def ctc_label_dense_to_sparse(labels, label_lengths):
"""Converts CTC labels from dense to sparse.
Arguments:
labels: dense CTC labels.
label_lengths: length of the labels.
Returns:
A sparse tensor representation of the lablels.
"""
label_shape = array_ops.shape(labels)
num_batches_tns = array_ops.stack([label_shape[0]])
max_num_labels_tns = array_ops.stack([label_shape[1]])
def range_less_than(_, current_input):
return array_ops.expand_dims(
math_ops.range(label_shape[1]), 0) < array_ops.fill(
max_num_labels_tns, current_input)
init = math_ops.cast(
array_ops.fill([1, label_shape[1]], 0), dtypes_module.bool)
dense_mask = functional_ops.scan(
range_less_than, label_lengths, initializer=init, parallel_iterations=1)
dense_mask = dense_mask[:, 0, :]
label_array = array_ops.reshape(
array_ops.tile(math_ops.range(0, label_shape[1]), num_batches_tns),
label_shape)
label_ind = array_ops.boolean_mask(label_array, dense_mask)
batch_array = array_ops.transpose(
array_ops.reshape(
array_ops.tile(math_ops.range(0, label_shape[0]), max_num_labels_tns),
reverse(label_shape, 0)))
batch_ind = array_ops.boolean_mask(batch_array, dense_mask)
indices = array_ops.transpose(
array_ops.reshape(concatenate([batch_ind, label_ind], axis=0), [2, -1]))
vals_sparse = array_ops.gather_nd(labels, indices)
return sparse_tensor.SparseTensor(
math_ops.to_int64(indices), vals_sparse, math_ops.to_int64(label_shape))
python类to_int64()的实例源码
def __compute_AP(self,c_scores,c_tp,c_fp,c_num_gbboxes):
aps_voc07 = {}
aps_voc12 = {}
for c in c_scores.keys():
num_gbboxes = c_num_gbboxes[c]
tp = c_tp[c]
fp = c_fp[c]
scores = c_scores[c]
#reshape data
num_gbboxes = math_ops.to_int64(num_gbboxes)
scores = math_ops.to_float(scores)
stype = tf.bool
tp = tf.cast(tp, stype)
fp = tf.cast(fp, stype)
# Reshape TP and FP tensors and clean away 0 class values.(difficult bboxes)
scores = tf.reshape(scores, [-1])
tp = tf.reshape(tp, [-1])
fp = tf.reshape(fp, [-1])
# Remove TP and FP both false.
mask = tf.logical_or(tp, fp)
rm_threshold = 1e-4
mask = tf.logical_and(mask, tf.greater(scores, rm_threshold))
scores = tf.boolean_mask(scores, mask)
tp = tf.boolean_mask(tp, mask)
fp = tf.boolean_mask(fp, mask)
num_gbboxes = tf.reduce_sum(num_gbboxes)
num_detections = tf.size(scores, out_type=tf.int32)
# Precison and recall values.
prec, rec = tfe.precision_recall(num_gbboxes, num_detections, tp, fp, scores)
v = tfe.average_precision_voc07(prec, rec)
aps_voc07[c] = v
# Average precision VOC12.
v = tfe.average_precision_voc12(prec, rec)
aps_voc12[c] = v
return aps_voc07, aps_voc12
def testMultiLabelTopKWithCustomMetrics(self):
"""Tests the cases where n_labels>1 top_k>1 and custom metrics on top_k."""
def _input_fn():
features = {
'language': tf.SparseTensor(values=['en', 'fr', 'zh'],
indices=[[0, 0], [0, 1], [2, 0]],
shape=[3, 2])
}
target = tf.constant([[0, 1], [0, 1], [0, 1]], dtype=tf.int64)
return features, target
def _my_metric_op(predictions, targets):
"""Simply adds the predictions and targets."""
