def compile(self, input_image: tf.Tensor, paf_label_layer: MakiTensor,
heatmap_label_layer: MakiTensor):
"""
Apply label correction for label heatmap and paf tensors by taken output from teacher
Parameters
----------
input_image : tf.Tensor
paf_label_layer : MakiTensor
heatmap_label_layer : MakiTensor
Returns
-------
paf : MakiLayerWrapper
Final paf layer with correction
heatmap : MakiLayerWrapper
Final heatmap layer with correction
"""
teacher_paf_tensor, teacher_heatmap_tensor = self.__init_teacher(
input_image)
paf_corrected = self.__label_correction_paf(
t_paf=teacher_paf_tensor, l_paf=paf_label_layer.get_data_tensor())
heatmap_corrected = self.__label_correction_heatmap(
t_heatmap=teacher_heatmap_tensor,
l_heatmap=heatmap_label_layer.get_data_tensor())
paf_mf = MakiLayerWrapper(paf_corrected)
heatmap_mf = MakiLayerWrapper(heatmap_corrected)
return paf_mf, heatmap_mf
def identity_block(x: MakiTensor,
block_id: int,
unit_id: int,
num_block=None,
in_f=None,
use_bias=False,
activation=tf.nn.relu,
bn_params={}):
"""
Parameters
----------
x : MakiTensor
Input MakiTensor.
in_f : int
Number of input feature maps. By default None (shape will be getted from tensor).
activation : tensorflow function
The function of activation, by default tf.nn.relu.
use_bias : bool
Use bias on layers or not.
block_id : int
Number of block (used in name of layers).
unit_id : int
Unit of block (used in name of layers).
num_block : int
Number of sum operation (used in name of layers).
bn_params : dict
Parameters for BatchNormLayer. If empty all parameters will have default valued.
Returns
---------
x : MakiTensor
Output MakiTensor.
"""
prefix_name = 'block' + str(block_id) + '/unit_' + str(unit_id)
if num_block is None:
num_block = prefix_name + '/sum_operation'
else:
num_block = 'add_' + str(num_block)
if in_f is None:
in_f = x.get_shape()[-1]
reduction = int(in_f / 4)
mx = ConvLayer(kw=1,
kh=1,
in_f=in_f,
out_f=reduction,
activation=None,
use_bias=use_bias,
name=prefix_name + '/bottleneck_v1/conv1/weights')(x)
mx = BatchNormLayer(D=reduction,
name=prefix_name + '/bottleneck_v1/conv1/BatchNorm',
**bn_params)(mx)
mx = ActivationLayer(activation=activation,
name=prefix_name + '/bottleneck_v1/conv1/activ')(mx)
mx = ConvLayer(kw=3,
kh=3,
in_f=reduction,
out_f=reduction,
activation=None,
use_bias=use_bias,
name=prefix_name + '/bottleneck_v1/conv2/weights')(mx)
mx = BatchNormLayer(D=reduction,
name=prefix_name + '/bottleneck_v1/conv2/BatchNorm',
**bn_params)(mx)
mx = ActivationLayer(activation=activation,
name=prefix_name + '/bottleneck_v1/conv2/activ')(mx)
mx = ConvLayer(kw=1,
kh=1,
in_f=reduction,
out_f=in_f,
activation=None,
use_bias=use_bias,
name=prefix_name + '/bottleneck_v1/conv3/weights')(mx)
mx = BatchNormLayer(D=in_f,
name=prefix_name + '/bottleneck_v1/conv3/BatchNorm',
**bn_params)(mx)
x = SumLayer(name=num_block)([mx, x])
return x
def without_pointwise_CB(x: MakiTensor,
block_id: int,
unit_id: int,
num_block=None,
in_f=None,
use_bias=False,
activation=tf.nn.relu,
stride=2,
out_f=None,
bn_params={}):
"""
Parameters
----------
x : MakiTensor
Input MakiTensor.
in_f : int
Number of input feature maps. By default is None (shape will be getted from tensor).
out_f : int
Number of output feature maps. By default is None which means out_f = 2 * in_f.
activation : tensorflow function
The function of activation. By default tf.nn.relu.
use_bias : bool
Use bias on layers or not.
block_id : int
Number of block (used in name of layers).
unit_id : int
Unit of block (used in name of layers).
num_block : int
Number of sum operation (used in name of layers).
bn_params : dict
Parameters for BatchNormLayer. If empty all parameters will have default valued.
Returns
---------
x : MakiTensor
Output MakiTensor.
"""
prefix_name = 'stage' + str(block_id) + '_unit' + str(unit_id) + '_'
if num_block is None:
num_block = prefix_name + '/sum_operation'
else:
num_block = 'add_' + str(num_block)
if in_f is None:
in_f = x.get_shape()[-1]
if out_f is None:
out_f = int(2 * in_f)
x = BatchNormLayer(D=in_f, name=prefix_name + 'bn1', **bn_params)(x)
x = ActivationLayer(activation=activation,
name=prefix_name + 'activation_1')(x)
mx = ZeroPaddingLayer(padding=[[1, 1], [1, 1]],
name=prefix_name + 'zero_pad_1')(x)
mx = ConvLayer(kw=3,
kh=3,
in_f=in_f,
out_f=out_f,
activation=None,
stride=stride,
padding='VALID',
use_bias=use_bias,
name=prefix_name + 'conv1')(mx)
mx = BatchNormLayer(D=out_f, name=prefix_name + 'bn2', **bn_params)(mx)
mx = ActivationLayer(activation=activation,
name=prefix_name + 'activation_2')(mx)
mx = ZeroPaddingLayer(padding=[[1, 1], [1, 1]],
name=prefix_name + 'zero_pad_2')(mx)
mx = ConvLayer(kw=3,
kh=3,
in_f=out_f,
out_f=out_f,
activation=None,
padding='VALID',
use_bias=use_bias,
name=prefix_name + 'conv2')(mx)
sx = ConvLayer(kw=1,
kh=1,
in_f=in_f,
out_f=out_f,
stride=stride,
padding='VALID',
activation=None,
use_bias=use_bias,
name=prefix_name + 'sc/conv')(x)
x = SumLayer(name=num_block)([mx, sx])
return x
def to_makitensor(x, name):
return MakiTensor(x,
parent_layer=None,
parent_tensor_names=[],
previous_tensors={},
name=name)