0
我試圖使用張量流來編碼RDD編碼器和解碼器，並使用不同長度的序列輸入，所以希望編碼器和解碼器都是動態的。此外，解碼器輸入由編碼器最終隱藏狀態（上下文向量）調節，其與Related Paper類似於第3頁的圖片a。解碼器嘗試在訓練期間完全推理，饋送先前的輸出和上下文向量作爲輸入，每一步。如何使用ScheduledOutputTrainingHelper調節RNN動態解碼器輸入端的編碼器最終隱藏狀態？
import tensorflow as tf
import copy
import math
from tensorflow.python.layers.core import Dense
class RNNEncoder_Decoder(object):
def __init__(self,input_dim,
context_dim,output_dim,hidden_dim,
layers_stacked_count,learning_rate):
self.graph = tf.get_default_graph()
self.input_dim = input_dim
self.output_dim = output_dim
self.context_dim = context_dim
self.hidden_dim = hidden_dim
self.layers_stacked_count = layers_stacked_count
self.learning_rate = learning_rate
self.sampling_probability = tf.constant(dtype=tf.float32,value=1.0)
# [batch_size,sequence_length,input_dimension]
self.enc_inp = tf.placeholder(tf.float32, [None,None,self.input_dim], name='encoder_inputs')
self.expected_out = tf.placeholder(tf.float32, [None,None,self.input_dim], name='expected_outs')
# fullly inference during trianing
self.dec_inp = tf.zeros_like(self.expected_out,dtype=tf.float32,name='decoder_inputs')
seq_length = tf.reduce_sum(tf.sign(tf.reduce_max(tf.abs(self.enc_inp), 2)), 1)
self.seq_length = tf.cast(seq_length, tf.int32)
with tf.variable_scope('RNNEncoderDecoder'):
with tf.variable_scope("Enocder") as encoder_varscope:
# create encoder LSTM cell
encoder_cells = []
for i in range(self.layers_stacked_count):
with tf.variable_scope('EncoderCell_{}'.format(i)):
encoder_cells.append(tf.nn.rnn_cell.LSTMCell(self.hidden_dim,
use_peepholes=True))
self.encoder_cell = tf.nn.rnn_cell.MultiRNNCell(encoder_cells)
# ruuning dynamic rnn encoder
_, enc_state = tf.nn.dynamic_rnn(cell = self.encoder_cell,
initial_state=None,
dtype=tf.float32,
inputs = self.enc_inp,
sequence_length = self.seq_length
)
# extract top layer hidden state as feature representation
self.context_vector = enc_state[-1].h
cell_state0 = tf.zeros_like(enc_state[0].c,dtype=tf.float32)
hidden_state0 = tf.zeros_like(enc_state[0].h,dtype=tf.float32)
dec_init_state = (enc_state[1], # pass the top layer state of enocder to the bottom layer of decoder
tf.nn.rnn_cell.LSTMStateTuple(cell_state0, hidden_state0))
# condition extracted features on decoder inputs
# with a shape that matches decoder inputs in all but (potentially) the final dimension.
# tile context vector from [batch_size,context_dim] to [batch_size,decoder_sequence_length,context_dim]
context_vector_shape = tf.shape(self.context_vector)
context_vector_reshaped = tf.reshape(self.context_vector,
[context_vector_shape[0], 1, context_vector_shape[1]]
)
enc_inp_shape = tf.shape(self.enc_inp)
self.auxiliary_inputs = tf.tile(context_vector_reshaped,
multiples=[1,enc_inp_shape[1],1]
)
with tf.variable_scope("Deocder") as decoder_varscope:
# create decoder LSTM cell
decoder_cells = []
for i in range(self.layers_stacked_count):
with tf.variable_scope('DecoderCell_{}'.format(i)):
decoder_cells.append(tf.nn.rnn_cell.LSTMCell(self.hidden_dim,
use_peepholes=True))
self.decoder_cell = tf.nn.rnn_cell.MultiRNNCell(decoder_cells)
dec_out_dense = Dense(units = self.output_dim,
activation = None,
use_bias = False,
kernel_initializer = tf.truncated_normal_initializer(
dtype=tf.float32,
stddev = 1.0/math.sqrt(float(self.hidden_dim))
),
name = 'dec_outp_linear_projection'
)
training_helper = tf.contrib.seq2seq.ScheduledOutputTrainingHelper(
inputs = self.dec_inp,
sequence_length = self.seq_length,
auxiliary_inputs = self.auxiliary_inputs, # condtional on inputs
sampling_probability = 1.0, # for fullly inference
name = 'feeding_conditional_input'
)
decoder = tf.contrib.seq2seq.BasicDecoder(
cell = self.decoder_cell,
helper = training_helper,
initial_state = dec_init_state,
output_layer = dec_out_dense
)
outputs, _ , final_seq_lengths = tf.contrib.seq2seq.dynamic_decode(decoder=decoder,
impute_finished = True
)
self.outputs = outputs
### optimize loss part
def get_decoder_prediction(self,X,session):
feed_dict = {
self.enc_inp:X
}
feed_dict.update({self.expected_out:X})
run = [self.outputs]
return session.run(run,feed_dict=feed_dict)
context_dim = 32
output_dim = input_dim = 1
hidden_dim = 32
layers_stacked_count = 2
learning_rate = 0.01
test = RNNEncoder_Decoder(input_dim=input_dim,
context_dim=context_dim,
output_dim=output_dim,
hidden_dim=hidden_dim,
layers_stacked_count=layers_stacked_count,
learning_rate=learning_rate
)
沒有 「auxiliary_inputs = self.auxiliary_inputs」，它成功運行，
但隨着auxiliary_inputs = self.auxiliary_inputs我得到了以下錯誤：
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-3-02522a01f0d8> in <module>()
9 hidden_dim=hidden_dim,
10 layers_stacked_count=layers_stacked_count,
---> 11 learning_rate=learning_rate
12 )
