Question

我正在使用LSTM进行分类。目前有两个标签。在将标签从int更改为一个热编码器之前，我的模型运行良好。

旧代码（有效）：

if i%2 == 0 :
   targets=targets.append({'seq_id':counter,'val':1},ignore_index=True)
else :
   targets=targets.append({'seq_id':counter,'val':-1},ignore_index=True) 

...

model.add(LSTM(int(num_features*4),input_shape=(num_rows,num_cols),  return_sequences=True))
model.add(LSTM(int(num_features*4),   dropout=0.5, recurrent_dropout=0.5))
model.add(Dense(1, activation='sigmoid'))

新代码（不起作用-停留在val_acc = 0.5）

if i%2 == 0 :
   targets=targets.append({'seq_id':counter,'val':'bic'},ignore_index=True)
else :
   targets=targets.append({'seq_id':counter,'val':'reno'},ignore_index=True)
...

targets = targets.values[:,1]
encoder = LabelEncoder()
encoder.fit(targets)
encoded_Y = encoder.transform(targets)
targets = np_utils.to_categorical(encoded_Y)

...

model.add(LSTM(int(num_features*4),input_shape=(num_rows,num_cols),  return_sequences=True))
model.add(LSTM(int(num_features*4),   dropout=0.5, recurrent_dropout=0.5))
model.add(Dense(2, activation='sigmoid')) #That line Changed

Answer 1

如果您要进行一次热编码，则应将最后一个Dense层的激活更改为softmax。

Answer 2

我将损失函数更改为categorical_crossentropy，它起作用了。

标签编码时LSTM模型无法学习

2 个答案: