Question

我想在特定的gpu设备下使用dynamic_rnn，但是会引发如下错误：

import tensorflow as tf

sequence = [[1, 2],[2, 3, 5]]
def generator():
    for el in sequence:
        yield el

dataset = tf.data.Dataset().from_generator(generator, tf.int32, tf.TensorShape([None]))
dataset = dataset.map(lambda seq: (seq, tf.size(seq)))
dataset = dataset.padded_batch(1, padded_shapes=(tf.TensorShape([None]), tf.TensorShape([])), padding_values=(0, 0))
iter = dataset.make_initializable_iterator()
seq, seq_len = iter.get_next()

embedding = tf.get_variable('embeddings', [10, 100])
cells = tf.contrib.rnn.GRUCell(5)
emb = tf.nn.embedding_lookup(embedding, seq)

with tf.device('/gpu:0'):  # If I assign the gpu device, the codes below will raise errors
    outputs, state = tf.nn.dynamic_rnn(cell=cells, inputs=emb, sequence_length=seq_len, dtype=tf.float32)

with tf.Session() as sess:
    sess.run(iter.initializer)

    sess.run(tf.global_variables_initializer())
    out = sess.run(outputs)

错误：

Caused by op 'rnn/Assert/Assert', defined at:
  File "test.py", line 22, in <module>
    outputs, state = tf.nn.dynamic_rnn(cell=cells, inputs=emb, sequence_length=seq_len, dtype=tf.float32)
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/ops/rnn.py", line 622, in dynamic_rnn
    [_assert_has_shape(sequence_length, [batch_size])]):
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/ops/rnn.py", line 617, in _assert_has_shape
    packed_shape, " but saw shape: ", x_shape])
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/util/tf_should_use.py", line 118, in wrapped
    return _add_should_use_warning(fn(*args, **kwargs))
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 172, in Assert
    return gen_logging_ops._assert(condition, data, summarize, name="Assert")
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/ops/gen_logging_ops.py", line 51, in _assert
    name=name)
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/op_def_library.py", line 787, in _apply_op_helper
    op_def=op_def)
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/ops.py", line 3290, in create_op
    op_def=op_def)
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/ops.py", line 1654, in __init__
    self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access

InvalidArgumentError (see above for traceback): Cannot assign a device for operation 'rnn/Assert/Assert': Could not satisfy explicit device specification '/device:GPU:0' because no supported kernel for GPU devices is available.
Registered kernels:
  device='CPU'

     [[Node: rnn/Assert/Assert = Assert[T=[DT_STRING, DT_INT32, DT_STRING, DT_INT32], summarize=3, _device="/device:GPU:0"](rnn/All, rnn/Assert/Assert/data_0, rnn/stack, rnn/Assert/Assert/data_2, rnn/Shape_1)]]

如果我不使用tf.device（'/ gpu：0'），则代码可以正常工作。但是出于某种原因，我需要使用它，例如多GPU支持。我应该如何解决？谢谢

Answer 1

一个简单的解决方案是在会话中使用“软放置”选项，即要求TensorFlow使用所指示的设备，除非没有内核，否则它将退回到CPU：

# ...
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
    # ...

另一种可能性是将函数传递给tf.device，以便选择GPU设备，除非给定的操作满足某种条件（通常在其具有某种特定类型时，在这种情况下为Assert，尽管您可能需要在模型中排除其他操作）

def assign_device(op):
    if op.type != 'Assert':
        return '/gpu:0'
    else:
        return '/cpu:0'

# ...
with tf.device(assign_device):
    # ...

Tensorflow：特定gpu设备下的dynamic_rnn无法工作，GPU设备不支持受支持的内核

1 个答案: