def middle_flow(self, data):
block_num = self.bottleneck_params["middle_flow"][0]
strides = self.bottleneck_params["middle_flow"][1]
chns = self.bottleneck_params["middle_flow"][2]
strides = check_data(strides, block_num)
chns = check_data(chns, block_num)
# params to control your flow
s = self.stride
block_point = self.block_point
output_stride = self.output_stride
with scope("middle_flow"):
for i in range(block_num):
block_point = block_point + 1
with scope("block" + str(i + 1)):
stride = strides[i] if check_stride(
s * strides[i], output_stride) else 1
data, short_cuts = self.xception_block(data,
chns[i],
[1, 1, strides[i]],
skip_conv=False)
s = s * stride
if check_points(block_point, self.decode_points):
self.short_cuts[block_point] = short_cuts[1]
self.stride = s
self.block_point = block_point
return data
def _decoder_with_concat(encode_data, decode_shortcut, param_attr):
with scope('concat'):
decode_shortcut = bn_relu(
conv(
decode_shortcut,
48,
1,
1,
groups=1,
padding=0,
param_attr=param_attr))
encode_data = fluid.layers.resize_bilinear(encode_data,
decode_shortcut.shape[2:])
encode_data = fluid.layers.concat([encode_data, decode_shortcut],
axis=1)
if cfg.MODEL.DEEPLAB.DECODER_USE_SEP_CONV:
with scope("separable_conv1"):
encode_data = separate_conv(
encode_data,
cfg.MODEL.DEEPLAB.DECODER.CONV_FILTERS,
1,
3,
dilation=1,
act=relu)
with scope("separable_conv2"):
encode_data = separate_conv(
encode_data,
cfg.MODEL.DEEPLAB.DECODER.CONV_FILTERS,
1,
3,
dilation=1,
act=relu)
else:
with scope("decoder_conv1"):
encode_data = bn_relu(
conv(
encode_data,
cfg.MODEL.DEEPLAB.DECODER.CONV_FILTERS,
stride=1,
filter_size=3,
dilation=1,
padding=1,
param_attr=param_attr))
with scope("decoder_conv2"):
encode_data = bn_relu(
conv(
encode_data,
cfg.MODEL.DEEPLAB.DECODER.CONV_FILTERS,
stride=1,
filter_size=3,
dilation=1,
padding=1,
param_attr=param_attr))
return encode_data
def xception_block(self,
input,
channels,
strides=1,
filters=3,
dilation=1,
skip_conv=True,
has_skip=True,
activation_fn_in_separable_conv=False):
repeat_number = 3
channels = check_data(channels, repeat_number)
filters = check_data(filters, repeat_number)
strides = check_data(strides, repeat_number)
data = input
results = []
for i in range(repeat_number):
with scope('separable_conv' + str(i + 1)):
if not activation_fn_in_separable_conv:
data = relu(data)
data = separate_conv(
data,
channels[i],
strides[i],
filters[i],
dilation=dilation)
else:
data = separate_conv(
data,
channels[i],
strides[i],
filters[i],
dilation=dilation,
act=relu)
results.append(data)
if not has_skip:
return data, results
if skip_conv:
param_attr = fluid.ParamAttr(
name=name_scope + 'weights',
regularizer=None,
initializer=fluid.initializer.TruncatedNormal(
loc=0.0, scale=0.09))
with scope('shortcut'):
skip = bn(
conv(
input,
channels[-1],
1,
strides[-1],
groups=1,
padding=0,
param_attr=param_attr))
else:
skip = input
return data + skip, results
def entry_flow(self, data):
param_attr = fluid.ParamAttr(
name=name_scope + 'weights',
regularizer=None,
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.09))
with scope("entry_flow"):
with scope("conv1"):
data = bn_relu(
conv(data,
32,
3,
stride=2,
padding=1,
param_attr=param_attr))
with scope("conv2"):
data = bn_relu(
conv(data,
64,
3,
stride=1,
padding=1,
param_attr=param_attr))
# get entry flow params
block_num = self.bottleneck_params["entry_flow"][0]
strides = self.bottleneck_params["entry_flow"][1]
chns = self.bottleneck_params["entry_flow"][2]
strides = check_data(strides, block_num)
chns = check_data(chns, block_num)
#print("entry:", block_num, strides, chns)
# params to control your flow
s = self.stride
block_point = self.block_point
output_stride = self.output_stride