<ipython-input-2-86494b8d99fa> in __init__(self, input_dim, context_dim, output_dim, hidden_dim, layers_stacked_count, learning_rate)
98
99 outputs, _ , final_seq_lengths = tf.contrib.seq2seq.dynamic_decode(decoder=decoder,
--> 100 impute_finished = True
101 )
102 self.outputs = outputs
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/contrib/seq2seq/python/ops/decoder.py in dynamic_decode(decoder, output_time_major, impute_finished, maximum_iterations, parallel_iterations, swap_memory, scope)
284 ],
285 parallel_iterations=parallel_iterations,
--> 286 swap_memory=swap_memory)
287
288 final_outputs_ta = res[1]
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/python/ops/control_flow_ops.py in while_loop(cond, body, loop_vars, shape_invariants, parallel_iterations, back_prop, swap_memory, name)
2773 context = WhileContext(parallel_iterations, back_prop, swap_memory, name)
2774 ops.add_to_collection(ops.GraphKeys.WHILE_CONTEXT, context)
-> 2775 result = context.BuildLoop(cond, body, loop_vars, shape_invariants)
2776 return result
2777
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/python/ops/control_flow_ops.py in BuildLoop(self, pred, body, loop_vars, shape_invariants)
2602 self.Enter()
2603 original_body_result, exit_vars = self._BuildLoop(
-> 2604 pred, body, original_loop_vars, loop_vars, shape_invariants)
2605 finally:
2606 self.Exit()
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/python/ops/control_flow_ops.py in _BuildLoop(self, pred, body, original_loop_vars, loop_vars, shape_invariants)
2552 structure=original_loop_vars,
2553 flat_sequence=vars_for_body_with_tensor_arrays)
-> 2554 body_result = body(*packed_vars_for_body)
2555 if not nest.is_sequence(body_result):
2556 body_result = [body_result]
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/contrib/seq2seq/python/ops/decoder.py in body(time, outputs_ta, state, inputs, finished, sequence_lengths)
232 """
233 (next_outputs, decoder_state, next_inputs,
--> 234 decoder_finished) = decoder.step(time, inputs, state)
235 next_finished = math_ops.logical_or(decoder_finished, finished)
236 if maximum_iterations is not None:
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/contrib/seq2seq/python/ops/basic_decoder.py in step(self, time, inputs, state, name)
137 """
138 with ops.name_scope(name, "BasicDecoderStep", (time, inputs, state)):
--> 139 cell_outputs, cell_state = self._cell(inputs, state)
140 if self._output_layer is not None:
141 cell_outputs = self._output_layer(cell_outputs)
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/python/ops/rnn_cell_impl.py in __call__(self, inputs, state, scope)
178 with vs.variable_scope(vs.get_variable_scope(),
179 custom_getter=self._rnn_get_variable):
--> 180 return super(RNNCell, self).__call__(inputs, state)
181
182 def _rnn_get_variable(self, getter, *args, **kwargs):
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/python/layers/base.py in __call__(self, inputs, *args, **kwargs)
448 # Check input assumptions set after layer building, e.g. input shape.
449 self._assert_input_compatibility(inputs)
--> 450 outputs = self.call(inputs, *args, **kwargs)
451
452 # Apply activity regularization.
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/python/ops/rnn_cell_impl.py in call(self, inputs, state)
936 [-1, cell.state_size])
937 cur_state_pos += cell.state_size
--> 938 cur_inp, new_state = cell(cur_inp, cur_state)
939 new_states.append(new_state)
940
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/python/ops/rnn_cell_impl.py in __call__(self, inputs, state, scope)
178 with vs.variable_scope(vs.get_variable_scope(),
179 custom_getter=self._rnn_get_variable):
--> 180 return super(RNNCell, self).__call__(inputs, state)
181
182 def _rnn_get_variable(self, getter, *args, **kwargs):
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/python/layers/base.py in __call__(self, inputs, *args, **kwargs)
448 # Check input assumptions set after layer building, e.g. input shape.
449 self._assert_input_compatibility(inputs)
--> 450 outputs = self.call(inputs, *args, **kwargs)
451
452 # Apply activity regularization.
/Library/Frameworks/Python.framework/Versions/3.5/lib/python3.5/site-packages/tensorflow/python/ops/rnn_cell_impl.py in call(self, inputs, state)
554 input_size = inputs.get_shape().with_rank(2)[1]
555 if input_size.value is None:
--> 556 raise ValueError("Could not infer input size from inputs.get_shape()[-1]")
557 scope = vs.get_variable_scope()
558 with vs.variable_scope(scope, initializer=self._initializer) as unit_scope:
ValueError: Could not infer input size from inputs.get_shape()[-1]
我剛開始使用tensforflow，所以任何人都可以幫助我： 這是一個正確的方式來調節解碼器輸入端上編碼器的最後隱藏狀態嗎？ 以及爲什麼在我將auxiliary_inputs作爲錯誤輸入之後解碼器的輸入變爲None？