def test_reshape_layer(self):
block_expand = layer.block_expand(
input=conv, num_channels=4, stride_x=1, block_x=1)
expand = layer.expand(
input=weight,
expand_as=pixel,
expand_level=layer.ExpandLevel.FROM_TIMESTEP)
repeat = layer.repeat(input=pixel, num_repeats=4)
reshape = layer.seq_reshape(input=pixel, reshape_size=4)
rotate = layer.rotate(input=pixel, height=16, width=49)
print layer.parse_network(block_expand, expand, repeat, reshape, rotate)
def test_reshape_layer(self):
block_expand = layer.block_expand(
input=conv, num_channels=4, stride_x=1, block_x=1)
expand = layer.expand(
input=weight,
expand_as=pixel,
expand_level=layer.ExpandLevel.FROM_NO_SEQUENCE)
repeat = layer.repeat(input=pixel, num_repeats=4)
reshape = layer.seq_reshape(input=pixel, reshape_size=4)
rotate = layer.rotate(input=pixel, height=16, width=49)
print layer.parse_network(
[block_expand, expand, repeat, reshape, rotate])
def __build_nn__(self):
'''
Build the network topology.
'''
# Get the image features with CNN.
conv_features = self.conv_groups(self.image, conf.filter_num,
conf.with_bn)
# Expand the output of CNN into a sequence of feature vectors.
sliced_feature = layer.block_expand(
input=conv_features,
num_channels=conf.num_channels,
stride_x=conf.stride_x,
stride_y=conf.stride_y,
block_x=conf.block_x,
block_y=conf.block_y)
# Use RNN to capture sequence information forwards and backwards.
gru_forward = simple_gru(
input=sliced_feature, size=conf.hidden_size, act=Relu())
gru_backward = simple_gru(
input=sliced_feature,
size=conf.hidden_size,
act=Relu(),
reverse=True)
# Map the output of RNN to character distribution.
self.output = layer.fc(input=[gru_forward, gru_backward],
size=self.num_classes + 1,
act=Linear())
self.log_probs = paddle.layer.mixed(
input=paddle.layer.identity_projection(input=self.output),
act=paddle.activation.Softmax())
# Use warp CTC to calculate cost for a CTC task.
if not self.is_infer:
self.cost = layer.warp_ctc(
input=self.output,
label=self.label,
size=self.num_classes + 1,
norm_by_times=conf.norm_by_times,
blank=self.num_classes)
self.eval = evaluator.ctc_error(input=self.output, label=self.label)
def __build_nn__(self):
'''
建立网络拓扑
'''
# 通过CNN获取图像特征
conv_features = self.conv_groups(self.image, conf.filter_num,
conf.with_bn)
# 将CNN的输出展开成一系列特征向量。
sliced_feature = layer.block_expand(
input=conv_features,
num_channels=conf.num_channels,
stride_x=conf.stride_x,
stride_y=conf.stride_y,
block_x=conf.block_x,
block_y=conf.block_y)
# 使用RNN向前和向后捕获序列信息。
gru_forward = simple_gru(
input=sliced_feature, size=conf.hidden_size, act=Relu())
gru_backward = simple_gru(
input=sliced_feature,
size=conf.hidden_size,
act=Relu(),
reverse=True)
# 将RNN的输出映射到字符分布。
self.output = layer.fc(input=[gru_forward, gru_backward],
size=self.num_classes + 1,
act=Linear())
self.log_probs = paddle.layer.mixed(
input=paddle.layer.identity_projection(input=self.output),
act=paddle.activation.Softmax())
# 使用扭曲CTC来计算CTC任务的成本。
if not self.is_infer:
self.cost = layer.warp_ctc(
input=self.output,
label=self.label,
size=self.num_classes + 1,
norm_by_times=conf.norm_by_times,
blank=self.num_classes)
self.eval = evaluator.ctc_error(input=self.output, label=self.label)