update cnn

conf/model/cnn.yaml

@@ -1,10 +1,10 @@
 model_name: cnn
 
-#in_channels: 100 # 使用 embedding 输出的结果，不需要指定
+in_channels: ???           # 使用 embedding 输出的结果，不需要指定
 out_channels: 100
-kernel_sizes: [3, 5, 7] # 必须为奇数，为了保证cnn的输出不改变句子长度
-activation: 'gelu'   # [relu, lrelu, prelu, selu, celu, gelu, sigmoid, tanh]
-pooling_strategy: 'max'  # [max, avg, cls]
+kernel_sizes: [3, 5, 7]    # 必须为奇数，为了保证cnn的输出不改变句子长度
+activation: 'gelu'         # [relu, lrelu, prelu, selu, celu, gelu, sigmoid, tanh]
+pooling_strategy: 'max'    # [max, avg, cls]
 dropout: 0.3
 
 # pcnn

models/PCNN.py

@@ -11,6 +11,10 @@ class PCNN(BasicModule):
         super(PCNN, self).__init__()
 
         self.use_pcnn = cfg.use_pcnn
+        if cfg.dim_strategy == 'cat':
+            cfg.in_channels = cfg.word_dim + 2 * cfg.pos_dim
+        else:
+            cfg.in_channels = cfg.word_dim
 
         self.embedding = Embedding(cfg)
         self.cnn = CNN(cfg)

module/CNN.py

@@ -30,12 +30,7 @@ class CNN(nn.Module):
         super(CNN, self).__init__()
 
         # self.xxx = config.xxx
-        # self.in_channels = config.in_channels
-        if config.dim_strategy == 'cat':
-            self.in_channels = config.word_dim + 2 * config.pos_dim
-        else:
-            self.in_channels = config.word_dim
-
+        self.in_channels = config.in_channels
         self.out_channels = config.out_channels
         self.kernel_sizes = config.kernel_sizes
         self.activation = config.activation

module/__init__.py

@@ -3,4 +3,5 @@ from .CNN import CNN
 from .RNN import RNN
 from .Attention import DotAttention, MultiHeadAttention
 from .Transformer import Transformer
-from .Capsule import Capsule
+from .Capsule import Capsule
+from .GCN import GCN