为什么CuDNNLSTM和LSTM在Keras中有不同的预测?

时间:2019-02-12 19:43:06

标签: python tensorflow keras lstm

这是我的RNN:

def make_cpu_regressor():

    regressor = Sequential()

    regressor.add(LSTM(units=50, return_sequences=True, input_shape=(X_train.shape[1], 1)))
    regressor.add(Dropout(0.2))

    regressor.add(LSTM(units=50, return_sequences=True))
    regressor.add(Dropout(0.2))

    regressor.add(LSTM(units=50, return_sequences=True))
    regressor.add(Dropout(0.2))

    regressor.add(LSTM(units=50))
    regressor.add(Dropout(0.2))

    regressor.add(Dense(units=1))

    regressor.compile(optimizer='adam', loss='mean_squared_error')

    regressor.fit(X_train, y_train, epochs=100, batch_size=32)
    regressor.save('model-cpu.h5')
    return regressor

我创建了第二个,只有一个区别,我使用CuDNNLSTM而不是LSTM,其他所有内容都是相同的。使用CuDNNLSTM的NN训练得更快,但是预测之间存在显着差异:

CuDNNLSTM vs LSTM differences in predictions

为什么预测有这样的差异?

当我将CuDNNLSTM修改为150和200个历元(蓝线)时,效果会更好:

CuDNNLSTM vs CuDNNLSTM vs LSTM differences in predictions

编辑: 这是CuDNNLSTM版本的代码:

def make_gpu_regressor():

    regressor = Sequential()
    regressor.add(CuDNNLSTM(units=50, return_sequences=True, input_shape=(X_train.shape[1], 1)))
    regressor.add(Dropout(0.2))

    regressor.add(CuDNNLSTM(units=50, return_sequences=True))
    regressor.add(Dropout(0.2))

    regressor.add(CuDNNLSTM(units=50, return_sequences=True))
    regressor.add(Dropout(0.2))

    regressor.add(CuDNNLSTM(units=50))
    regressor.add(Dropout(0.2))

    regressor.add(Dense(units=1))

    regressor.compile(optimizer='adam', loss='mean_squared_error')
    regressor.fit(x=X_train, y=y_train, epochs=100, batch_size=32)
    regressor.save('model_gpu.h5')
    return regressor

regressor_gpu = make_gpu_regressor()
regressor_cpu = make_cpu_regressor()

predicted_stock_price_gpu = regressor_gpu.predict(X_test)
predicted_stock_price_gpu = sc.inverse_transform(predicted_stock_price_gpu)
predicted_stock_price_cpu = regressor_cpu.predict(X_test)
predicted_stock_price_cpu = sc.inverse_transform(predicted_stock_price_cpu)

0 个答案:

没有答案