训练集对lstm的影响大小

Question

我正在使用以下模型求解分类时间序列。当我使用小的训练数据集（近5000行）时，模型的精度约为0.9，损失约为0.02。但是，当我使用超过900000行的训练数据集时，准确性和损失会分别变为0.3和1.5。对于这种现象有什么解释吗，我应该增加LSTM单位或batch_size吗？那么到哪个值范围？

    ```
   # using keras (tensorflow backend)
   timesteps = 6
   LSTM_units = 45
   dropout = 0.2
   epochs = 100
   batch_size = 100

    model_input = Input(shape=(timesteps, features))
    batchNormalization = BatchNormalization()(model_input)
    lstm1 = LSTM(units = LSTM_units, return_sequences = True)(batchNormalization)
    dropout1 = Dropout(dropout)(lstm1)
    lstm2 = LSTM(units = LSTM_units)(dropout1)
    dropout2 = Dropout(dropout)(lstm2)

    dense1 = Dense(150, activation='relu')(dropout2)
    outputs = Dense(7, activation='softmax')(dense1)
    model = Model(inputs=[model_input], outputs=outputs)

    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

    model.fit_generator(generator, epochs=epochs, steps_per_epoch=batch_size, workers=2, verbose=2)
    ```