Add Non-Static Implementation

Author: practicingman

README.md

@@ -37,11 +37,18 @@ python3 main.py
         early stop by 1000 steps, acc: 94.0000%
 - [x] CNN-static 使用预训练的静态词向量
     ```bash
-      python main.py -static
+      python main.py -static=true
     ```
     >
         Batch[1900] - loss: 0.011894  acc: 100.0000%(128/128)
         Evaluation - loss: 0.000018  acc: 95.0000%(6679/7000)
         early stop by 1000 steps, acc: 95.0000%
-
+- [x] CNN-static 微调预训练的静态词向量
+    ```bash
+      python main.py -static=true -no-static=true
+    ```
+    >
+        Batch[1500] - loss: 0.008823  acc: 99.0000%(127/128))
+        Evaluation - loss: 0.000016  acc: 96.0000%(6729/7000)
+        early stop by 1000 steps, acc: 96.0000%
 - [ ] CNN-multichannel

dataset.py

@@ -1,6 +1,8 @@
 import re
 from torchtext import data
 import jieba
+import logging
+jieba.setLogLevel(logging.INFO)
 
 regex = re.compile(r'[^\u4e00-\u9fa5aA-Za-z0-9]')
 

main.py

@@ -29,6 +29,7 @@
                     help='comma-separated filter sizes to use for convolution')
 
 parser.add_argument('-static', type=bool, default=False, help='whether to use static pre-trained word vectors')
+parser.add_argument('-non-static', type=bool, default=False, help='whether to fine-tune static pre-trained word vectors')
 parser.add_argument('-pretrained-name', type=str, default='sgns.zhihu.word',
                     help='filename of pre-trained word vectors')
 parser.add_argument('-pretrained-path', type=str, default='pretrained', help='path of pre-trained word vectors')

model.py

@@ -17,7 +17,7 @@ def __init__(self, args):
         embedding_dimension = args.embedding_dim
         self.embedding = nn.Embedding(vocabulary_size, embedding_dimension)
         if args.static:
-            self.embedding = self.embedding.from_pretrained(args.vectors)
+            self.embedding = self.embedding.from_pretrained(args.vectors, freeze=not args.non_static)
 
         self.convs = nn.ModuleList(
             [nn.Conv2d(chanel_num, filter_num, (size, embedding_dimension)) for size in filter_sizes])