python - 在基础 Tensorflow 2.0 中运行简单回归

Question

我正在学习 Tensorflow 2.0，我认为在 Tensorflow 中实现最基本的简单线性回归是个好主意。不幸的是，我遇到了几个问题，我想知道这里是否有人可以提供帮助。

考虑以下设置:

import tensorflow as tf # 2.0.0-alpha0
import numpy as np

x_data = np.random.randn(2000, 1)

w_real = [0.7] # coefficients
b_real = -0.2 # global bias
noise = np.random.randn(1, 2000) * 0.5 # level of noise
y_data = np.matmul(w_real, x_data.T) + b_real + noise

现在开始模型定义:

# modelling this data with tensorflow (manually!)
class SimpleRegressionNN(tf.keras.Model):

    def __init__(self):
        super(SimpleRegressionNN, self).__init__()
        self.input_layer = tf.keras.layers.Input
        self.output_layer = tf.keras.layers.Dense(1)


    def call(self, data_input):
        model = self.input_layer(data_input)
        model = self.output_layer(model)
        # open question: how to account for the intercept/bias term?
        # Ideally, we'd want to generate preds as matmult(X,W) + b
        return model

nn_regressor = SimpleRegressionNN()

reg_loss = tf.keras.losses.MeanSquaredError()
reg_optimiser = tf.keras.optimizers.SGD(0.1)
metric_accuracy = tf.keras.metrics.mean_squared_error

# define forward step
@tf.function
def train_step(x_sample, y_sample):
  with tf.GradientTape() as tape:
    predictions = nn_regressor(x_sample)
    loss = reg_loss(y_sample, predictions)
    gradients = tape.gradient(loss, nn_regressor.trainable_variables) # had to indent this!
  reg_optimiser.apply_gradients(zip(gradients, nn_regressor.trainable_variables))  
  metric_accuracy(y_sample, predictions)

#%%
# run the model
for epoch in range(10):
    for x_point, y_point in zip(x_data.T[0], y_data[0]): # batch of 1
        train_step(x_sample=x_point, y_sample=y_point)
    print("MSE: {}".format(metric_accuracy.result()))

不幸的是，我收到以下错误:

TypeError: You are attempting to use Python control flow in a layer that was not declared to be dynamic. Pass `dynamic=True` to the class constructor.
Encountered error:
"""
Tensor objects are only iterable when eager execution is enabled. To iterate over this tensor use tf.map_fn.
"""

完整的错误输出在这里:

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

TypeError                                 Traceback (most recent call last)

/anaconda3/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py in __call__(self, inputs, *args, **kwargs)
    611                                                       inputs)) as auto_updater:
--> 612                 outputs = self.call(inputs, *args, **kwargs)
    613                 auto_updater.set_outputs(outputs)

<ipython-input-5-8464ad8bcf07> in call(self, data_input)
      7     def call(self, data_input):
----> 8         model = self.input_layer(data_input)
      9         model = self.output_layer(model)

/anaconda3/lib/python3.6/site-packages/tensorflow/python/keras/engine/input_layer.py in Input(shape, batch_size, name, dtype, sparse, tensor, **kwargs)
    232       sparse=sparse,
--> 233       input_tensor=tensor)
    234   # Return tensor including `_keras_history`.

/anaconda3/lib/python3.6/site-packages/tensorflow/python/keras/engine/input_layer.py in __init__(self, input_shape, batch_size, dtype, input_tensor, sparse, name, **kwargs)
     93       if input_shape is not None:
---> 94         batch_input_shape = (batch_size,) + tuple(input_shape)
     95       else:

/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in __iter__(self)
    448       raise TypeError(
--> 449           "Tensor objects are only iterable when eager execution is "
    450           "enabled. To iterate over this tensor use tf.map_fn.")

TypeError: Tensor objects are only iterable when eager execution is enabled. To iterate over this tensor use tf.map_fn.


