因此，经过一番调查，我找到了与该线程相关的所有问题的答案。self.ids_mask主要问题是：为什么InferenceSampler不更新？原因在于内部dynamic_decode。根据Tensorflow 的问题跟踪器中的这个答案：
...每次循环迭代都会评估循环内定义的张量。所有在循环外定义的张量都将只计算一次。
就我而言，self.ids_mask是在循环之外指定的。这意味着我需要重写dynamic_decode以获得我想要的东西。下面的代码是初始任务的一些修改版本，但几乎相同。
让我们从一个新的开始，它应该创建和更新已经预测dynamic_decode的掩码收集。sample_ids我删除了我没有修改的代码，遵循initial_maskandmask变量。
新dynamic_decode：
def dynamic_decode(decoder,
output_time_major=False,
impute_finished=False,
maximum_iterations=None,
parallel_iterations=32,
swap_memory=False,
scope=None):
...
initial_finished, initial_inputs, initial_mask, initial_state = decoder.initialize()
...
def body(time, outputs_ta, state, inputs, finished, sequence_lengths, mask):
"""Internal while_loop body.
Args:
time: scalar int32 tensor.
outputs_ta: structure of TensorArray.
state: (structure of) state tensors and TensorArrays.
inputs: (structure of) input tensors.
finished: bool tensor (keeping track of what's finished).
sequence_lengths: int32 tensor (keeping track of time of finish).
mask: SparseTensor to remove already predicted items
Returns:
`(time + 1, outputs_ta, next_state, next_inputs, next_finished,
next_sequence_lengths, next_mask)`.
```
"""
(next_outputs, decoder_state, next_inputs, next_mask,
decoder_finished) = decoder.step(time, inputs, state, mask)
...
nest.assert_same_structure(state, decoder_state)
nest.assert_same_structure(outputs_ta, next_outputs)
nest.assert_same_structure(inputs, next_inputs)
nest.assert_same_structure(mask, next_mask)
...
return (time + 1, outputs_ta, next_state, next_inputs, next_finished,
next_sequence_lengths, next_mask)
res = control_flow_ops.while_loop(
condition,
body,
loop_vars=[
initial_time, initial_outputs_ta, initial_state, initial_inputs,
initial_finished, initial_sequence_lengths, initial_mask,
],
parallel_iterations=parallel_iterations,
swap_memory=swap_memory)
...
return final_outputs, final_state, final_sequence_lengths
在下一步mask应该传递给Decoderand Helper。以下是 和 的更新BasicDecoder版本InferenceHelper：
MaskedDecoder：
class MaskedDecoder(BasicDecoder):
def step(self, time, inputs, state, mask, name=None):
with ops.name_scope(name, "MaskedDecoderStep", (time, inputs, state, mask)):
cell_outputs, cell_state = self._cell(inputs, state)
if self._output_layer is not None:
cell_outputs = self._output_layer(cell_outputs)
sample_ids = self._helper.sample(
time=time,
outputs=cell_outputs,
state=cell_state,
mask=mask)
(finished, next_inputs, next_state, next_mask) = self._helper.next_inputs(
time=time,
outputs=cell_outputs,
state=cell_state,
mask=mask,
sample_ids=sample_ids)
outputs = BasicDecoderOutput(cell_outputs, sample_ids)
return (outputs, next_state, next_inputs, next_mask, finished)
MaskedInferenceHelper：
class MaskedInferenceHelper(Helper):
"""A helper to use during inference with a custom sampling function."""
def __init__(self, norm_track_embs, features, start_sample_ids):
self._norm_track_embs = norm_track_embs
self._batch_size = tf.shape(start_sample_ids)[0]
self._n_tracks = tf.shape(norm_track_embs)[0]
self._start_sample_ids = start_sample_ids
self._sample_shape = tf.TensorShape([])
self._sample_dtype = tf.int32
self._features = features
def _get_sparse_mask(self, sample_ids):
_mask_shape = tf.convert_to_tensor([
tf.cast(self._batch_size, dtype=tf.int64),
tf.cast(self._n_tracks, dtype=tf.int64)
])
_st_rows = tf.range(0, self._batch_size)
_st_cols = sample_ids
_st_indices = tf.cast(tf.stack([_st_rows, _st_cols], axis=1), dtype=tf.int64)
_st_values = tf.fill([self._batch_size], np.inf)
return tf.SparseTensor(_st_indices, _st_values, _mask_shape)
...
def initialize(self, name=None):
finished = tf.tile([False], [self._batch_size])
start_embs = tf.nn.embedding_lookup(self._norm_track_embs, self._start_sample_ids)
start_inputs = tf.concat([start_embs, self._features], axis=1)
mask = self._get_sparse_mask(self._start_sample_ids)
return finished, start_inputs, mask
def sample(self, time, outputs, state, mask, name=None):
del time, state # unused by sample
outputs = tf.nn.l2_normalize(outputs, axis=-1)
cos_sims = tf.matmul(outputs, self._norm_track_embs, transpose_b=True)
cos_sims = cos_sims - tf.sparse_tensor_to_dense(mask)
sample_ids = tf.cast(tf.argmax(cos_sims, axis=-1), tf.int32)
return sample_ids
def next_inputs(self, time, outputs, state, sample_ids, mask, name=None):
del time, outputs # unused by next_inputs
finished = tf.tile([False], [self._batch_size])
next_embs = tf.nn.embedding_lookup(self._norm_track_embs, sample_ids)
next_inputs = tf.concat([next_embs, self._features], axis=1)
next_mask = tf.sparse_add(mask, self._get_sparse_mask(sample_ids))
return finished, next_inputs, state, next_mask
所以，现在我可以在不重复已经预测的项目的情况下生成推论。