v1.2: optimize model arch

cnocr/__version__.py

-__version__ = '1.1.1'
+__version__ = '1.2.0'

cnocr/fit/fit.py

@@ -23,8 +23,8 @@ def fit(network, data_train, data_val, metrics, args, hp, data_names=None):
     logging.info('hp: %s', hp)
 
     sym, arg_params, aux_params = _load_model(args)
-    if sym is not None:
-        assert sym.tojson() == network.tojson()
+    # if sym is not None:
+    #     assert sym.tojson() == network.tojson()
     if not os.path.exists(os.path.dirname(args.prefix)):
         os.makedirs(os.path.dirname(args.prefix))
 

cnocr/hyperparams/cn_hyperparams.py

@@ -25,15 +25,15 @@ class CnHyperparams(object):
 
         # LSTM hyper parameters
         self.seq_model_type = 'lstm'
-        self._num_hidden = 100
-        self._num_lstm_layer = 2
+        self._num_hidden = 128
+        self._num_lstm_layer = 1
 
         # 模型对于图片宽度压缩的比例（模型中的卷积层造成的）；由模型决定，不同模型不一样
         self.seq_len_cmpr_ratio = None
         # 序列长度；由模型决定，不同模型不一样
         self._seq_length = None
         self._num_label = 20
-        self._drop_out = 0.5
+        self._drop_out = 0.3
 
     def __repr__(self):
         return str(self.__dict__)

cnocr/symbols/crnn.py

@@ -77,7 +77,7 @@ def gen_seq_model(hp):
 
         def fc_seq_model(data):
             fc = mx.sym.FullyConnected(
-                data, num_hidden=2 * hp.num_hidden, flatten=False, name='seq-fc'
+                data, num_hidden=hp.num_hidden, flatten=False, name='seq-fc'
             )
             net = mx.sym.Activation(data=fc, act_type='relu', name='seq-relu')
             return net