During handling of the above exception, another exception occurred:

TypeError                                 Traceback (most recent call last)

<ipython-input-22-e1bde858b0fc> in <module>()
      3     #train_step(x_sample=x_data.T[0], y_sample=y_data[0])
      4     for x_point, y_point in zip(x_data.T[0], y_data[0]):
----> 5         train_step(x_sample=x_point, y_sample=y_point)
      6     print("MSE: {}".format(metric_accuracy.result()))
      7 

/anaconda3/lib/python3.6/site-packages/tensorflow/python/eager/def_function.py in __call__(self, *args, **kwds)
    416       # In this case we have not created variables on the first call. So we can
    417       # run the first trace but we should fail if variables are created.
--> 418       results = self._stateful_fn(*args, **kwds)
    419       if self._created_variables:
    420         raise ValueError("Creating variables on a non-first call to a function"

/anaconda3/lib/python3.6/site-packages/tensorflow/python/eager/function.py in __call__(self, *args, **kwargs)
   1285   def __call__(self, *args, **kwargs):
   1286     """Calls a graph function specialized to the inputs."""
-> 1287     graph_function, args, kwargs = self._maybe_define_function(args, kwargs)
   1288     return graph_function._filtered_call(args, kwargs)  # pylint: disable=protected-access
   1289 

/anaconda3/lib/python3.6/site-packages/tensorflow/python/eager/function.py in _maybe_define_function(self, args, kwargs)
   1609           relaxed_arg_shapes)
   1610       graph_function = self._create_graph_function(
-> 1611           args, kwargs, override_flat_arg_shapes=relaxed_arg_shapes)
   1612       self._function_cache.arg_relaxed[rank_only_cache_key] = graph_function
   1613 

/anaconda3/lib/python3.6/site-packages/tensorflow/python/eager/function.py in _create_graph_function(self, args, kwargs, override_flat_arg_shapes)
   1510             arg_names=arg_names,
   1511             override_flat_arg_shapes=override_flat_arg_shapes,
-> 1512             capture_by_value=self._capture_by_value),
   1513         self._function_attributes)
   1514 

/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/func_graph.py in func_graph_from_py_func(name, python_func, args, kwargs, signature, func_graph, autograph, autograph_options, add_control_dependencies, arg_names, op_return_value, collections, capture_by_value, override_flat_arg_shapes)
    692                                           converted_func)
    693 
--> 694       func_outputs = python_func(*func_args, **func_kwargs)
    695 
    696       # invariant: `func_outputs` contains only Tensors, IndexedSlices,

/anaconda3/lib/python3.6/site-packages/tensorflow/python/eager/def_function.py in wrapped_fn(*args, **kwds)
    315         # __wrapped__ allows AutoGraph to swap in a converted function. We give
    316         # the function a weak reference to itself to avoid a reference cycle.
--> 317         return weak_wrapped_fn().__wrapped__(*args, **kwds)
    318     weak_wrapped_fn = weakref.ref(wrapped_fn)
    319 

/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/func_graph.py in wrapper(*args, **kwargs)
    684                   optional_features=autograph_options,
    685                   force_conversion=True,
--> 686               ), args, kwargs)
    687 
    688         # Wrapping around a decorator allows checks like tf_inspect.getargspec

/anaconda3/lib/python3.6/site-packages/tensorflow/python/autograph/impl/api.py in converted_call(f, owner, options, args, kwargs)
    390     return _call_unconverted(f, args, kwargs)
    391 
--> 392   result = converted_f(*effective_args, **kwargs)
    393 
    394   # The converted function's closure is simply inserted into the function's

/var/folders/8_/pl9fgq297ld3b7kgy5tmvf700000gn/T/tmpluzodr7d.py in tf__train_step(x_sample, y_sample)
      2 def tf__train_step(x_sample, y_sample):
      3   with tf.GradientTape() as tape:
----> 4     predictions = ag__.converted_call(nn_regressor, None, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(tf.function, defun, ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=(), internal_convert_user_code=True), (x_sample,), {})
      5     loss = ag__.converted_call(reg_loss, None, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(tf.function, defun_1, ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=(), internal_convert_user_code=True), (y_sample, predictions), {})
      6     gradients = ag__.converted_call('gradient', tape, ag__.ConversionOptions(recursive=True, verbose=0, strip_decorators=(tf.function, defun_2, ag__.convert, ag__.do_not_convert, ag__.converted_call), force_conversion=False, optional_features=(), internal_convert_user_code=True), (loss, nn_regressor.trainable_variables), {})