#print("entry:", s, block_point, output_stride)
with scope("entry_flow"):
for i in range(block_num):
block_point = block_point + 1
with scope("block" + str(i + 1)):
stride = strides[i] if check_stride(
s * strides[i], output_stride) else 1
data, short_cuts = self.xception_block(
data, chns[i], [1, 1, stride])
s = s * stride
if check_points(block_point, self.decode_points):
#print("decode shortcut:", block_point)
self.short_cuts[block_point] = short_cuts[1]
#print("entry:", i, data.shape)
self.stride = s
self.block_point = block_point
#print("entry:", s, block_point, output_stride)
return data
def decode(data, short_cuts):
# 解码器设置，与编码器对称
with scope("decode"):
with scope("decode1"):
data = up(data, short_cuts[3], 256)
with scope("decode2"):
data = up(data, short_cuts[2], 128)
with scope("decode3"):
data = up(data, short_cuts[1], 64)
with scope("decode4"):
data = up(data, short_cuts[0], 64)
return data
def double_conv(data, out_ch):
param_attr = fluid.ParamAttr(
name='weights',
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0),
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.33))
with scope("conv0"):
data = bn_relu(
conv(data, out_ch, 3, stride=1, padding=1, param_attr=param_attr))
with scope("conv1"):
data = bn_relu(
conv(data, out_ch, 3, stride=1, padding=1, param_attr=param_attr))
return data
def sub_net_1(input):
with scope("conv1_sub1"):
tmp = conv(input, 32, 3, 2, padding=1)
tmp = bn(tmp, act='relu')
with scope("conv2_sub1"):
tmp = conv(tmp, 32, 3, 2, padding=1)
tmp = bn(tmp, act='relu')
with scope("conv3_sub1"):
tmp = conv(tmp, 64, 3, 2, padding=1)
tmp = bn(tmp, act='relu')
with scope("conv3_sub1_proj"):
tmp = conv(tmp, 128, 1, 1)
tmp = bn(tmp)
return tmp
def sub_net_4(input, input_shape):
tmp = pyramis_pooling(input, input_shape)
with scope("conv5_4_k1"):
tmp = conv(tmp, 256, 1, 1)
tmp = bn(tmp, act='relu')
tmp = interp(tmp, out_shape=np.ceil(input_shape / 16))
return tmp
def psp_module(input, out_features):
cat_layers = []
sizes = (1, 2, 3, 6)
for size in sizes:
psp_name = "psp" + str(size)
with scope(psp_name):
pool = fluid.layers.adaptive_pool2d(input,
pool_size=[size, size],
pool_type='avg',
name=psp_name + '_adapool')
data = conv(pool,
out_features,
filter_size=1,
bias_attr=False,
name=psp_name + '_conv')
data_bn = bn(data, act='relu')
interp = fluid.layers.resize_bilinear(data_bn,
out_shape=input.shape[2:],
name=psp_name + '_interp',
align_mode=0)
cat_layers.append(interp)
cat_layers = [input] + cat_layers
out = fluid.layers.concat(cat_layers, axis=1, name='psp_cat')
return out
def _decoder_with_sum_merge(encode_data, decode_shortcut, param_attr):
encode_data = fluid.layers.resize_bilinear(encode_data,
decode_shortcut.shape[2:])
encode_data = conv(
encode_data,
#cfg.MODEL.DEEPLAB.DECODER.CONV_FILTERS,
256,
1,
1,
groups=1,
padding=0,
param_attr=param_attr)
with scope('merge'):
decode_shortcut = conv(
decode_shortcut,
#cfg.MODEL.DEEPLAB.DECODER.CONV_FILTERS,
256,
1,
1,
groups=1,
padding=0,
param_attr=param_attr)
return encode_data + decode_shortcut
def deeplabv3p_nas(img, num_classes, arch=None):
data, decode_shortcut = nas_backbone(img, arch)
# 编码器解码器设置
cfg.MODEL.DEFAULT_EPSILON = 1e-5
if cfg.MODEL.DEEPLAB.ENCODER_WITH_ASPP:
data = encoder(data)
if cfg.MODEL.DEEPLAB.ENABLE_DECODER:
data = decoder(data, decode_shortcut)
# 根据类别数设置最后一个卷积层输出，并resize到图片原始尺寸
param_attr = fluid.ParamAttr(
name=name_scope + 'weights',
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0),
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.01))
with scope('logit'):
logit = conv(data,
num_classes,
1,
stride=1,
padding=0,
bias_attr=True,
param_attr=param_attr)
logit = fluid.layers.resize_bilinear(logit, img.shape[2:])
return logit
def deeplabv3p(img, num_classes):
