在 TensorFlow 中实现 Theano 操作

Question

在此paper on domain adaptation 之后，我正在尝试在 Tensorflow 中实现以下梯度反转层（为带有 Theano 后端的 Keras 编写，如在此 Keras issue 中找到），因为我的模型在 Theano 上运行不佳。

class GradientReversalLayer(Layer):
    """ Reverse a gradient
    <feedforward> return input x
    <backward> return -lambda * delta
    """
    def __init__(self, hp_lambda, **kwargs):
        super(GradientReversalLayer, self).__init__(**kwargs)
        self.hp_lambda = hp_lambda
        self.gr_op = ReverseGradient(self.hp_lambda)

    def build(self, input_shape):
        self.trainable_weights = []

    def call(self, x, mask=None):
        return self.gr_op(x)

    def get_output_shape_for(self, input_shape):
        return input_shape

    def get_config(self):
        config = {"name": self.__class__.__name__,
                         "lambda": self.hp_lambda}
        base_config = super(GradientReversalLayer, self).get_config()
        return dict(list(base_config.items()) + list(config.items()))

层执行这个操作：

 import theano
    from keras.engine import Layer

    class ReverseGradient(theano.Op):
        """ theano operation to reverse the gradients
        Introduced in http://arxiv.org/pdf/1409.7495.pdf
        """

        view_map = {0: [0]}

        __props__ = ('hp_lambda', )

        def __init__(self, hp_lambda):
            super(ReverseGradient, self).__init__()
            self.hp_lambda = hp_lambda

        def make_node(self, x):
            assert hasattr(self, '_props'), "Your version of theano is too old to support __props__."
            x = theano.tensor.as_tensor_variable(x)
            return theano.Apply(self, [x], [x.type()])

        def perform(self, node, inputs, output_storage):
            xin, = inputs
            xout, = output_storage
            xout[0] = xin

        def grad(self, input, output_gradients):
            return [-self.hp_lambda * output_gradients[0]]

        def infer_shape(self, node, i0_shapes):
            return i0_shapes

为什么不能这样用？

如果我使用 tf 后端运行我的模型并使用 Theano 编写的此函数，我会收到以下错误：

theano.tensor.var.AsTensorError: ('Cannot convert Tensor("concatenate_1/concat:0", shape=(?, ?, 128), dtype=float32) to TensorType', <class 'tensorflow.python.framework.ops.Tensor'>)

这样调用之后：

lstm_concat = concatenate([hidden_out_1, hidden_out_2])
lstm_concat = FlipGradientKeras.GradientReversalLayer(0.31)(lstm_concat)

如何将此操作转换为 TF operation？

关于adding a new operation 的文档只建议用C++ 实现它。

ops codes 显示了一般框架，但我想确保我正在实现的所有东西都与 Theano 操作一样。

我认为它会是这样的：

def ReverseGradient(input_tensor, hp_lambda):

    with ops.name_scope(name, "ReverseGradient", [input_tensor, hp_lambda]) as name:
        input_tensor = ops.convert_to_tensor(input_tensor, name="input_tensor")

但我真的不确定其余的。

提前致谢！

Answer 1

我通过扩展 here 所做的工作解决了这个问题。

这是工作代码：

import tensorflow as tf
from keras.engine import Layer
import keras.backend as K

def reverse_gradient(X, hp_lambda):
    '''Flips the sign of the incoming gradient during training.'''
    try:
        reverse_gradient.num_calls += 1
    except AttributeError:
        reverse_gradient.num_calls = 1

    grad_name = "GradientReversal%d" % reverse_gradient.num_calls

    @tf.RegisterGradient(grad_name)
    def _flip_gradients(op, grad):
        return [tf.negative(grad) * hp_lambda]

    g = K.get_session().graph
    with g.gradient_override_map({'Identity': grad_name}):
        y = tf.identity(X)

    return y

class GradientReversal(Layer):
    '''Flip the sign of gradient during training.'''
    def __init__(self, hp_lambda, **kwargs):
        super(GradientReversal, self).__init__(**kwargs)
        self.supports_masking = False
        self.hp_lambda = hp_lambda

    def build(self, input_shape):
        self.trainable_weights = []

    def call(self, x, mask=None):
        return reverse_gradient(x, self.hp_lambda)

    def get_output_shape_for(self, input_shape):
        return input_shape

    def get_config(self):
        config = {}
        base_config = super(GradientReversal, self).get_config()
        return dict(list(base_config.items()) + list(config.items()))