return tf.add(math_ops.to_int64(predictions), targets)
sparse_column = tf.contrib.layers.sparse_column_with_hash_bucket(
'language', hash_bucket_size=20)
embedding_features = [
tf.contrib.layers.embedding_column(sparse_column, dimension=1)
]
classifier = dnn_sampled_softmax_classifier._DNNSampledSoftmaxClassifier(
n_classes=3,
n_samples=2,
n_labels=2,
top_k=2,
feature_columns=embedding_features,
hidden_units=[4, 4],
optimizer=tf.train.AdamOptimizer(learning_rate=0.01),
config=tf.contrib.learn.RunConfig(tf_random_seed=5))
classifier.fit(input_fn=_input_fn, steps=50)
# evaluate() without custom metrics.
evaluate_output = classifier.evaluate(input_fn=_input_fn, steps=1)
self.assertGreater(evaluate_output['precision_at_1'], 0.4)
self.assertGreater(evaluate_output['recall_at_1'], 0.4)
self.assertGreater(evaluate_output['precision_at_2'], 0.4)
self.assertGreater(evaluate_output['recall_at_2'], 0.4)
# evaluate() with custom metrics.
metrics = {('my_metric', 'top_k'): _my_metric_op}
evaluate_output = classifier.evaluate(input_fn=_input_fn, steps=1,
metrics=metrics)
# This test's output is flaky so just testing that 'my_metric' is indeed
# part of the evaluate_output.
self.assertTrue('my_metric' in evaluate_output)
# predict() with top_k.
predict_output = classifier.predict(input_fn=_input_fn, get_top_k=True)
self.assertListEqual([3, 2], list(predict_output.shape))
# TODO(dnivara): Setup this test such that it is not flaky and predict() and
# evaluate() outputs can be tested.
def sparse_feature_cross(inputs, hashed_output=False, num_buckets=0,
name=None):
"""Crosses a list of Tensor or SparseTensor objects.
See sparse_feature_cross_kernel.cc for more details.
Args:
inputs: List of `SparseTensor` or `Tensor` to be crossed.
hashed_output: If true, returns the hash of the cross instead of the string.
This will allow us avoiding string manipulations.
num_buckets: It is used if hashed_output is true.
output = hashed_value%num_buckets if num_buckets > 0 else hashed_value.
name: A name prefix for the returned tensors (optional).
Returns:
A `SparseTensor` with the crossed features.
Return type is string if hashed_output=False, int64 otherwise.
Raises:
TypeError: If the inputs aren't either SparseTensor or Tensor.
"""
if not isinstance(inputs, list):
raise TypeError("Inputs must be a list")
if not all(isinstance(i, ops.SparseTensor) or
isinstance(i, ops.Tensor) for i in inputs):
raise TypeError("All inputs must be SparseTensors")
sparse_inputs = [i for i in inputs if isinstance(i, ops.SparseTensor)]
dense_inputs = [i for i in inputs if not isinstance(i, ops.SparseTensor)]
indices = [sp_input.indices for sp_input in sparse_inputs]
values = [sp_input.values for sp_input in sparse_inputs]
shapes = [sp_input.shape for sp_input in sparse_inputs]
out_type = dtypes.int64 if hashed_output else dtypes.string
internal_type = dtypes.string
for i in range(len(values)):
if values[i].dtype != dtypes.string:
values[i] = math_ops.to_int64(values[i])
internal_type = dtypes.int64
for i in range(len(dense_inputs)):
if dense_inputs[i].dtype != dtypes.string:
dense_inputs[i] = math_ops.to_int64(dense_inputs[i])
internal_type = dtypes.int64
indices_out, values_out, shape_out = (
_sparse_feature_cross_op.sparse_feature_cross(indices,
values,
shapes,
dense_inputs,
hashed_output,
num_buckets,
out_type=out_type,
internal_type=internal_type,
name=name))
return ops.SparseTensor(indices_out, values_out, shape_out)
def _call_cell(self, inputs, initial_cell_state, initial_output, dtype,
sequence_length):
"""Run this LSTM on inputs, starting from the given state.
Args:
inputs: `3-D` tensor with shape `[time_len, batch_size, input_size]`
initial_cell_state: initial value for cell state, shape `[batch_size,
self._num_units]`
initial_output: initial value of cell output, shape `[batch_size,
self._num_units]`
dtype: The data type for the initial state and expected output.
sequence_length: Specifies the length of each sequence in inputs. An
`int32` or `int64` vector (tensor) size `[batch_size]`, values in `[0,
time_len)` or None.
