python - Keras tensorflow 自定义损失使用带有 Eager 执行的中间 tensorflow 概率层

Question

当使用中间张量流概率DistributionLambda层为模型创建自定义损失时，无需 Eager 执行很容易，因为该层可以存储在变量中，然后从损失内部使用。启用 Eager 执行时，它会失败，因为它说没有评估网络变量的损失。如果DistributionLambda我在损失中使用前一层的输入参数而不是层，就地创建分布，那么它会失败，因为该输入层是一个符号张量。有关如何执行此操作的任何帮助？

这是一个完整的 colab 示例，这里是示例代码：

import numpy as np
from tensorflow.keras import layers
from tensorflow.keras import models
from tensorflow import math
from tensorflow import version
from tensorflow_probability import distributions
from tensorflow_probability import layers as tfp_layers

print(version.GIT_VERSION)
print(version.VERSION)

x = np.array([[1.0], [2.0], [3.0], [4.0], [5.0]])
y = np.array([[6.0, 7.0], [7.0, 8.0], [8.0, 9.0], [9.0, 10.0], [10.0, 11.0]])


def normal_distribution(param):
    """Normal distribution parameterization from input."""
    return distributions.Normal(param[:, 0], math.sqrt(math.exp(param[:, 1])))


class DistribNN(object):
    """Neural net with a distributional output."""

    def __init__(self):
        """Constructor of DistribNN."""
        self._distribution_layer = None
        self._predist = None
        self._the_model = None

    def _loss(self):
        """Correct final loss for this model."""

        def likelihood_loss(y_true, y_pred):
            """Adding negative log likelihood loss."""
            # return -distributions.Normal(self._predist[:, 0],
            #                              math.sqrt(math.exp(self._predist[:, 1]))).log_prob(
            #     y_true + y_pred - y_pred + 0.001)
            # return -normal_distribution(self._predist).log_prob(
            #     y_true + y_pred - y_pred + 0.001)
            return -self._distribution_layer.log_prob(y_true + 0.001)

        return likelihood_loss

    def create_net(self):
        """Create neural net model.."""
        input_layer = layers.Input((1,))
        dense_layer = layers.Dense(10)(input_layer)
        params_dense_layer = layers.Dense(4)(dense_layer)
        self._predist = layers.Reshape((2, 2))(params_dense_layer)
        self._distribution_layer = tfp_layers.DistributionLambda(
            make_distribution_fn=normal_distribution)(self._predist)
        last_layer = layers.Lambda(lambda param: param - 0.001)(self._distribution_layer)
        self._the_model = models.Model(inputs=input_layer, outputs=last_layer)
        self._the_model.compile(loss=self._loss(), optimizer='Adam')

    def fit(self, x, y):
        self._the_model.fit(x=x, y=y, epochs=2)


one_model = DistribNN()
one_model.create_net()
one_model.fit(x, y)

给出错误：

v2.3.0-0-gb36436b087
2.3.0
Epoch 1/2

---------------------------------------------------------------------------

ValueError                                Traceback (most recent call last)

<ipython-input-3-6b2d34187a56> in <module>()
     40 one_model = DistribNN()
     41 one_model.create_net()
---> 42 one_model.fit(x, y)

11 frames

/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/func_graph.py in wrapper(*args, **kwargs)
    971           except Exception as e:  # pylint:disable=broad-except
    972             if hasattr(e, "ag_error_metadata"):
--> 973               raise e.ag_error_metadata.to_exception(e)
    974             else:
    975               raise

ValueError: in user code:

    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:806 train_function  *
        return step_function(self, iterator)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:796 step_function  **
        outputs = model.distribute_strategy.run(run_step, args=(data,))
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:1211 run
        return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:2585 call_for_each_replica
        return self._call_for_each_replica(fn, args, kwargs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:2945 _call_for_each_replica
        return fn(*args, **kwargs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:789 run_step  **
        outputs = model.train_step(data)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:757 train_step
        self.trainable_variables)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:2737 _minimize
        trainable_variables))
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:562 _aggregate_gradients
        filtered_grads_and_vars = _filter_grads(grads_and_vars)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:1271 _filter_grads
        ([v.name for _, v in grads_and_vars],))

    ValueError: No gradients provided for any variable: ['dense/kernel:0', 'dense/bias:0', 'dense_1/kernel:0', 'dense_1/bias:0'].

标签： pythontensorflowkeraslosstensorflow-probability