此代码用作Seq2Seq模型...
# imports
from utils import get_sorted_buckets
import logging
from six.moves import xrange
import numpy as np
import tensorflow as tf
# classes
class Seq2Seq:
def __init__(self,
input_vocab_size,
output_vocab_size,
buckets,
layer_size=256,
n_layers=3,
max_gradient_norm=5.0,
batch_size=64,
learning_rate=0.5,
learning_rate_decay_factor=0.99,
rnn_cell=tf.contrib.rnn.GRUCell,
n_samples=512,
forward_only=False):
logging.info('initializing Seq2Seq model')
buckets = get_sorted_buckets(buckets)
self.input_vocab_size = input_vocab_size
self.output_vocab_size = output_vocab_size
self.buckets = buckets
self.layer_size = layer_size
self.n_layers = n_layers
self.max_gradient_norm = max_gradient_norm
self.batch_size = batch_size
self.learning_rate = learning_rate
self.learning_rate_decay_factor = learning_rate_decay_factor
self.rnn_cell = rnn_cell
self.n_samples = n_samples
self.forward_only = forward_only
self.learning_rate = tf.Variable(float(self.learning_rate),
trainable=False)
self.learning_rate_decay_operation = self.learning_rate.assign(
self.learning_rate * self.learning_rate_decay_factor)
self.global_step = tf.Variable(0, trainable=False)
self.encoder_inputs = [self._get_val_placeholder('encoder', i)
for i in xrange(buckets[-1][0])]
self.decoder_inputs = [self._get_val_placeholder('decoder', i)
for i in xrange(buckets[-1][1] + 1)]
self.target_weights = [self._get_val_placeholder('weight', i,
dtype=tf.float32)
for i in xrange(buckets[-1][1] + 1)]
logging.debug('getting model')
self.outputs, self.losses = self._get_model_with_buckets()
self.saver = tf.train.Saver(tf.all_variables())
if not forward_only:
logging.debug('setting gradient norms and updates')
out = self._get_gradient_norms_and_updates()
self.gradient_norms, self.updates = out
def _get_softmax_loss_func_and_output_proj(self):
logging.debug('function: _get_softmax_loss_func_and_output_proj')
use_sampled_softmax = self.n_samples > 0 and \
self.n_samples < self.output_vocab_size
if use_sampled_softmax:
w = tf.get_variable('proj_w',
[self.layer_size, self.output_vocab_size],
dtype=tf.int32)
w_t = tf.transpose(w)
b = tf.get_variable('proj_b',
[self.output_vocab_size],
dtype=tf.int32)
def get_sampled_loss(inputs, labels):
labels = tf.reshape(labels, [-1, 1])
print(w_t)
print(b)
print(labels)
print(inputs)
return tf.nn.sampled_softmax_loss(
weights=w_t, biases=b, labels=labels, inputs=inputs,
num_sampled=self.n_samples,
num_classes=self.output_vocab_size)
softmax_loss_function = get_sampled_loss
output_projection = (w, b)
else:
softmax_loss_function = None
output_projection = None
return softmax_loss_function, output_projection
def _get_cell(self):
logging.debug('function: _get_cell')
single_cell = self.rnn_cell(self.layer_size)
# single_cell = self.rnn_cell
if self.n_layers > 1:
cell = tf.contrib.rnn.MultiRNNCell([single_cell] * self.n_layers)
else:
cell = single_cell
return cell
def _get_val_placeholder(self, name, idx, dtype=tf.int32):
return tf.placeholder(dtype, shape=[None], name='{}_{}'.format(name,
idx))
def _get_model_with_buckets(self):
targets = [self.decoder_inputs[i + 1]
for i in xrange(len(self.decoder_inputs) - 1)]
out = self._get_softmax_loss_func_and_output_proj()
softmax_loss_function, output_projection = out
cell = self._get_cell()
def seq2seq_func(encoder_inputs, decoder_inputs, do_decode):
print('seq2seq: {} {} {}'.format(encoder_inputs, decoder_inputs, do_decode))
return tf.contrib.legacy_seq2seq.embedding_attention_seq2seq(
encoder_inputs, decoder_inputs, cell,
num_encoder_symbols=self.input_vocab_size,
num_decoder_symbols=self.output_vocab_size,
embedding_size=self.layer_size,
output_projection=output_projection,
feed_previous=do_decode)
logging.debug('getting model with buckets')
outputs, losses = tf.contrib.legacy_seq2seq.model_with_buckets(
self.encoder_inputs, self.decoder_inputs, targets,
self.target_weights, self.buckets,
lambda x, y: seq2seq_func(x, y, self.forward_only),
softmax_loss_function=softmax_loss_function)
logging.debug('forward only')
if self.forward_only:
if output_projection is not None:
for i in xrange(len(self.buckets)):
w = output_projection[0]
b = output_projection[1]
new_bucket_outputs = [tf.matmul(output, w) + b
for output in outputs[i]]
self.outputs[i] = new_bucket_outputs
return outputs, losses
def _get_gradient_norms_and_updates(self):
params = tf.trainable_variables()
gradient_norms = []
updates = []
optimizer = tf.train.GradientDescentOptimizer(self.learning_rate)
for i in xrange(len(self.buckets)):
gradients = tf.gradients(self.losses[i], params)
clipped_gradients, norm = tf.clip_by_global_norm(
gradients, self.max_gradient_norm)
gradient_norms.append(norm)
updates.append(optimizer.apply_gradients(
zip(clipped_gradients, params), global_step=self.global_step))
return gradient_norms, updates
def step(self, session, encoder_inputs, decoder_inputs, target_weights,
bucket_id, forward_only):
encoder_size, decoder_size = self.buckets[bucket_id]
if len(encoder_inputs) != encoder_size:
raise ValueError('Encoder length must be equal to one in bucket.')
elif len(decoder_inputs) != decoder_size:
raise ValueError('Decoder length must be equal to one in bucket.')
elif len(target_weights) != decoder_size:
raise ValueError('Weights length must be equal to one in bucket.')
input_feed = {}
for i in xrange(encoder_size):
input_feed[self.encoder_inputs[i].name] = encoder_inputs[i]
for i in xrange(decoder_size):
input_feed[self.decoder_inputs[i].name] = decoder_inputs[i]
input_feed[self.target_weights[i].name] = target_weights[i]
last_target = self.decoder_inputs[decoder_size].name
input_feed[last_target] = np.zeros([self.batch_size], dtype=np.int32)
if forward_only:
output_feed = [self.losses[bucket_id]]
for i in xrange(decoder_size):
output_feed.append(self.outputs[bucket_id][i])
else:
output_feed = [self.updates[bucket_id],
self.gradient_norms[bucket_id],
self.losses[bucket_id]]
outputs = session.run(output_feed, input_feed)
if not forward_only:
return outputs[1], outputs[2], None
else:
return None, outputs[0], outputs[1:]
生成此错误...
INFO:root:initializing Seq2Seq model
DEBUG:root:getting model
DEBUG:root:function: _get_softmax_loss_func_and_output_proj
DEBUG:root:function: _get_cell
DEBUG:root:getting model with buckets
seq2seq: [<tf.Tensor 'encoder_0:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_1:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_2:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_3:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_4:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_5:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_6:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_7:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_8:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_9:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_10:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_11:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_12:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_13:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_14:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_15:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_16:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_17:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_18:0' shape=(?,) dtype=int32>, <tf.Tensor 'encoder_19:0' shape=(?,) dtype=int32>] [<tf.Tensor 'decoder_0:0' shape=(?,) dtype=int32>, <tf.Tensor 'decoder_1:0' shape=(?,) dtype=int32>, <tf.Tensor 'decoder_2:0' shape=(?,) dtype=int32>, <tf.Tensor 'decoder_3:0' shape=(?,) dtype=int32>, <tf.Tensor 'decoder_4:0' shape=(?,) dtype=int32>, <tf.Tensor 'decoder_5:0' shape=(?,) dtype=int32>, <tf.Tensor 'decoder_6:0' shape=(?,) dtype=int32>, <tf.Tensor 'decoder_7:0' shape=(?,) dtype=int32>, <tf.Tensor 'decoder_8:0' shape=(?,) dtype=int32>, <tf.Tensor 'decoder_9:0' shape=(?,) dtype=int32>] False
Tensor("transpose:0", shape=(20000, 256), dtype=int32)
Tensor("proj_b/read:0", shape=(20000,), dtype=int32)
Tensor("model_with_buckets/sequence_loss/sequence_loss_by_example/Reshape:0", shape=(?, 1), dtype=float32)
Tensor("decoder_1:0", shape=(?,), dtype=int32)
Traceback (most recent call last):
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/common_shapes.py", line 670, in _call_cpp_shape_fn_impl
status)
File "/usr/local/Cellar/python3/3.6.0_1/Frameworks/Python.framework/Versions/3.6/lib/python3.6/contextlib.py", line 89, in __exit__
next(self.gen)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/errors_impl.py", line 469, in raise_exception_on_not_ok_status
pywrap_tensorflow.TF_GetCode(status))
tensorflow.python.framework.errors_impl.InvalidArgumentError: Shape must be rank 2 but is rank 1 for 'model_with_buckets/sequence_loss/sequence_loss_by_example/sampled_softmax_loss/MatMul_1' (op: 'MatMul') with input shapes: [?], [?,256].
