seq2seq-推理模型产生的结果与在相同验证集上的训练模型产生的结果大不相同

Question

我正在研究时间序列seq2seq问题。对于我的方法，我使用带有教师强制的LSTM seq2seq RNN。如您所知，出于任务的目的，应该对模型进行训练，然后使用经过训练的层来构建推理模型以解决任务（即共享层）。这是我用于定义共享层的代码：

# Define the shared layers for the train and inference models
encoder_lstm = LSTM(latent_dim, return_state=True, name='encoder_lstm')
# Define the shared layers for the train and inference models
encoder_lstm = LSTM(latent_dim, return_state=True, name='encoder_lstm')
decoder_lstm = LSTM(latent_dim, return_sequences=True, 
                    return_state=True, name='decoder_lstm')
decoder_dense = Dense(decoder_output_dim, 
                      activation='linear', name='decoder_dense')
decoder_reshape = Reshape((decoder_output_dim, ), name='decoder_reshape')

接下来，我使用共享图层定义火车模型。

# Define an input for the encoder
encoder_inputs = Input(shape=(Tx, encoder_input_dim), name='encoder_input')

# We discard output and keep the states only.
_, h, c = encoder_lstm(encoder_inputs)

# Define an input for the decoder
decoder_inputs = Input(shape=(Ty, decoder_input_dim), name='decoder_input')

# Obtain all the outputs from the decoder (return_sequences = True)
decoder_outputs, _, _  = decoder_lstm(decoder_inputs, initial_state=[h, c])

# Apply dense layer to each output
decoder_outputs = decoder_dense(decoder_outputs)

train_model = Model(inputs=[encoder_inputs, decoder_inputs], outputs=decoder_outputs)

值得一提的是，我正在使用自定义损失函数，该函数基本上是均方误差，但是我掩盖了某些条目。

def masked_mse(y_true, y_pred):
    return K.mean(
                  K.mean(((y_true[:,:,0] - y_pred[:,:,0])**2)*(1-y_true[:,:,1]),
                         axis=0),
                  axis=0)

经过几个时期的训练之后，输出如下所示：

Train on 67397 samples, validate on 3389 samples
Epoch 1/10
67397/67397 [==============================] - 36s 536us/sample - loss: 0.1981 - val_loss: 0.0713
Epoch 2/10
67397/67397 [==============================] - 34s 499us/sample - loss: 0.0755 - val_loss: 0.0535
Epoch 3/10
67397/67397 [==============================] - 31s 456us/sample - loss: 0.0633 - val_loss: 0.0494
Epoch 4/10
67397/67397 [==============================] - 29s 429us/sample - loss: 0.0595 - val_loss: 0.0478

我们注意到验证集的损失约为0.045。现在，我创建从上面的共享层派生的推理模型：

# Define an input for the encoder
encoder_inputs = Input(shape=(Tx, encoder_input_dim), name='encoder_input')

# We discard output and keep the states only.
_, h, c = encoder_lstm(encoder_inputs)

# Define an input for the decoder
decoder_input = Input(shape=(1, decoder_input_dim), name='decoder_input')
current_input = decoder_input

# Obtain the outputs for each of the Ty timesteps
decoder_outputs = []
for _ in range(Ty):
    # apply a single step of recurrence
    out, h, c = decoder_lstm(current_input, initial_state=[h, c])

    # pass the LSTM output through a dense layer
    out = decoder_dense(out)

    # The input in the next timestep (its shape is (?, 1, 1))
    current_input = out

    # reshape the decoder output as (?, 1) for convenience
    out = decoder_reshape(out)

    # append the output to the model's outputs
    decoder_outputs.append(out)

inference_model = Model(inputs=[encoder_inputs, decoder_input], outputs=decoder_outputs)

使用此推理模型，我尝试在训练期间使用的same验证集中对其进行评估，以重新创建最后的结果：

# The input for the first timestep in the decoder is -1,
# (consistently, the same was applied during training)
decoder_input = -1 * np.ones((len(X_valid), 1, 1))

# Obtain the predictions, the resulting shape is (Ty, ?, 1)
y_pred = np.array(inference_model.predict([X_valid, decoder_input]))

# Reshape the output in the shape (?, Ty, 1)
y_pred = np.swapaxes(y_pred, axis1=0, axis2=1)

loss = masked_mse(K.constant(y_valid), K.constant(y_pred))
K.eval(loss)

评估损失的结果为0.1637。继续训练，它从未跌破0.14。

这很奇怪，因为我使用相同的验证集进行评估。我怀疑错误可能在推理模型的构建方式中，但是我不确定。您有什么想法？

Answer 1

一方面，您的LSTM复发应该在应用Dense层之前发生。

decoder_outputs = []
for _ in range(Ty):
    out, h, c = decoder_lstm(current_input, initial_state=[h, c])

    # This line moved to before the decoder_dense call.
    current_input = out

    out = decoder_dense(out)
    out = decoder_reshape(out)
    decoder_outputs.append(out)