Tensorflow 2.0 - LSTM 状态和输入大小

Posted 2023-02-16

【中文标题】Tensorflow 2.0 - LSTM 状态和输入大小

【英文标题】：Tensorflow 2.0 - LSTM statefulness and input size

【发布时间】：2020-09-10 17:42:36

【问题描述】：

对于强化学习中的一个特定问题（受this paper 的启发），我使用了一个 RNN，它输入了形状为 (batch_size, time_steps, features) = (1,1,1) 的数据，用于 L 数据-点，然后一个“循环”结束；使用 LSTM 单元。我使用的是 lstm.stateful = True，在 L 馈送到网络后，我调用 lstm.reset_states()。

在一个周期和另一个周期之间，并且在调用 lstm.reset_states() 之后，我想在形状 (batch_size, time_steps, features) = (L,1) 的输入数据上评估网络的输出,1);然后继续使用输入为 batch_size = 1 的 RNN。

此外，我希望代码尽可能优化，为此我尝试通过 @tf.function 装饰器使用 AutoGraph。

问题是我遇到了一个错误，可以通过以下示例重新创建（请注意，如果删除 @tf.function，一切正常，但我不明白为什么？）

import tensorflow as tf
import numpy as np


class Actor(tf.keras.Model):
    def __init__(self):
        super(Actor,self).__init__()
        self.lstm = tf.keras.layers.LSTM(5, return_sequences=True, stateful=True, input_shape=(None,None,1))#, input_shape=(None,None,1))

    def call(self, inputs):
        feat= self.lstm(inputs)
        return feat

actor = Actor()

@tf.function
def g(actor):
    context1 = tf.reshape(np.array([0.]*10),(10,1,1))
    actor(context1)
    actor.reset_states()
    actor.lstm.stateful=False
    context = tf.reshape(np.array([0.]),(1,1,1))
    actor(context)

g(actor)    



---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-28-4487772bee64> in <module>
     23     actor(context)
     24 
---> 25 g(actor)

~/.local/lib/python3.6/site-packages/tensorflow/python/eager/def_function.py in __call__(self, *args, **kwds)
    578         xla_context.Exit()
    579     else:
--> 580       result = self._call(*args, **kwds)
    581 
    582     if tracing_count == self._get_tracing_count():

~/.local/lib/python3.6/site-packages/tensorflow/python/eager/def_function.py in _call(self, *args, **kwds)
    625       # This is the first call of __call__, so we have to initialize.
    626       initializers = []
--> 627       self._initialize(args, kwds, add_initializers_to=initializers)
    628     finally:
    629       # At this point we know that the initialization is complete (or less

~/.local/lib/python3.6/site-packages/tensorflow/python/eager/def_function.py in _initialize(self, args, kwds, add_initializers_to)
    504     self._concrete_stateful_fn = (
    505         self._stateful_fn._get_concrete_function_internal_garbage_collected(  # pylint: disable=protected-access
--> 506             *args, **kwds))
    507 
    508     def invalid_creator_scope(*unused_args, **unused_kwds):

~/.local/lib/python3.6/site-packages/tensorflow/python/eager/function.py in _get_concrete_function_internal_garbage_collected(self, *args, **kwargs)
   2444       args, kwargs = None, None
   2445     with self._lock:
-> 2446       graph_function, _, _ = self._maybe_define_function(args, kwargs)
   2447     return graph_function
   2448 

~/.local/lib/python3.6/site-packages/tensorflow/python/eager/function.py in _maybe_define_function(self, args, kwargs)
   2775 
   2776       self._function_cache.missed.add(call_context_key)
-> 2777       graph_function = self._create_graph_function(args, kwargs)
   2778       self._function_cache.primary[cache_key] = graph_function
   2779       return graph_function, args, kwargs

