python - 如何在 tensorflow 2.0 中使用额外输入进行自定义损失

Question

我在使用 tf.keras 和数据集获取带有额外参数的自定义损失函数以在 TF 2.0 中工作时遇到了很多麻烦。

在下面的例子中，额外的参数是模型的输入数据，它包含在 Dataset 中。在 1.14 的情况下，我会在数据集上运行 .make_one_shot_iterator().get_next()，然后将我得到的张量传递给损失函数。同样的事情在 2.0 中不起作用。

class WeightedSDRLoss(keras.losses.Loss):

    def __init__(self, noisy_signal, reduction=keras.losses.Reduction.AUTO, name='WeightedSDRLoss'):
        super().__init__(reduction=reduction, name=name)
        self.noisy_signal = noisy_signal

    def sdr_loss(self, sig_true, sig_pred):
        return (-tf.reduce_mean(sig_true * sig_pred) /
                tf.reduce_mean(tf.norm(tensor=sig_pred) * tf.norm(tensor=sig_true)))

    def call(self, y_true, y_pred):
        noise_true = self.noisy_signal - y_true
        noise_pred = self.noisy_signal - y_pred
        alpha = (tf.reduce_mean(tf.square(y_true)) /
                 tf.reduce_mean(tf.square(y_true) + tf.square(self.noisy_signal - y_pred)))
        return alpha * self.sdr_loss(y_true, y_pred) + (1 - alpha) * self.sdr_loss(noise_true, noise_pred)

data_x = np.random.rand(5, 4, 1)
data_y = np.random.rand(5, 4, 1)

x = keras.layers.Input([4, 1])
y = keras.layers.Activation('tanh')(x)
model = keras.models.Model(inputs=x, outputs=y)

train_dataset = tf.data.Dataset.from_tensor_slices((data_x, data_y))
x_dataset = train_dataset.map(lambda x, y: x)

model.compile(loss=WeightedSDRLoss(x_dataset), optimizer='Adam')
model.fit(train_dataset)

但是我在 tensorflow 中得到以下错误:

_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/training/tracking/base.py:457: in _method_wrapper
    result = method(self, *args, **kwargs)
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py:377: in compile
    self._compile_weights_loss_and_weighted_metrics()
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/training/tracking/base.py:457: in _method_wrapper
    result = method(self, *args, **kwargs)
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py:1618: in _compile_weights_loss_and_weighted_metrics
    self.total_loss = self._prepare_total_loss(masks)
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py:1678: in _prepare_total_loss
    per_sample_losses = loss_fn.call(y_true, y_pred)
...:144: in call
    noise_true = self.noisy_signal - y_true
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/ops/math_ops.py:924: in r_binary_op_wrapper
    x = ops.convert_to_tensor(x, dtype=y.dtype.base_dtype, name="x")
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py:1184: in convert_to_tensor
    return convert_to_tensor_v2(value, dtype, preferred_dtype, name)
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py:1242: in convert_to_tensor_v2
    as_ref=False)
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py:1296: in internal_convert_to_tensor
    ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/framework/constant_op.py:286: in _constant_tensor_conversion_function
    return constant(v, dtype=dtype, name=name)
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/framework/constant_op.py:227: in constant
    allow_broadcast=True)
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/framework/constant_op.py:265: in _constant_impl
    allow_broadcast=allow_broadcast))
../../anaconda3/envs/.../lib/python3.6/site-packages/tensorflow_core/python/framework/tensor_util.py:449: in make_tensor_proto
    _AssertCompatible(values, dtype)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

values = <MapDataset shapes: (...), types: tf.float32>
dtype = tf.float32

    def _AssertCompatible(values, dtype):
      if dtype is None:
        fn = _check_not_tensor
      else:
        try:
          fn = _TF_TO_IS_OK[dtype]
        except KeyError:
          # There isn't a specific fn, so we try to do the best possible.
          if dtype.is_integer:
            fn = _check_int
          elif dtype.is_floating:
            fn = _check_float
          elif dtype.is_complex:
            fn = _check_complex
          elif dtype.is_quantized:
            fn = _check_quantized
          else:
            fn = _check_not_tensor

      try:
        fn(values)
      except ValueError as e:
        [mismatch] = e.args
        if dtype is None:
          raise TypeError("List of Tensors when single Tensor expected")
        else:
          raise TypeError("Expected %s, got %s of type '%s' instead." %
>                         (dtype.name, repr(mismatch), type(mismatch).__name__))
E         TypeError: Expected float32, got <MapDataset shapes: (...), types: tf.float32> of type 'MapDataset' instead.

