我知道类似的问题已经在stackoverflow和互联网上多次被问过，但是我无法找到解决以下问题的方法：我正在尝试在tensorflow中构建一个有状态的LSTM模型并且它的 Estimator API 。
我尝试了Tensorflow, best way to save state in RNNs?的解决方案，只要我使用静态batch_size，它就会起作用。拥有动态batch_size会导致以下问题：
ValueError：initial_value必须具有指定的形状： 张量（＆＃34; DropoutWrapperZeroState / MultiRNNCellZeroState / DropoutWrapperZeroState / LSTMCellZeroState /零点：0＆＃34 ;, shape =（？，200），dtype = float32）
设置tf.Variable(...., validate_shape=False)只是将问题进一步向下移动图表：
Traceback (most recent call last):
File "model.py", line 576, in <module>
tf.app.run(main=run_experiment)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/platform/app.py", line 48, in run
_sys.exit(main(_sys.argv[:1] + flags_passthrough))
File "model.py", line 137, in run_experiment
hparams=params # HParams
File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/learn_runner.py", line 210, in run
return _execute_schedule(experiment, schedule)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/learn_runner.py", line 47, in _execute_schedule
return task()
File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/experiment.py", line 495, in train_and_evaluate
self.train(delay_secs=0)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/experiment.py", line 275, in train
hooks=self._train_monitors + extra_hooks)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/experiment.py", line 660, in _call_train
hooks=hooks)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/estimator/estimator.py", line 241, in train
loss = self._train_model(input_fn=input_fn, hooks=hooks)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/estimator/estimator.py", line 560, in _train_model
model_fn_lib.ModeKeys.TRAIN)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/estimator/estimator.py", line 545, in _call_model_fn
features=features, labels=labels, **kwargs)
File "model.py", line 218, in model_fn
output, state = get_model(features, params)
File "model.py", line 567, in get_model
model = lstm(inputs, params)
File "model.py", line 377, in lstm
output, new_states = tf.nn.dynamic_rnn(multicell, inputs=inputs, initial_state = states)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn.py", line 574, in dynamic_rnn
dtype=dtype)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn.py", line 737, in _dynamic_rnn_loop
swap_memory=swap_memory)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 2770, in while_loop
result = context.BuildLoop(cond, body, loop_vars, shape_invariants)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 2599, in BuildLoop
pred, body, original_loop_vars, loop_vars, shape_invariants)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 2549, in _BuildLoop
body_result = body(*packed_vars_for_body)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn.py", line 722, in _time_step
(output, new_state) = call_cell()
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn.py", line 708, in <lambda>
call_cell = lambda: cell(input_t, state)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn_cell_impl.py", line 752, in __call__
output, new_state = self._cell(inputs, state, scope)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn_cell_impl.py", line 180, in __call__
return super(RNNCell, self).__call__(inputs, state)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/layers/base.py", line 441, in __call__
outputs = self.call(inputs, *args, **kwargs)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn_cell_impl.py", line 916, in call
cur_inp, new_state = cell(cur_inp, cur_state)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn_cell_impl.py", line 752, in __call__
output, new_state = self._cell(inputs, state, scope)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn_cell_impl.py", line 180, in __call__
return super(RNNCell, self).__call__(inputs, state)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/layers/base.py", line 441, in __call__
outputs = self.call(inputs, *args, **kwargs)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn_cell_impl.py", line 542, in call
lstm_matrix = _linear([inputs, m_prev], 4 * self._num_units, bias=True)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn_cell_impl.py", line 1002, in _linear
raise ValueError("linear is expecting 2D arguments: %s" % shapes)
ValueError: linear is expecting 2D arguments: [TensorShape([Dimension(None), Dimension(62)]), TensorShape(None)]
根据github issue 2838，不推荐使用不可训练的变量（???），这就是我继续寻找其他解决方案的原因。
现在我在model_fn中使用占位符和类似的东西（在github线程中也有建议）：
def rnn_placeholders(state):
"""Convert RNN state tensors to placeholders with the zero state as default."""
if isinstance(state, tf.contrib.rnn.LSTMStateTuple):
c, h = state
c = tf.placeholder_with_default(c, c.shape, c.op.name)
h = tf.placeholder_with_default(h, h.shape, h.op.name)
return tf.contrib.rnn.LSTMStateTuple(c, h)
elif isinstance(state, tf.Tensor):
h = state
h = tf.placeholder_with_default(h, h.shape, h.op.name)
return h
else:
structure = [rnn_placeholders(x) for x in state]
return tuple(structure)
state = rnn_placeholders(cell.zero_state(batch_size, tf.float32))
for tensor in flatten(state):
tf.add_to_collection('rnn_state_input', tensor)
x, new_state = tf.nn.dynamic_rnn(...)
for tensor in flatten(new_state):
tf.add_to_collection('rnn_state_output', tensor)
但遗憾的是，我不知道如何使用占位符new_state在每次迭代时将其值反馈给占位符state，在使用tf.Estimator API等时。由于我对Tensorflow很新，我认为我在这里缺乏概念知识。可能使用自定义SessionRunHook？：
class UpdateHook(tf.train.SessionRunHook):
def before_run(self, run_context):
run_args = super(UpdateHook, self).before_run(run_context)
run_args = tf.train.SessionRunArgs(new_state)
#print(run_args)
return run_args
def after_run(self, run_context, run_values):
#run_values gives the actual value of new_state.
# How to update now the state placeholder??
有没有人知道如何解决这个问题？提示和技巧非常感谢!!! 非常感谢！
PS：如果有什么不清楚，请告诉我;）编辑：不幸的是我正在使用新的tf.data API，并且不能像Eugene建议的那样使用StateSavingRNNEstimator。
答案 0 :(得分：1)
您可以使用batch_sequences_with_states建立代码的估算工具。它被称为StateSavingRNNEstimator。除非您使用新的tf.contrib.data / tf.data API，否则它应该足以让您入门。
答案 1 :(得分：1)
这个答案可能会迟到。 几个月前我遇到了类似的问题。 我使用自定义的SessionRunHook解决了它。它在性能方面可能并不完美，但你可以尝试一下。
class LSTMStateHook(tf.train.SessionRunHook):
def __init__(self, params):
self.init_states = None
self.current_state = np.zeros((params.rnn_layers, 2, params.batch_size, params.state_size))
def before_run(self, run_context):
run_args = tf.train.SessionRunArgs([tf.get_default_graph().get_tensor_by_name('LSTM/output_states:0')],{self.init_states:self.current_state,},)
return run_args
def after_run(self, run_context, run_values):
self.current_state = run_values[0][0] //depends on your session run arguments!!!!!!!
def begin(self):
self.init_states = tf.get_default_graph().get_tensor_by_name('LSTM/init_states:0')
在您定义lstm图的代码中，您需要这样的内容：
if self.stateful is True:
init_states = multicell.zero_state(self.batch_size, tf.float32)
init_states = tf.identity(init_states, "init_states")
l = tf.unstack(init_states, axis=0)
rnn_tuple_state = tuple([tf.nn.rnn_cell.LSTMStateTuple(l[idx][0], l[idx][1]) for idx in range(self.rnn_layers)])
else:
rnn_tuple_state = multicell.zero_state(self.batch_size, tf.float32)
# Unroll RNN
output, output_states = tf.nn.dynamic_rnn(multicell, inputs=inputs, initial_state = rnn_tuple_state)
if self.stateful is True:
output_states = tf.identity(output_states, "output_states")
return output