自定义Tensorflow Keras -TypeError中的model.fit()：TypeError: compile()获得意外的关键字参数'optimizer‘

Question

我正在尝试通过修改Keras中的model.fit()来实现我自己的重量训练算法。我读过来自Keras的this文章，它基于GANs很好地解释了这一点。不过，我现在面对的问题是：

我遵循了与文章中解释的完全相同的步骤和程序，但我用我自己的训练算法替换了他们的训练算法。我得到的错误是

TypeError: compile() got an unexpected keyword argument 'optimizer'

我创建的模型使用Tensorflow 1.15。另外，我使用来自Tensorflow hub的ELMO嵌入层，它需要TensorFlow1.15(因为它还不支持V2+ )

def Predictor():
  input_sentence = Input(shape=(250,), dtype=tf.int32)
  embedding = ELMoEmbedding(index_word=index_word, trainable=False)(input_sentence)
  lstm = LSTM(1024)(embedding)
  dense = Dense(64, activation='relu')(lstm)
  pred = Dense(5, activation='softmax')(dense)

  model = Model(inputs=[input_sentence], outputs=pred)

  return model

def Adversary():
  input_features = Input((1,))
  dense1 = Dense(64, activation='relu')(input_features)
  pred = Dense(5, activation='sigmoid')(dense1)

  model = Model(inputs=[input_features], outputs=[pred])

  return model

class Mitigator(Model):

    def __init__(self, predictor, adversary, debias_param):
        super(Mitigator, self).__init__()
        self.predictor = predictor
        self.adversary = adversary
        self.debias_param = debias_param

    def compile(self, a_optimizer, p_optimizer, a_loss, p_loss):
        super(Mitigator, self).compile()
        self.a_optimizer = a_optimizer
        self.p_optimizer = p_optimizer
        self.a_loss = a_loss
        self.p_loss = p_loss

    def train_step(self, data):
        # Pass tf.data.Dataset element based on numpy arrays
        x, train_y, z_true = data

        # Generate prediction and compute loss for train_step
        y_pred = self.predictor(x)
        pred_loss = self.p_loss(train_y, y_pred)

        # Input data for Adversary
        adv_x = tf.concat([y_pred, train_y], axis=1)

        # Train Adversary
        z_pred = self.adversary(adv_x)
        adv_loss = self.a_loss(z_true, z_pred)
        adv_grad = {v: g for (g, v) in self.a_optimizer.compute_gradients(adv_loss)}
        adv_min = self.a_optimizer.minimize(adv_loss, var_list=self.adversary.trainable_weights)

        # Train Predictor
        pred_grad = []

        for (g, v) in self.p_optimizer.compute_gradients(pred_loss):
            unit_adv = tf_normalize(adv_grad[v])
            g -= tf.math.reduce_sum(g * unit_adv) * unit_adv
            g -= self.debias_param * adv_grad[v]
            pred_grad.append((g, v))
            pred_min = self.p_optimizer.apply_gradients(pred_grad)

        return {"pred_loss": pred_loss, "adv_loss": adv_loss}

    def test_step(self, data):
        # Pass tf.data.Dataset element based on numpy arrays
        x, y_true, z_true = data

        # Compute predictions for Predictor
        y_pred = self.predictor(x, training=False)

        # Input data for Adversary
        adv_x = tf.concat([y_pred, y_true], axis=1)

        # Compute predictions for Adversary
        z_pred = self.adversary(adv_x, training=False)

        # Update losses
        self.p_loss(y_pred, y_true)
        self.a_loss(z_pred, z_true)

当我运行以下代码时出现错误：

mitigator = Mitigator(predictor=Predictor, adversary=Adversary, debias_param=1)

mitigator.compile(a_optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001),
                  p_optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001),
                  a_loss='binary_crossentropy',
                  p_loss='categorical_crossentropy')

mitigator.fit(x=train_text, y=[train_y, z_true], epochs=5)

到目前为止，我已经运行了Keras文章中的代码来检查是否发生了相同的错误，但没有。GAN示例运行良好。

有人知道我如何解决这个错误吗？如果需要更多信息来帮助我，请让我知道。

更新

根据要求，下面给出了错误的完整回溯。

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-19-4f1a8c7cee8b> in <module>()
----> 1 mitigator.fit(x=train_text, y=[train_y, z_true], epochs=5)

3 frames
/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)
    725         max_queue_size=max_queue_size,
    726         workers=workers,
--> 727         use_multiprocessing=use_multiprocessing)
    728 
    729   def evaluate(self,

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training_arrays.py in fit(self, model, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, **kwargs)
    641         steps=steps_per_epoch,
    642         validation_split=validation_split,
--> 643         shuffle=shuffle)
    644 
    645     if validation_data:

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, batch_size, check_steps, steps_name, steps, validation_split, shuffle, extract_tensors_from_dataset)
   2430     is_compile_called = False
   2431     if not self._is_compiled and self.optimizer:
-> 2432       self._compile_from_inputs(all_inputs, y_input, x, y)
   2433       is_compile_called = True
   2434 

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training.py in _compile_from_inputs(self, all_inputs, target, orig_inputs, orig_target)
   2665         sample_weight_mode=self.sample_weight_mode,
   2666         run_eagerly=self.run_eagerly,
-> 2667         experimental_run_tf_function=self._experimental_run_tf_function)
   2668 
   2669   # TODO(omalleyt): Consider changing to a more descriptive function name.

TypeError: compile() got an unexpected keyword argument 'optimizer'

Answer 1

您可以在compile()或train_step()函数中设置optimizer。

class Mitigator(Model):

    def __init__(self, predictor, adversary, debias_param):
        super(Mitigator, self).__init__()
        self.predictor = predictor
        self.adversary = adversary
        self.debias_param = debias_param
             
    def compile(self):
        super(Mitigator, self).compile()
        self.a_optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)
        self.p_optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)

def train_step(self, data):
        a_optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)
        p_optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)
        # Pass tf.data.Dataset element based on numpy arrays
        x, y_true, z_true = data

设置self._is_compiled = True只是让代码认为你的代码已经编译了，我认为这是行不通的。

Answer 2

检查您的tensorflow版本

导入tensorflow print(tensorflow.version)

必须高于v2.4

因此，如果您的版本低于此版本，您有两个选择，要么安装最新版本的Tensorflow和keras，要么使用google colabs。

由于这些定制的训练方法和一些预处理API包含在最近版本的keras和tensorflow中。