问题似乎是我将数据集传递给损失函数，但它需要一个急切求值的张量。

相反，我尝试将输入层传递到自定义损失中，但这也不起作用:

data_x = np.random.rand(5, 4, 1)
data_y = np.random.rand(5, 4, 1)

x = keras.layers.Input(shape=[4, 1])
y = keras.layers.Activation('tanh')(x)
model = keras.models.Model(inputs=x, outputs=y)

train_dataset = tf.data.Dataset.from_tensor_slices((data_x, data_y)).batch(1)

model.compile(loss=WeightedSDRLoss(x), optimizer='Adam')
model.fit(train_dataset)

相反，我得到了错误:

op_name = '__inference_distributed_function_169', num_outputs = 2
inputs = [<tf.Tensor: id=82, shape=(), dtype=resource, numpy=<unprintable>>, <tf.Tensor: id=83, shape=(), dtype=variant, numpy=<unprintable>>, <tf.Tensor 'input_1:0' shape=(None, 4, 1) dtype=float32>]
attrs = ('executor_type', '', 'config_proto', b'\n\x07\n\x03GPU\x10\x00\n\x07\n\x03CPU\x10\x012\x02J\x008\x01')
ctx = <tensorflow.python.eager.context.Context object at 0x11785f4e0>
name = None

    def quick_execute(op_name, num_outputs, inputs, attrs, ctx, name=None):
      """Execute a TensorFlow operation.

      Args:
        op_name: Name of the TensorFlow operation (see REGISTER_OP in C++ code) to
          execute.
        num_outputs: The number of outputs of the operation to fetch.
                     (Explicitly provided instead of being inferred for performance
                     reasons).
        inputs: A list of inputs to the operation. Each entry should be a Tensor, or
          a value which can be passed to the Tensor constructor to create one.
        attrs: A tuple with alternating string attr names and attr values for this
          operation.
        ctx: The value of context.context().
        name: Customized name for the operation.

      Returns:
        List of output Tensor objects. The list is empty if there are no outputs

      Raises:
        An exception on error.
      """
      device_name = ctx.device_name
      # pylint: disable=protected-access
      try:
        ctx.ensure_initialized()
        tensors = pywrap_tensorflow.TFE_Py_Execute(ctx._handle, device_name,
                                                   op_name, inputs, attrs,
>                                                  num_outputs)
E                                                  TypeError: An op outside of the function building code is being passed
E                                                  a "Graph" tensor. It is possible to have Graph tensors
E                                                  leak out of the function building context by including a
E                                                  tf.init_scope in your function building code.
E                                                  For example, the following function will fail:
E                                                    @tf.function
E                                                    def has_init_scope():
E                                                      my_constant = tf.constant(1.)
E                                                      with tf.init_scope():
E                                                        added = my_constant * 2
E                                                  The graph tensor has name: input_1:0

../../../lib/python3.6/site-packages/tensorflow_core/python/eager/execute.py:61: TypeError

During handling of the above exception, another exception occurred:

    def test_loss():

        data_x = np.random.rand(5, 4, 1)
        data_y = np.random.rand(5, 4, 1)

        x = keras.layers.Input(shape=[4, 1])
        y = keras.layers.Activation('tanh')(x)
        model = keras.models.Model(inputs=x, outputs=y)

        train_dataset = tf.data.Dataset.from_tensor_slices((data_x, data_y)).batch(1)
        print(train_dataset)

        model.compile(loss=WeightedSDRLoss(x))
>       model.fit(train_dataset)

