未验证 提交 2360dd20 编写于 作者: W whs 提交者: GitHub

Merge pull request #7438 from wanghaoshuang/ctc_py

Add python API for Warp-CTC op
...@@ -22,36 +22,13 @@ from ..param_attr import ParamAttr ...@@ -22,36 +22,13 @@ from ..param_attr import ParamAttr
from tensor import concat from tensor import concat
__all__ = [ __all__ = [
'fc', 'fc', 'embedding', 'dynamic_lstm', 'gru_unit', 'linear_chain_crf',
'embedding', 'crf_decoding', 'cos_sim', 'cross_entropy', 'square_error_cost', 'accuracy',
'dynamic_lstm', 'chunk_eval', 'sequence_conv', 'conv2d', 'sequence_pool', 'pool2d',
'gru_unit', 'batch_norm', 'beam_search_decode', 'conv2d_transpose', 'sequence_expand',
'linear_chain_crf', 'lstm_unit', 'reduce_sum', 'reduce_mean', 'reduce_max', 'reduce_min',
'crf_decoding', 'sequence_first_step', 'sequence_last_step', 'dropout', 'split',
'cos_sim', 'l2_normalize', 'matmul', 'warpctc'
'cross_entropy',
'square_error_cost',
'accuracy',
'chunk_eval',
'sequence_conv',
'conv2d',
'sequence_pool',
'pool2d',
'batch_norm',
'beam_search_decode',
'conv2d_transpose',
'sequence_expand',
'lstm_unit',
'reduce_sum',
'reduce_mean',
'reduce_max',
'reduce_min',
'sequence_first_step',
'sequence_last_step',
'dropout',
'split',
'l2_normalize',
'matmul',
] ]
...@@ -1721,29 +1698,29 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None): ...@@ -1721,29 +1698,29 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None):
def matmul(x, y, transpose_x=False, transpose_y=False, name=None): def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
""" """
Applies matrix multipication to two tensors. Currently only rank 1 to rank Applies matrix multipication to two tensors. Currently only rank 1 to rank
3 input tensors are supported. 3 input tensors are supported.
The actual behavior depends on the shapes of :math:`x`, :math:`y` and the The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically: flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
- If a transpose flag is specified, the last two dimensions of the tensor - If a transpose flag is specified, the last two dimensions of the tensor
are transposed. If the tensor is rank-1 of shape :math:`[D]`, then for are transposed. If the tensor is rank-1 of shape :math:`[D]`, then for
:math:`x` it is treated as :math:`[1, D]` in nontransposed form and as :math:`x` it is treated as :math:`[1, D]` in nontransposed form and as
:math:`[D, 1]` in transposed form, whereas for :math:`y` it is the :math:`[D, 1]` in transposed form, whereas for :math:`y` it is the
opposite: It is treated as :math:`[D, 1]` in nontransposed form and as opposite: It is treated as :math:`[D, 1]` in nontransposed form and as
:math:`[1, D]` in transposed form. :math:`[1, D]` in transposed form.
- After transpose, the two tensors are 2-D or 3-D and matrix multipication - After transpose, the two tensors are 2-D or 3-D and matrix multipication
performs in the following way. performs in the following way.
- If both are 2-D, they are multiplied like conventional matrices. - If both are 2-D, they are multiplied like conventional matrices.
- If either is 3-D, it is treated as a stack of matrices residing in the - If either is 3-D, it is treated as a stack of matrices residing in the
last two dimensions and a batched matrix multiply supporting broadcast last two dimensions and a batched matrix multiply supporting broadcast
applies on the two tensors. applies on the two tensors.
Also note that if the raw tensor :math:`x` or :math:`y` is rank-1 and Also note that if the raw tensor :math:`x` or :math:`y` is rank-1 and
nontransposed, the prepended or appended dimension :math:`1` will be nontransposed, the prepended or appended dimension :math:`1` will be
removed after matrix multipication. removed after matrix multipication.
Args: Args:
...@@ -1751,7 +1728,7 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None): ...@@ -1751,7 +1728,7 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
y (Variable): The input variable which is a Tensor or LoDTensor. y (Variable): The input variable which is a Tensor or LoDTensor.
transpose_x (bool): Whether to transpose :math:`x` before multiplication. transpose_x (bool): Whether to transpose :math:`x` before multiplication.
transpose_y (bool): Whether to transpose :math:`y` before multiplication. transpose_y (bool): Whether to transpose :math:`y` before multiplication.
name(str|None): A name for this layer(optional). If set None, the layer name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically. will be named automatically.
Returns: Returns:
...@@ -1788,3 +1765,56 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None): ...@@ -1788,3 +1765,56 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
attrs={'transpose_X': transpose_x, attrs={'transpose_X': transpose_x,
'transpose_Y': transpose_y}) 'transpose_Y': transpose_y})
return out return out
def warpctc(input, label, blank=0, norm_by_times=False, **kwargs):
"""
An operator integrating the open source Warp-CTC library
(https://github.com/baidu-research/warp-ctc)
to compute Connectionist Temporal Classification (CTC) loss.
It can be aliased as softmax with CTC, since a native softmax activation is
interated to the Warp-CTC library, to to normlize values for each row of the
input tensor.
Args:
input(Variable): (LodTensor, default: LoDTensor<float>),
the unscaled probabilities of variable-length sequences,
which is a 2-D Tensor with LoD information.
It's shape is [Lp, num_classes + 1], where Lp is the sum of all input
sequences' length and num_classes is the true number of classes.
(not including the blank label).
label(Variable): (LodTensor, default: LoDTensor<int>), the ground truth
of variable-length sequence, which is a 2-D Tensor with LoD
information. It is of the shape [Lg, 1], where Lg is th sum of
all labels' length.
blank: (int, default: 0), the blank label index of Connectionist
Temporal Classification (CTC) loss, which is in the
half-opened interval [0, num_classes + 1).
norm_by_times: (bool, default: false), whether to normalize
the gradients by the number of time-step,which is also the
sequence's length. There is no need to normalize the gradients
if warpctc layer was follewed by a mean_op.
Returns:
Variable: The Connectionist Temporal Classification (CTC) loss,
which is a 2-D Tensor of the shape [batch_size, 1].
Examples:
.. code-block:: python
y = layers.data(name='y', shape=[11, 8], dtype='float32', lod_level=1)
y_predict = layers.data(name='y_predict', shape=[11, 1], dtype='float32')
cost = layers.warpctc(input=y_predict, label=y)
"""
helper = LayerHelper('warpctc', **kwargs)
loss_out = helper.create_tmp_variable(dtype=input.dtype)
grad_out = helper.create_tmp_variable(dtype=input.dtype)
helper.append_op(
type='warpctc',
inputs={'Logits': [input],
'Label': [label]},
outputs={'WarpCTCGrad': [grad_out],
'Loss': [loss_out]},
attrs={'blank': blank,
'norm_by_times': norm_by_times})
return loss_out
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册