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "train.py", line 87, in <module>
main()
File "train.py", line 82, in main
model = get_model()
File "train.py", line 76, in get_model
get_rnn_cell(), FLAGS.n_samples, FLAGS.forward_only)
File "/Users/edholm/Desktop/Seq2Seq/model.py", line 58, in __init__
self.outputs, self.losses = self._get_model_with_buckets()
File "/Users/edholm/Desktop/Seq2Seq/model.py", line 141, in _get_model_with_buckets
softmax_loss_function=softmax_loss_function)
File "/usr/local/lib/python3.6/site-packages/tensorflow/contrib/legacy_seq2seq/python/ops/seq2seq.py", line 1195, in model_with_buckets
softmax_loss_function=softmax_loss_function))
File "/usr/local/lib/python3.6/site-packages/tensorflow/contrib/legacy_seq2seq/python/ops/seq2seq.py", line 1110, in sequence_loss
softmax_loss_function=softmax_loss_function))
File "/usr/local/lib/python3.6/site-packages/tensorflow/contrib/legacy_seq2seq/python/ops/seq2seq.py", line 1067, in sequence_loss_by_example
crossent = softmax_loss_function(target, logit)
File "/Users/edholm/Desktop/Seq2Seq/model.py", line 91, in get_sampled_loss
num_classes=self.output_vocab_size)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/nn_impl.py", line 1191, in sampled_softmax_loss
name=name)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/nn_impl.py", line 995, in _compute_sampled_logits
inputs, sampled_w, transpose_b=True) + sampled_b
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/math_ops.py", line 1855, in matmul
a, b, transpose_a=transpose_a, transpose_b=transpose_b, name=name)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/gen_math_ops.py", line 1454, in _mat_mul
transpose_b=transpose_b, name=name)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py", line 763, in apply_op
op_def=op_def)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 2397, in create_op
set_shapes_for_outputs(ret)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1757, in set_shapes_for_outputs
shapes = shape_func(op)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1707, in call_with_requiring
return call_cpp_shape_fn(op, require_shape_fn=True)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/common_shapes.py", line 610, in call_cpp_shape_fn
debug_python_shape_fn, require_shape_fn)
File "/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/common_shapes.py", line 675, in _call_cpp_shape_fn_impl
raise ValueError(err.message)
ValueError: Shape must be rank 2 but is rank 1 for 'model_with_buckets/sequence_loss/sequence_loss_by_example/sampled_softmax_loss/MatMul_1' (op: 'MatMul') with input shapes: [?], [?,256].
该错误可能与采样的softmax损失有关，我之前有很多错误，这与Tensorflow的新更新，参数的新语法顺序等有关。代码中可能存在更多错误，但在遇到更多错误之前我必须解决这个问题。
两天过去了，我还不知道该怎么做。我应该对代码进行哪些更改才能使其正常工作？
答案 0 :(得分：1)
在新的Tensorflow更新中，他们颠倒了sampled_softmax_loss中的参数顺序。
不要使用它。
def get_sampled_loss(inputs, labels):
labels = tf.reshape(labels, [-1, 1])
return tf.nn.sampled_softmax_loss(
weights=w_t, biases=b, labels=labels, inputs=inputs,
num_sampled=self.n_samples,
num_classes=self.output_vocab_size)
而是使用它。
def get_sampled_loss(labels, inputs):
labels = tf.reshape(labels, [-1, 1])
return tf.nn.sampled_softmax_loss(
weights=w_t, biases=b, labels=labels, inputs=inputs,
num_sampled=self.n_samples,
num_classes=self.output_vocab_size)
两个片段之间的区别在于不是按此顺序包含参数：get_sampled_loss(inputs, labels)。
使用此订单：get_sampled_loss(labels, inputs)