test_preprocess.py:162: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
../../../lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py:734: in fit
    use_multiprocessing=use_multiprocessing)
../../../lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training_v2.py:324: in fit
    total_epochs=epochs)
../../../lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training_v2.py:123: in run_one_epoch
    batch_outs = execution_function(iterator)
../../../training_v2_utils.py:86: in execution_function
    distributed_function(input_fn))
../../../def_function.py:445: in __call__
    return self._concrete_stateful_fn._filtered_call(canon_args, canon_kwds)  # pylint: disable=protected-access
../../../function.py:1141: in _filtered_call
    self.captured_inputs)
../../../function.py:1224: in _call_flat
    ctx, args, cancellation_manager=cancellation_manager)
../../../function.py:511: in call
    ctx=ctx)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

op_name = '__inference_distributed_function_169', num_outputs = 2
inputs = [<tf.Tensor: id=82, shape=(), dtype=resource, numpy=<unprintable>>, <tf.Tensor: id=83, shape=(), dtype=variant, numpy=<unprintable>>, <tf.Tensor 'input_1:0' shape=(None, 4, 1) dtype=float32>]
attrs = ('executor_type', '', 'config_proto', b'\n\x07\n\x03GPU\x10\x00\n\x07\n\x03CPU\x10\x012\x02J\x008\x01')
ctx = <tensorflow.python.eager.context.Context object at 0x11785f4e0>
name = None

    def quick_execute(op_name, num_outputs, inputs, attrs, ctx, name=None):
      """Execute a TensorFlow operation.

      Args:
        op_name: Name of the TensorFlow operation (see REGISTER_OP in C++ code) to
          execute.
        num_outputs: The number of outputs of the operation to fetch.
                     (Explicitly provided instead of being inferred for performance
                     reasons).
        inputs: A list of inputs to the operation. Each entry should be a Tensor, or
          a value which can be passed to the Tensor constructor to create one.
        attrs: A tuple with alternating string attr names and attr values for this
          operation.
        ctx: The value of context.context().
        name: Customized name for the operation.

      Returns:
        List of output Tensor objects. The list is empty if there are no outputs

      Raises:
        An exception on error.
      """
      device_name = ctx.device_name
      # pylint: disable=protected-access
      try:
        ctx.ensure_initialized()
        tensors = pywrap_tensorflow.TFE_Py_Execute(ctx._handle, device_name,
                                                   op_name, inputs, attrs,
                                                   num_outputs)
      except core._NotOkStatusException as e:
        if name is not None:
          message = e.message + " name: " + name
        else:
          message = e.message
        six.raise_from(core._status_to_exception(e.code, message), None)
      except TypeError as e:
        keras_symbolic_tensors = [
            x for x in inputs if ops._is_keras_symbolic_tensor(x)
        ]
        if keras_symbolic_tensors:
          raise core._SymbolicException(
              "Inputs to eager execution function cannot be Keras symbolic "
>             "tensors, but found {}".format(keras_symbolic_tensors))
E         tensorflow.python.eager.core._SymbolicException: Inputs to eager execution function cannot be Keras symbolic tensors, but found [<tf.Tensor 'input_1:0' shape=(None, 4, 1) dtype=float32>]

关于如何让它发挥作用的任何想法？我不想使用自定义训练循环，因为那样我就失去了 keras 的很多便利。

Answer 1

仅 TF 2.0.0-beta1 不 rc0

对我来说你的第二次尝试

data_x = np.random.rand(5, 4, 1)
data_y = np.random.rand(5, 4, 1)

x = keras.layers.Input([4, 1])
y = keras.layers.Activation('tanh')(x)
model = keras.models.Model(inputs=x, outputs=y)

train_dataset = tf.data.Dataset.from_tensor_slices((data_x, data_y)).batch(1)

model.compile(loss=WeightedSDRLoss(x), optimizer='Adam')
model.fit(train_dataset)

工作正常。 我只需要指定一个优化器。

我只收到警告 Expected a shuffled dataset but input dataset `x` is not shuffled。请在输入数据集上调用 `shuffle()`。 这可以通过在训练前添加 train_dataset = train_dataset.shuffle(1) 来避免。