提交 9f3cbed2 编写于 作者: Q qingqing01 提交者: Yu Yang

Add more details for CTC layer, fix CTC evalutor and add their interface test (#74)

* Add some comments for CTC layer and fix CTC evalutors, also add interface test
上级 93006787
......@@ -25,7 +25,7 @@ repo or just head straight to the command line:
```shell
# Clone your fork to your local machine
git clone git@github.com:USERNAME/paddle.git
git clone git@github.com:USERNAME/Paddle.git
```
Then you can start to develop.
......@@ -52,7 +52,7 @@ To do this, you'll need to add a remote at first:
# see the current configured remote repository
git remote -v
# add upstream repository
git remote add upstream https://github.com/paddle/paddle.git
git remote add upstream https://github.com/baidu/Paddle.git
# verify the new upstream
git remote -v
```
......
......@@ -296,6 +296,7 @@ def precision_recall_evaluator(
@wrap_name_default()
def ctc_error_evaluator(
input,
label,
name=None,
):
"""
......@@ -305,16 +306,19 @@ def ctc_error_evaluator(
.. code-block:: python
eval = ctc_error_evaluator(input)
eval = ctc_error_evaluator(input=input, label=lbl)
:param name: Evaluator name.
:type name: None|basestring
:param input: Input Layer.
:type input: LayerOutput
:param label: input label, which is a data_layer.
:type label: LayerOutput
"""
evaluator_base(name=name,
type="ctc_edit_distance",
input=input)
input=input,
label=label)
@evaluator(EvaluatorAttribute.FOR_CLASSIFICATION)
@wrap_name_default()
......
......@@ -2944,7 +2944,7 @@ def linear_comb_layer(weights, vectors, size, name=None):
.. math::
z = x^T Y
z = x^\mathrm{T} Y
In this formular:
- :math:`x`: weights
......@@ -3064,6 +3064,17 @@ def ctc_layer(input, label, size, name=None, norm_by_times=False):
classication task. That is, for sequence labeling problems where the
alignment between the inputs and the target labels is unknown.
More details can be found by referring to `Connectionist Temporal
Classification: Labelling Unsegmented Sequence Data with Recurrent
Neural Networks <http://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf>`_
Note:
Considering the 'blank' label needed by CTC, you need to use
(num_classes + 1) as the input size. num_classes is the category number.
And the 'blank' is the last category index. So the size of 'input' layer, such as
fc_layer with softmax activation, should be num_classes + 1. The size of ctc_layer
should also be num_classes + 1.
The simple usage:
.. code-block:: python
......@@ -3077,7 +3088,7 @@ def ctc_layer(input, label, size, name=None, norm_by_times=False):
:type input: LayerOutput
:param label: The data layer of label with variable length.
:type label: LayerOutput
:param size: category numbers.
:param size: category numbers + 1.
:type size: int
:param name: The name of this layer, which can not specify.
:type name: string|None
......
......@@ -34,6 +34,15 @@ out = fc_layer(input=[cos1, cos3, linear_comb],
outputs(classification_cost(out, data_layer(name="label", size=num_classes)))
# for ctc
tmp = fc_layer(input=x1,
size=num_classes + 1,
act=SoftmaxActivation())
ctc = ctc_layer(input=tmp,
label=y,
size=num_classes + 1)
ctc_eval = ctc_error_evaluator(input=ctc, label=y)
settings(
batch_size=10,
learning_rate=2e-3,
......
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