Update pipeline.py

pipeline.py

@@ -431,11 +431,7 @@ class FairseqStyleMultiheadAttention(nn.Module):
 
 
 class TransformerSentenceEncoderLayer(nn.Module):
-    # Returns (x, layer_result) where layer_result is the FFN output
-    # BEFORE the residual add -- this specific pre-residual signal is what
-    # the fine-tuned layer_weights in LinearAdapter were trained to combine
-    # across all 12 layers. Using the post-residual hidden state instead
-    # would silently feed the adapter the wrong signal.
+    
     def __init__(self, embedding_dim, ffn_embedding_dim, num_attention_heads,
                  dropout=0.1, attention_dropout=0.1, activation_dropout=0.1,
                  layer_norm_first=False):
@@ -589,15 +585,12 @@ class LinearAdapter(nn.Module):
 
 
 class SignLanguageByT5Encoder(nn.Module):
-    def __init__(self, t5_config):
+    def __init__(self, config):
         super().__init__()
         self.adapter = LinearAdapter()
-        self.block = nn.ModuleList([
-            T5Block(t5_config, has_relative_attention_bias=(i == 0))
-            for i in range(t5_config.num_layers)
-        ])
-        self.final_layer_norm = T5LayerNorm(t5_config.d_model, eps=t5_config.layer_norm_epsilon)
-        self.dropout = nn.Dropout(t5_config.dropout_rate)
+        self.block = nn.ModuleList([T5Block(config, has_relative_attention_bias=(i == 0)) for i in range(config.num_layers)])
+        self.final_layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
 
     def forward(self, face, left_hand, right_hand, body_posture):
         inputs_embeds = self.adapter(face, left_hand, right_hand, body_posture)
@@ -611,7 +604,7 @@ class SignLanguageByT5Encoder(nn.Module):
         for block in self.block:
             layer_outputs = block(hidden_states, attention_mask=extended_attention_mask, position_bias=position_bias)
             hidden_states = layer_outputs[0]
-            position_bias = layer_outputs[2]
+            position_bias = layer_outputs[-1]
 
         hidden_states = self.final_layer_norm(hidden_states)
         hidden_states = self.dropout(hidden_states)
@@ -621,15 +614,11 @@ class SignLanguageByT5Encoder(nn.Module):
 class SignLanguageByT5(nn.Module):
     def __init__(self, encoder_config, decoder_config):
         super().__init__()
-        # self.decoder_config stored explicitly -- generate_greedy needs it,
-        # and referencing a module-level variable instead (as in an earlier
-        # draft) breaks the moment this class is imported into a different
-        # file than wherever that variable happened to be defined.
+       
         self.decoder_config = decoder_config
         self.decoder_emb = nn.Embedding(decoder_config.vocab_size, decoder_config.d_model)
         self.encoder = SignLanguageByT5Encoder(encoder_config)
-        # T5Stack in current `transformers` versions doesn't accept
-        # embed_tokens as a constructor arg -- construct bare, then assign.
+       
         self.decoder = T5Stack(decoder_config)
         self.decoder.embed_tokens = self.decoder_emb
         self.lm_head = nn.Linear(decoder_config.d_model, decoder_config.vocab_size, bias=False)