Returns:
A pair containing:
- Cell state (cs): A `3-D` tensor of shape `[time_len, batch_size,
output_size]`
- Output (h): A `3-D` tensor of shape `[time_len, batch_size,
output_size]`
"""
inputs_shape = inputs.get_shape().with_rank(3)
time_len = inputs_shape[0].value
if time_len is None:
time_len = array_ops.shape(inputs)[0]
input_size = inputs_shape[2].value
w = vs.get_variable(
"W_0", [input_size + self._num_units, self._num_units * 4], dtype=dtype)
b = vs.get_variable(
"B", [w.get_shape().with_rank(2)[1]],
initializer=init_ops.constant_initializer(0.0),
dtype=dtype)
if self._use_peephole:
wci = vs.get_variable("W_I_diag", [self._num_units], dtype=dtype)
wco = vs.get_variable("W_O_diag", [self._num_units], dtype=dtype)
wcf = vs.get_variable("W_F_diag", [self._num_units], dtype=dtype)
else:
wci = wco = wcf = array_ops.zeros([self._num_units], dtype=dtype)
if sequence_length is None:
max_seq_len = time_len
else:
max_seq_len = math_ops.to_int64(math_ops.reduce_max(sequence_length))
_, cs, _, _, _, _, h = _lstm_ops_so.block_lstm(
seq_len_max=max_seq_len,
x=inputs,
cs_prev=initial_cell_state,
h_prev=initial_output,
w=w,
wci=wci,
wco=wco,
wcf=wcf,
b=b,
forget_bias=self._forget_bias,
cell_clip=self._cell_clip,
use_peephole=self._use_peephole)
return cs, h
def _reverse_seq(input_seq, lengths):
"""Reverse a list of Tensors up to specified lengths.
Args:
input_seq: Sequence of seq_len tensors of dimension (batch_size, n_features)
or nested tuples of tensors.
lengths: A `Tensor` of dimension batch_size, containing lengths for each
sequence in the batch. If "None" is specified, simply reverses
the list.
Returns:
time-reversed sequence
"""
if lengths is None:
return list(reversed(input_seq))
flat_input_seq = tuple(nest.flatten(input_) for input_ in input_seq)
flat_results = [[] for _ in range(len(input_seq))]
for sequence in zip(*flat_input_seq):
input_shape = tensor_shape.unknown_shape(
ndims=sequence[0].get_shape().ndims)
for input_ in sequence:
input_shape.merge_with(input_.get_shape())
input_.set_shape(input_shape)
# Join into (time, batch_size, depth)
s_joined = array_ops.pack(sequence)
# TODO(schuster, ebrevdo): Remove cast when reverse_sequence takes int32
if lengths is not None:
lengths = math_ops.to_int64(lengths)
# Reverse along dimension 0
s_reversed = array_ops.reverse_sequence(s_joined, lengths, 0, 1)
# Split again into list
result = array_ops.unpack(s_reversed)
for r, flat_result in zip(result, flat_results):
r.set_shape(input_shape)
flat_result.append(r)
results = [nest.pack_sequence_as(structure=input_, flat_sequence=flat_result)
for input_, flat_result in zip(input_seq, flat_results)]
return results
def _reverse_seq(input_seq, lengths):
"""Reverse a list of Tensors up to specified lengths.
Args:
input_seq: Sequence of seq_len tensors of dimension (batch_size, n_features)
or nested tuples of tensors.
lengths: A `Tensor` of dimension batch_size, containing lengths for each
sequence in the batch. If "None" is specified, simply reverses
the list.
Returns:
time-reversed sequence
"""
if lengths is None:
return list(reversed(input_seq))
flat_input_seq = tuple(nest.flatten(input_) for input_ in input_seq)
flat_results = [[] for _ in range(len(input_seq))]
for sequence in zip(*flat_input_seq):
input_shape = tensor_shape.unknown_shape(
ndims=sequence[0].get_shape().ndims)
for input_ in sequence:
input_shape.merge_with(input_.get_shape())
input_.set_shape(input_shape)
# Join into (time, batch_size, depth)
s_joined = array_ops.pack(sequence)
# TODO(schuster, ebrevdo): Remove cast when reverse_sequence takes int32
if lengths is not None:
lengths = math_ops.to_int64(lengths)
# Reverse along dimension 0
s_reversed = array_ops.reverse_sequence(s_joined, lengths, 0, 1)
# Split again into list
result = array_ops.unpack(s_reversed)
for r, flat_result in zip(result, flat_results):
r.set_shape(input_shape)
flat_result.append(r)
results = [nest.pack_sequence_as(structure=input_, flat_sequence=flat_result)
for input_, flat_result in zip(input_seq, flat_results)]
return results
