如何创建张量列表并在for循环中使用tensorflow2中的tf.stack

Question

我正在尝试创建一个张量列表，并在tensorflow2中使用for循环将它们堆叠在一起。我创建了一个测试示例，并尝试如下所示。

import tensorflow as tf

@tf.function
def test(x):
    tensor_list = []
    for i in tf.range(x):
        tensor_list.append(tf.ones(4)*tf.cast(i, tf.float32))
    return  tf.stack(tensor_list)

result = test(5)
print(result)

但是我得到了如下的错误：

Traceback (most recent call last):
  File "test.py", line 10, in <module>
    result = test(5)
  File "/root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/eager/def_function.py", line 457, in __call__
    result = self._call(*args, **kwds)
  File "/root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/eager/def_function.py", line 503, in _call
    self._initialize(args, kwds, add_initializers_to=initializer_map)
  File "/root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/eager/def_function.py", line 408, in _initialize
    *args, **kwds))
  File "/root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/eager/function.py", line 1848, in _get_concrete_function_internal_garbage_collected
    graph_function, _, _ = self._maybe_define_function(args, kwargs)
  File "/root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/eager/function.py", line 2150, in _maybe_define_function
    graph_function = self._create_graph_function(args, kwargs)
  File "/root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/eager/function.py", line 2041, in _create_graph_function
    capture_by_value=self._capture_by_value),
  File "/root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/framework/func_graph.py", line 915, in func_graph_from_py_func
    func_outputs = python_func(*func_args, **func_kwargs)
  File "/root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/eager/def_function.py", line 358, in wrapped_fn
    return weak_wrapped_fn().__wrapped__(*args, **kwds)
  File "/root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/framework/func_graph.py", line 905, in wrapper
    raise e.ag_error_metadata.to_exception(e)
tensorflow.python.framework.errors_impl.InaccessibleTensorError: in converted code:

    test.py:8 test  *
        return  tf.stack(tensor_list)
    /root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/util/dispatch.py:180 wrapper
        return target(*args, **kwargs)
    /root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/ops/array_ops.py:1165 stack
        return gen_array_ops.pack(values, axis=axis, name=name)
    /root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/ops/gen_array_ops.py:6304 pack
        "Pack", values=values, axis=axis, name=name)
    /root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/framework/op_def_library.py:793 _apply_op_helper
        op_def=op_def)
    /root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/framework/func_graph.py:544 create_op
        inp = self.capture(inp)
    /root/.pyenv/versions/summarization-abstractive/lib/python3.6/site-packages/tensorflow_core/python/framework/func_graph.py:603 capture
        % (tensor, tensor.graph, self))

    InaccessibleTensorError: The tensor 'Tensor("mul:0", shape=(4,), dtype=float32)' cannot be accessed here: it is defined in another function or code block. Use return values, explicit Python locals or TensorFlow collections to access it. Defined in: FuncGraph(name=while_body_8, id=139870442952744); accessed from: FuncGraph(name=test, id=139870626510608).

有人知道我做错了什么吗？如何在tensorflow 2中创建张量列表并将它们与for循环堆叠在一起？

Answer 1

在张量上的循环通常应该通过使用´tf.map_fn´来完成。以下是一个有效的解决方案：

import tensorflow as tf
import numpy as np

@tf.function
def test(x):

    tensor_list = tf.map_fn(lambda inp: tf.ones(4)*tf.cast(inp, tf.float32), x, dtype=tf.dtypes.float32)

    return  tf.stack(tensor_list)

result = test(np.arange(5))
print(result)

但是，您必须在test()函数中提供一个实数组，但也可以在tf.function中调用tf.range()将标量转换为张量。

Answer 2

使用tf.TensorArray而不是list