如何为 state_is_tuple = True 设置 RNN？

Posted 2023-02-16

【中文标题】如何为 state_is_tuple = True 设置 RNN？

【英文标题】：How to set RNN for the state_is_tuple = True?

【发布时间】：2020-08-28 05:24:39

【问题描述】：

class Model:
    def __init__(
        self,
        learning_rate,
        num_layers,
        size,
        size_layer,
        output_size,
        forget_bias = 0.1,
    ):
        def lstm_cell(size_layer):
            return tf.compat.v1.nn.rnn_cell.LSTMCell(size_layer, state_is_tuple = False)

        rnn_cells = tf.compat.v1.nn.rnn_cell.MultiRNNCell(
            [lstm_cell(size_layer) for _ in range(num_layers)],
            state_is_tuple = False,
        )
        self.X = tf.compat.v1.placeholder(tf.float32, (None, None, size))
        self.Y = tf.compat.v1.placeholder(tf.float32, (None, output_size))
        drop = tf.compat.v1.nn.rnn_cell.DropoutWrapper(
            rnn_cells, output_keep_prob = forget_bias
        )
        self.hidden_layer = tf.compat.v1.placeholder(
            tf.float32, (None, num_layers * 2 * size_layer)
        )
        self.outputs, self.last_state = tf.compat.v1.nn.dynamic_rnn(
            drop, self.X, initial_state = self.hidden_layer, dtype = tf.float32
        )
        self.logits = tf.compat.v1.layers.dense(self.outputs[-1], output_size)
        self.cost = tf.reduce_mean(tf.square(self.Y - self.logits))
        self.optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate).minimize(
            self.cost
        )

我上面有这段代码，我一直收到这个警告：

WARNING:tensorflow:<tensorflow.python.ops.rnn_cell_impl.LSTMCell object at 0x0000021BF7EBFEF0>: Using a concatenated state is slower and will soon be deprecated.  Use state_is_tuple=True.

当我更改 state_is_tuple=True 时，它会显示如下错误：

<ipython-input-922-91f013941f83> in __init__(self, learning_rate, num_layers, size, size_layer, output_size, forget_bias)
     25         )
     26         self.outputs, self.last_state = tf.compat.v1.nn.dynamic_rnn(
---> 27             drop, self.X, initial_state = self.hidden_layer, dtype = tf.float32
     28         )
     29         self.logits = tf.compat.v1.layers.dense(self.outputs[-1], output_size)

~\Anaconda3\lib\site-packages\tensorflow_core\python\util\deprecation.py in new_func(*args, **kwargs)
    322               'in a future version' if date is None else ('after %s' % date),
    323               instructions)
--> 324       return func(*args, **kwargs)
    325     return tf_decorator.make_decorator(
    326         func, new_func, 'deprecated',

~\Anaconda3\lib\site-packages\tensorflow_core\python\ops\rnn.py in dynamic_rnn(cell, inputs, sequence_length, initial_state, dtype, parallel_iterations, swap_memory, time_major, scope)
    705         swap_memory=swap_memory,
    706         sequence_length=sequence_length,
--> 707         dtype=dtype)
    708 
    709     # Outputs of _dynamic_rnn_loop are always shaped [time, batch, depth].

~\Anaconda3\lib\site-packages\tensorflow_core\python\ops\rnn.py in _dynamic_rnn_loop(cell, inputs, initial_state, parallel_iterations, swap_memory, sequence_length, dtype)
    914       parallel_iterations=parallel_iterations,
    915       maximum_iterations=time_steps,
--> 916       swap_memory=swap_memory)
    917 
    918   # Unpack final output if not using output tuples.

~\Anaconda3\lib\site-packages\tensorflow_core\python\ops\control_flow_ops.py in while_loop(cond, body, loop_vars, shape_invariants, parallel_iterations, back_prop, swap_memory, name, maximum_iterations, return_same_structure)
   2673         name=name,
   2674         return_same_structure=return_same_structure,
-> 2675         back_prop=back_prop)
   2676 
   2677   with ops.name_scope(name, "while", loop_vars):

~\Anaconda3\lib\site-packages\tensorflow_core\python\ops\while_v2.py in while_loop(cond, body, loop_vars, shape_invariants, parallel_iterations, maximum_iterations, name, return_same_structure, back_prop)
    192         func_graph=util.WhileBodyFuncGraph(
    193             body_name, collections=ops.get_default_graph()._collections),  # pylint: disable=protected-access
--> 194         add_control_dependencies=add_control_dependencies)
    195     # Add external captures of body to the list of loop vars.
    196     # Note that external tensors will be treated as loop invariants, i.e.,

~\Anaconda3\lib\site-packages\tensorflow_core\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)
    976                                           converted_func)
    977 
--> 978       func_outputs = python_func(*func_args, **func_kwargs)
    979 
    980       # invariant: `func_outputs` contains only Tensors, CompositeTensors,

~\Anaconda3\lib\site-packages\tensorflow_core\python\ops\while_v2.py in wrapped_body(loop_counter, maximum_iterations_arg, *args)
    170       # `orig_loop_vars` and `args`, converts flows in `args` to TensorArrays
    171       # and packs it into the structure of `orig_loop_vars`.
--> 172       outputs = body(*_pack_sequence_as(orig_loop_vars, args))
    173       if not nest.is_sequence_or_composite(outputs):
    174         outputs = [outputs]

~\Anaconda3\lib\site-packages\tensorflow_core\python\ops\rnn.py in _time_step(time, output_ta_t, state)
    882           skip_conditionals=True)
    883     else:
--> 884       (output, new_state) = call_cell()
    885 
    886     # Keras cells always wrap state as list, even if it's a single tensor.

~\Anaconda3\lib\site-packages\tensorflow_core\python\ops\rnn.py in <lambda>()
    868     if is_keras_rnn_cell and not nest.is_sequence(state):
    869       state = [state]
--> 870     call_cell = lambda: cell(input_t, state)
    871 
    872     if sequence_length is not None:

~\Anaconda3\lib\site-packages\tensorflow_core\python\ops\rnn_cell_impl.py in __call__(self, inputs, state, scope)
   1138     """
   1139     return self._call_wrapped_cell(
-> 1140         inputs, state, cell_call_fn=self.cell.__call__, scope=scope)
   1141 
   1142   def get_config(self):

~\Anaconda3\lib\site-packages\tensorflow_core\python\ops\rnn_cell_wrapper_impl.py in _call_wrapped_cell(self, inputs, state, cell_call_fn, **kwargs)
    275       inputs = self._dropout(inputs, "input", self._recurrent_input_noise,
    276                              self._input_keep_prob)
--> 277     output, new_state = cell_call_fn(inputs, state, **kwargs)
    278     if _should_dropout(self._state_keep_prob):
    279       # Identify which subsets of the state to perform dropout on and

~\Anaconda3\lib\site-packages\tensorflow_core\python\ops\rnn_cell_impl.py in __call__(self, inputs, state, scope)
    242         setattr(self, scope_attrname, scope)
    243       with scope:
--> 244         return super(RNNCell, self).__call__(inputs, state)
    245 
    246   def _rnn_get_variable(self, getter, *args, **kwargs):

~\Anaconda3\lib\site-packages\tensorflow_core\python\layers\base.py in __call__(self, inputs, *args, **kwargs)
    545 
    546       # Actually call layer
--> 547       outputs = super(Layer, self).__call__(inputs, *args, **kwargs)
    548 
    549     if not context.executing_eagerly():

~\Anaconda3\lib\site-packages\tensorflow_core\python\keras\engine\base_layer.py in __call__(self, inputs, *args, **kwargs)
    776                     outputs = base_layer_utils.mark_as_return(outputs, acd)
    777                 else:
--> 778                   outputs = call_fn(cast_inputs, *args, **kwargs)
    779 
    780             except errors.OperatorNotAllowedInGraphError as e:

~\Anaconda3\lib\site-packages\tensorflow_core\python\autograph\impl\api.py in wrapper(*args, **kwargs)
    235       except Exception as e:  # pylint:disable=broad-except
    236         if hasattr(e, 'ag_error_metadata'):
--> 237           raise e.ag_error_metadata.to_exception(e)
    238         else:
    239           raise

ValueError: in converted code:

    C:\Users\ThinkPad\Anaconda3\lib\site-packages\tensorflow_core\python\ops\rnn_cell_impl.py:1306 call
        (len(self.state_size), state))

    ValueError: Expected state to be a tuple of length 2, but received: Tensor("Placeholder_2:0", shape=(None, 256), dtype=float32)

我该如何克服这个问题，以便 state_is_tuple 不会显示任何错误，因为 TensorFlow 版本指示将其更改为 True？因为我尝试过 LSTMStateTuple 但它不起作用，可能是我的方法不正确，请帮忙。

【问题讨论】：

欢迎来到 *** 社区，这个问题已经被问到检查这些链接 prev question 1 prev question 2 并且请在 [提出你的问题] (***.com/help/how-to-ask) 之前阅读并阅读此内容！ /跨度>

首先亲爱的@techPirate99先生，我没有batch_size参数，其次，我问了这个问题，因为我已经通过保存初始状态的张量尝试了给定链接中的解决方案，但它不起作用，仍然出现错误，例如：ValueError: Expected state to be a tuple of length 2, but received: Tensor("Placeholder_2:0", shape=(None, 256), dtype=float32)

这就是代码的样子 @techPirate99 : class Model: def __init__( self, learning_rate, num_layers, size, size_layer, output_size, forget_bias = 0.1, ): def lstm_cell(size_layer): return tf .compat.v1.nn.rnn_cell.LSTMCell(size_layer, state_is_tuple = True) self.initial_state = np.zeros((num_layers, 2, 32, size_layer)) 我在 lstm_cell 函数之后添加 self.initial_state

您有什么建议可以通过编辑上面的代码使 state_is_tuple 正常工作，先生？ @techPirate99

我的建议是在 tf.compat.v1.nn.rnn_cell.LSTMCell(num_units, use_peepholes=False, cell_clip=None, initializer=None, num_proj=None, proj_clip=None, num_unit_shards =None, num_proj_shards=None, forget_bias=1.0, state_is_tuple=True, activation=None, reuse=None, name=None, dtype=None, **kwargs ) 并查看此文档以了解 tensorflow rnn 单元 link for doc

【参考方案1】：

tf.compat.v1.nn 已弃用。较新的 Keras API 具有等效的功能并得到积极维护。请参阅https://www.tensorflow.org/guide/keras/overview 了解更多信息。