~/.local/lib/python3.6/site-packages/tensorflow/python/eager/function.py in _create_graph_function(self, args, kwargs, override_flat_arg_shapes)
   2665             arg_names=arg_names,
   2666             override_flat_arg_shapes=override_flat_arg_shapes,
-> 2667             capture_by_value=self._capture_by_value),
   2668         self._function_attributes,
   2669         # Tell the ConcreteFunction to clean up its graph once it goes out of

~/.local/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)
    979         _, original_func = tf_decorator.unwrap(python_func)
    980 
--> 981       func_outputs = python_func(*func_args, **func_kwargs)
    982 
    983       # invariant: `func_outputs` contains only Tensors, CompositeTensors,

~/.local/lib/python3.6/site-packages/tensorflow/python/eager/def_function.py in wrapped_fn(*args, **kwds)
    439         # __wrapped__ allows AutoGraph to swap in a converted function. We give
    440         # the function a weak reference to itself to avoid a reference cycle.
--> 441         return weak_wrapped_fn().__wrapped__(*args, **kwds)
    442     weak_wrapped_fn = weakref.ref(wrapped_fn)
    443 

~/.local/lib/python3.6/site-packages/tensorflow/python/framework/func_graph.py in wrapper(*args, **kwargs)
    966           except Exception as e:  # pylint:disable=broad-except
    967             if hasattr(e, "ag_error_metadata"):
--> 968               raise e.ag_error_metadata.to_exception(e)
    969             else:
    970               raise

ValueError: in user code:

    <ipython-input-28-4487772bee64>:23 g  *
        actor(context)
    <ipython-input-28-4487772bee64>:11 call  *
        feat= self.lstm(inputs)
    /home/cooper-cooper/.local/lib/python3.6/site-packages/tensorflow/python/keras/layers/recurrent.py:654 __call__  **
        return super(RNN, self).__call__(inputs, **kwargs)
    /home/cooper-cooper/.local/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py:886 __call__
        self.name)
    /home/cooper-cooper/.local/lib/python3.6/site-packages/tensorflow/python/keras/engine/input_spec.py:227 assert_input_compatibility
        ', found shape=' + str(shape))

    ValueError: Input 0 is incompatible with layer lstm_7: expected shape=(10, None, 1), found shape=[1, 1, 1]

【参考方案1】：

如果有人感兴趣，我在以下帖子中找到了答案，我的解决方法如下：

import tensorflow as tf
import numpy as np


class Actor(tf.keras.Model):
    def __init__(self):
        super(Actor,self).__init__()
        self.lstm = tf.keras.layers.LSTM(5, return_sequences=True, stateful=True,input_shape=(1,1))#, input_shape=(None,None,1))

    def call(self, inputs):
        feat= self.lstm(inputs)
        return feat

    def reset_states_workaround(self, new_batch_size=1):
        self.lstm.states = [tf.Variable(tf.random.uniform((new_batch_size,5))), tf.Variable(tf.random.uniform((new_batch_size,5)))]
        self.lstm.input_spec = [tf.keras.layers.InputSpec(shape=(new_batch_size,None,1), ndim=3)]

然后，在使用 @tf.function 的两个不同调用之间，我会这样做：

actor = Actor()
@tf.function
def g(actor):
    context1 = tf.reshape(np.array([0.]*10),(10,1,1))
    preds = actor(context1)
    return preds

g(actor)    
actor.reset_states_workaround(new_batch_size=1)
@tf.function
def g2(actor):
    context1 = tf.reshape(np.array([0.]*1),(1,1,1))
    preds = actor(context1)
    return preds

g2(actor)    

在@tf.function 内部使用actor.reset_states_workaround(new_batch_size=1) 会出现问题：ValueError: tf.function-decorated function tried to create variables on non-first call.，这就是我在外部使用它的原因。

【讨论】：

你能创建变量而不是__init__而不是reset_states_workaround吗？然后reset_states 将只分配给这些变量，您将避免“函数试图创建变量”错误。