            x = self.enc_dec_attention(q=x, k=enc, v=enc, mask=src_mask)
+            x = self.enc_dec_attention(q=x, k=enc, v=enc, mask=s_mask)
 
             # 4. add and norm
             x = self.norm2(x + _x)

models/blocks/encoder_layer.py

@@ -25,7 +25,7 @@ def __init__(self, d_model, ffn_hidden, n_head, drop_prob):
     def forward(self, x, s_mask):
         # 1. compute self attention
         _x = x
-        x = self.attention(q=x, k=x, v=x, mask=src_mask)
+        x = self.attention(q=x, k=x, v=x, mask=s_mask)
 
         # 2. add and norm
         x = self.norm1(x + _x)

models/embedding/token_embeddings.py

@@ -19,4 +19,4 @@ class for token embedding that included positional information
         :param vocab_size: size of vocabulary
         :param d_model: dimensions of model
         """
-        super(TokenEmbedding, self).__init__(vocab_size, d_model, padding_idx=0)
+        super(TokenEmbedding, self).__init__(vocab_size, d_model, padding_idx=1)

models/layers/scale_dot_product_attention.py

@@ -30,6 +30,8 @@ def forward(self, q, k, v, mask=None, e=1e-12):
         k_t = k.view(batch_size, head, d_tensor, length)  # transpose
         score = (q @ k_t) / math.sqrt(d_tensor)  # scaled dot product
 
+        print("score : {}" .format(score.size()))
+        print("mask : {}" .format(mask.size()))
         # 2. apply masking (opt)
         if mask is not None:
             score = score.masked_fill(mask == 0, -e)

models/model/transformer.py

@@ -38,31 +38,37 @@ def __init__(self, src_pad_idx, trg_pad_idx, trg_sos_idx, enc_voc_size, dec_voc_
                                device=device)
 
     def forward(self, src, trg):
-        src_mask = self.make_src_mask(src)
-        trg_mask = self.make_trg_mask(trg)
+        src_mask = self.make_pad_mask(src, src)
+
+        src_trg_mask = self.make_pad_mask(trg, src)
+
+        trg_mask = self.make_pad_mask(trg, trg) * \
+                   self.make_no_peak_mask(trg, trg)
+
         enc_src = self.encoder(src, src_mask)
-        output = self.decoder(trg, enc_src, trg_mask, src_mask)
+        output = self.decoder(trg, enc_src, trg_mask, src_trg_mask)
         return output
 
-    def make_src_mask(self, src):
-        batch_size, length = src.size()
+    def make_pad_mask(self, q, k):
+        len_q, len_k = q.size(1), k.size(1)
 
         # batch_size x 1 x 1 x len_k
-        src_k = src.ne(self.src_pad_idx).unsqueeze(1).unsqueeze(2)
+        k = k.ne(self.src_pad_idx).unsqueeze(1).unsqueeze(2)
         # batch_size x 1 x len_q x len_k
-        src_k = src_k.repeat(1, 1, length, 1)
+        k = k.repeat(1, 1, len_q, 1)
 
         # batch_size x 1 x len_q x 1
-        src_q = src.ne(self.src_pad_idx).unsqueeze(1).unsqueeze(3)
+        q = q.ne(self.src_pad_idx).unsqueeze(1).unsqueeze(3)
         # batch_size x 1 x len_q x len_k
-        src_q = src_q.repeat(1, 1, 1, length)
+        q = q.repeat(1, 1, 1, len_k)
+
+        mask = k & q
+        return mask
+
+    def make_no_peak_mask(self, q, k):
+        len_q, len_k = q.size(1), k.size(1)
 
-        src_mask = src_k & src_q
-        return src_mask
+        # len_q x len_k
+        mask = torch.tril(torch.ones(len_q, len_k)).type(torch.BoolTensor).to(self.device)
 
-    def make_trg_mask(self, trg):
-        trg_pad_mask = (trg != self.trg_pad_idx).unsqueeze(1).unsqueeze(3)
-        trg_len = trg.shape[1]
-        trg_sub_mask = torch.tril(torch.ones(trg_len, trg_len)).type(torch.ByteTensor).to(self.device)
-        trg_mask = trg_pad_mask & trg_sub_mask
-        return trg_mask
+        return mask

train.py

@@ -47,7 +47,6 @@ def initialize_weights(m):
 scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer=optimizer,
                                                  verbose=True,
                                                  factor=factor,
-                                                 min_lr=min_lr,
                                                  patience=patience)
 
 criterion = nn.CrossEntropyLoss(ignore_index=src_pad_idx)

util/data_loader.py

@@ -3,8 +3,8 @@
 @when : 2019-10-29
 @homepage : https://github.com/gusdnd852
 """
-from torchtext.data import Field, BucketIterator
-from torchtext.datasets.translation import Multi30k
+from torchtext.legacy.data import Field, BucketIterator
+from torchtext.legacy.datasets.translation import Multi30k
 
 
 class DataLoader:

util/tokenizer.py

@@ -9,8 +9,8 @@
 class Tokenizer:
 
     def __init__(self):
-        self.spacy_de = spacy.load('de')
-        self.spacy_en = spacy.load('en')
+        self.spacy_de = spacy.load('de_core_news_sm')
+        self.spacy_en = spacy.load('en_core_web_sm')
 
     def tokenize_de(self, text):
         """