add model

modeling_lddbert.py

@@ -383,10 +383,18 @@ class LddBertModel(LddBertPreTrainedModel):
         self.embeddings = Embeddings(config)  # Embeddings
         self.transformer = Transformer(config)  # Encoder
         self.gru = nn.GRU(config.dim , config.dim//2, config.n_gru_layers, batch_first=True, bidirectional=True)
-        self.cnn = nn.Sequential(*(
-            nn.Conv1d(config.max_position_embeddings, config.max_position_embeddings, config.cnn_kernel_size, padding=(config.cnn_kernel_size-1)//2)
+
+        self.activation_cnn = get_activation('relu')
+        self.cnn = nn.ModuleList([
+            nn.Sequential(
+                nn.Conv2d(in_channels=1,
+                          out_channels=1,
+                          kernel_size=config.cnn_kernel_size,
+                          padding=(config.cnn_kernel_size-1)//2),
+                self.activation_cnn
+            )
             for _ in range(config.n_cnn_layers)
-        ))
+        ])
 
         # Initialize weights and apply final processing
         self.post_init()
@@ -511,9 +519,12 @@ class LddBertModel(LddBertPreTrainedModel):
 
         gru_output, _ = self.gru(bert_output[0])
 
-        cnn_output = self.cnn(bert_output[0])
+        cnn_output =  bert_output[0].view(input_shape[0], 1, input_shape[1], -1)
+        for i, layer_module in enumerate(self.cnn):
+            cnn_output = layer_module(cnn_output)
+        cnn_output = cnn_output.view(input_shape[0], input_shape[1], -1)
 
-        output = gru_output + cnn_output
+        output = gru_output + cnn_output        
         if not return_dict:
             return (output, ) + bert_output[1:]
 

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:151f439844ff10c523e93c90fbce4a543ab1bcce6f660822748eae4bd2e9c94c
-size 363280885
+oid sha256:189cb4b46e7ca027e3dd89c6f57b1c15e77bc2a58dd1620dcd7dde62d8f42816
+size 331811701