/anaconda3/lib/python3.6/site-packages/tensorflow/python/autograph/impl/api.py in converted_call(f, owner, options, args, kwargs)
    265 
    266   if not options.force_conversion and conversion.is_whitelisted_for_graph(f):
--> 267     return _call_unconverted(f, args, kwargs)
    268 
    269   # internal_convert_user_code is for example turned off when issuing a dynamic

/anaconda3/lib/python3.6/site-packages/tensorflow/python/autograph/impl/api.py in _call_unconverted(f, args, kwargs)
    186     return f.__self__.call(args, kwargs)
    187 
--> 188   return f(*args, **kwargs)
    189 
    190 

/anaconda3/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py in __call__(self, inputs, *args, **kwargs)
    623                                   'dynamic. Pass `dynamic=True` to the class '
    624                                   'constructor.\nEncountered error:\n"""\n' +
--> 625                                   exception_str + '\n"""')
    626               raise
    627           else:

TypeError: You are attempting to use Python control flow in a layer that was not declared to be dynamic. Pass `dynamic=True` to the class constructor.
Encountered error:
"""
Tensor objects are only iterable when eager execution is enabled. To iterate over this tensor use tf.map_fn.
"""

麻烦的是，2.0默认设置为eager execution!

除了这个问题，我还有几个额外的问题:

1. 在这里解释截距项的最佳方法是什么？
2. 一般方法是否合理，或者我在这里做了什么奇怪的事情？ (忽略批量大小和我必须验证数据的事实，这只是一个玩具示例)

非常感谢!

Answer 1

我有以下评论:

• 您的SimpleRegression 模型中不需要Input 层。另外，不要通过“model”名称调用层的张量输出(就像在 call() 方法中所做的那样)。这真的很令人困惑。
• 您没有将正确的形状传递给您的 train_step 函数。它期望在您传递 (input_dim, ) 时收到 (n_samples, input_dim)。
• 请记住，在 tensorflow 中，张量的第一个维度始终是批量大小(即样本数)。永远这样使用它，不要转置。
• 为什么调用 metric_accuracy = tf.keras.metrics.mean_squared_error 精度？你有一个回归问题，回归中没有准确性这样的东西。另外，为什么要定义两次并计算两次 mse？
• 如果您使用 tf.convert_to_tensor() 转换数据，执行速度会更快。
• 函数 train_step() 执行前向和后向传递，而不仅仅是前向传递。
• 使用小型数据集作为玩具示例(2-10 个样本，而不是 2000 个)，尤其是当您不知道自己的代码是否有效时!
• 您的函数 train_step() 不返回任何内容，您希望如何打印 mse 损失的值。

这是您的代码的更正版本:

import tensorflow as tf # 2.0.0-alpha0
import numpy as np

x_data = np.random.randn(5, 2)

w_real = 0.7 # coefficients
b_real = -0.2 # global bias
noise = np.random.randn(5, 2) * 0.01 # level of noise
y_data = w_real * x_data + b_real + noise

class SimpleRegressionNN(tf.keras.Model):

    def __init__(self):
        super(SimpleRegressionNN, self).__init__()

        self.output_layer = tf.keras.layers.Dense(1, input_shape=(2, ))

    def call(self, data_input):
        result = self.output_layer(data_input)
        return result

reg_loss = tf.keras.losses.MeanSquaredError()
reg_optimiser = tf.keras.optimizers.SGD(0.1)

nn_regressor = SimpleRegressionNN()

@tf.function
def train_step(x_sample, y_sample):
    with tf.GradientTape() as tape:
        predictions = nn_regressor(x_sample)
        loss = reg_loss(y_sample, predictions)
    gradients = tape.gradient(loss, nn_regressor.trainable_variables) # had to indent this!
    reg_optimiser.apply_gradients(zip(gradients, nn_regressor.trainable_variables))  

    return loss

for x_point, y_point in zip(x_data, y_data): # batch of 1
    x_point, y_point = tf.convert_to_tensor([x_point]), tf.convert_to_tensor([y_point])
    mse = train_step(x_sample=x_point, y_sample=y_point)
    print("MSE: {}".format(mse.numpy()))