python - TensorFlow 2.0：无法在图形模式下训练具有自定义拟合的子类模型

Question

下面的代码片段是 TensorFlow 模型的 vanila 实现，我在其中使用子类模型和自定义拟合函数（通过train_step和实现test_step）。该代码在急切执行模式（TF2.0 中的默认执行模式）下工作正常，但在图形模式下失败。

import numpy as np
import tensorflow as tf

class Encoder(tf.keras.Model):
    def __init__(self):
        super(Encoder, self).__init__(name = 'Encoder')
        self.input_layer   = tf.keras.layers.Dense(10)
        self.hidden_layer1 = tf.keras.layers.Dense(10)
        self.dropout_laye1 = tf.keras.layers.Dropout(0.2)
        self.hidden_layer2 = tf.keras.layers.Dense(10)        
        self.dropout_laye2 = tf.keras.layers.Dropout(0.2)
        self.hidden_layer3 = tf.keras.layers.Dense(10)
        self.dropout_laye3 = tf.keras.layers.Dropout(0.2)           
        self.output_layer  = tf.keras.layers.Dense(1)
        
    def call(self, input_data, training):
        fx = self.input_layer(input_data)        
        fx = self.hidden_layer1(fx)
        if training:
            fx = self.dropout_laye1(fx)     
        fx = self.hidden_layer2(fx)
        if training:
            fx = self.dropout_laye2(fx) 
        fx = self.hidden_layer3(fx)
        if training:
            fx = self.dropout_laye3(fx) 
        return self.output_layer(fx)

class CustomModelV1(tf.keras.Model):
    def __init__(self):
        super(CustomModelV1, self).__init__()
        self.encoder = Encoder()
    
    def train_step(self, data):
        # Unpack the data. Its structure depends on your model and
        # on what you pass to `fit()`.
        x, y = data

        with tf.GradientTape() as tape:
            y_pred = self.encoder(x, training=True)  # Forward pass
            # Compute the loss value
            # (the loss function is configured in `compile()`)
            loss = self.compiled_loss(y, y_pred, regularization_losses=self.losses)

        # Compute gradients
        trainable_vars = self.trainable_variables
        gradients = tape.gradient(loss, trainable_vars)
        
        # Update weights
        self.optimizer.apply_gradients(zip(gradients, trainable_vars))
        
        # Update metrics (includes the metric that tracks the loss)
        self.compiled_metrics.update_state(y, y_pred)
        
        # Return a dict mapping metric names to current value
        return {m.name: m.result() for m in self.metrics} 

# Just use `fit` as usual
x = tf.data.Dataset.from_tensor_slices(np.random.random((1000, 32)))

y_numpy = np.random.random((1000, 1))
y = tf.data.Dataset.from_tensor_slices(y_numpy)

x_window = x.window(30, shift=10, stride=1)
flat_x = x_window.flat_map(lambda t: t)
flat_x_scaled = flat_x.map(lambda t: t * 2)

y_window = y.window(30, shift=10, stride=1)
flat_y = y_window.flat_map(lambda t: t)
flat_y_scaled = flat_y.map(lambda t: t * 2)

z = tf.data.Dataset.zip((flat_x_scaled, flat_y_scaled)).batch(32).cache().shuffle(buffer_size=32).prefetch(buffer_size=tf.data.experimental.AUTOTUNE)


# Construct and compile an instance of CustomModel
model = CustomModelV1()
model.compile(optimizer="adam", loss="mse", metrics=["mae"])

model.fit(z, epochs=3)

该代码在急切模式下运行良好，但在图形模式下会引发以下错误。我使用禁用了急切执行tf.compat.v1.disable_eager_execution()。

AttributeError                            Traceback (most recent call last)
<ipython-input-4-f7a5b420f08f> in <module>
     27 model.compile(optimizer="adam", loss="mse", metrics=["mae"])
     28 
---> 29 model.fit(z, epochs=3)

~\Anaconda3\envs\tf\lib\site-packages\tensorflow\python\keras\engine\training_v1.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)
    793 
    794     func = self._select_training_loop(x)
--> 795     return func.fit(
    796         self,
    797         x=x,

~\Anaconda3\envs\tf\lib\site-packages\tensorflow\python\keras\engine\training_arrays_v1.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)
    617                                                      steps_per_epoch, x)
    618 
--> 619     x, y, sample_weights = model._standardize_user_data(
    620         x,
    621         y,

~\Anaconda3\envs\tf\lib\site-packages\tensorflow\python\keras\engine\training_v1.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)
   2328     is_compile_called = False
   2329     if not self._is_compiled and self.optimizer:
-> 2330       self._compile_from_inputs(all_inputs, y_input, x, y)
   2331       is_compile_called = True
   2332 

~\Anaconda3\envs\tf\lib\site-packages\tensorflow\python\keras\engine\training_v1.py in _compile_from_inputs(self, all_inputs, target, orig_inputs, orig_target)
   2548       # We need to use `y` to set the model targets.
   2549       if training_utils_v1.has_tensors(target):
-> 2550         target = training_utils_v1.cast_if_floating_dtype_and_mismatch(
   2551             target, self.outputs)
   2552       training_utils_v1.validate_input_types(

~\Anaconda3\envs\tf\lib\site-packages\tensorflow\python\keras\engine\training_utils_v1.py in cast_if_floating_dtype_and_mismatch(targets, outputs)
   1377   if tensor_util.is_tf_type(targets):
   1378     # There is one target, so output[0] should be the only output.
-> 1379     return cast_single_tensor(targets, dtype=outputs[0].dtype)
   1380   new_targets = []
   1381   for target, out in zip(targets, outputs):

AttributeError: 'NoneType' object has no attribute 'dtype'

标签： pythontf.kerastensorflow2.xeager-executiontf.data.dataset