# Backbone设置：xception 或 mobilenetv2
if 'xception' in cfg.MODEL.DEEPLAB.BACKBONE:
data, decode_shortcut = xception(img)
elif 'mobilenet' in cfg.MODEL.DEEPLAB.BACKBONE:
data, decode_shortcut = mobilenetv2(img)
else:
raise Exception("deeplab only support xception and mobilenet backbone")
# 编码器解码器设置
cfg.MODEL.DEFAULT_EPSILON = 1e-5
if cfg.MODEL.DEEPLAB.ENCODER_WITH_ASPP:
data = encoder(data)
if cfg.MODEL.DEEPLAB.ENABLE_DECODER:
data = decoder(data, decode_shortcut)
# 根据类别数设置最后一个卷积层输出，并resize到图片原始尺寸
param_attr = fluid.ParamAttr(
name=name_scope + 'weights',
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0),
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.01))
with scope('logit'):
with fluid.name_scope('last_conv'):
logit = conv(data,
num_classes,
1,
stride=1,
padding=0,
bias_attr=True,
param_attr=param_attr)
logit = fluid.layers.resize_bilinear(logit, img.shape[2:])
return logit
def deeplabv3p(img, num_classes):
# Backbone设置：xception 或 mobilenetv2
data, decode_shortcut = resnet_vd(img)
# 编码器解码器设置
#cfg.MODEL.DEFAULT_EPSILON = 1e-5
#if cfg.MODEL.DEEPLAB.ENCODER_WITH_ASPP:
data = encoder(data)
#if cfg.MODEL.DEEPLAB.ENABLE_DECODER:
data = decoder(data, decode_shortcut)
# 根据类别数设置最后一个卷积层输出，并resize到图片原始尺寸
param_attr = fluid.ParamAttr(
name=name_scope + 'weights',
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0),
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.01))
#if not cfg.MODEL.DEEPLAB.DECODER.OUTPUT_IS_LOGITS:
with scope('logit'):
with fluid.name_scope('last_conv'):
logit = conv(
data,
num_classes,
1,
stride=1,
padding=0,
bias_attr=True,
param_attr=param_attr)
#else:
# logit = data
logit = fluid.layers.resize_bilinear(logit, img.shape[2:])
return logit
def exit_flow(self, data):
block_num = self.bottleneck_params["exit_flow"][0]
strides = self.bottleneck_params["exit_flow"][1]
chns = self.bottleneck_params["exit_flow"][2]
strides = check_data(strides, block_num)
chns = check_data(chns, block_num)
#print("exit:", block_num, strides, chns)
assert (block_num == 2)
# params to control your flow
s = self.stride
block_point = self.block_point
output_stride = self.output_stride
#print("exit:", s, block_point, output_stride)
with scope("exit_flow"):
with scope('block1'):
block_point += 1
stride = strides[0] if check_stride(s * strides[0],
output_stride) else 1
data, short_cuts = self.xception_block(data, chns[0],
[1, 1, stride])
s = s * stride
if check_points(block_point, self.decode_points):
#print("decode shortcut:", block_point)
self.short_cuts[block_point] = short_cuts[1]
#print("exit:", 0, data.shape)
with scope('block2'):
block_point += 1
stride = strides[1] if check_stride(s * strides[1],
output_stride) else 1
data, short_cuts = self.xception_block(
data,
chns[1], [1, 1, stride],
dilation=2,
has_skip=False,
activation_fn_in_separable_conv=True)
s = s * stride
if check_points(block_point, self.decode_points):
#print("decode shortcut:", block_point)
self.short_cuts[block_point] = short_cuts[1]
#print("exit:", 1, data.shape)
self.stride = s
self.block_point = block_point
#print("exit:", s, block_point, output_stride)
return data
def net(self, x):
with scope('dsconv1'):
x = separate_conv(x,
self.dw_channels,
stride=self.stride,
filter=3,
act=fluid.layers.relu)
with scope('dsconv2'):
x = separate_conv(x,
self.dw_channels,
stride=self.stride,
filter=3,
act=fluid.layers.relu)
x = dropout2d(x, 0.1, is_train=cfg.PHASE == 'train')
x = conv(x, self.num_classes, 1, bias_attr=True)
return x
def CCF24(sub2_out, sub4_out, input_shape):
with scope("conv_sub4"):
tmp = conv(sub4_out, 128, 3, dilation=2, padding=2)
tmp = bn(tmp)
tmp = tmp + sub2_out
tmp = fluid.layers.relu(tmp)
tmp = interp(tmp, np.ceil(input_shape / 8))
return tmp
def learning_to_downsample(x,
dw_channels1=32,
dw_channels2=48,
out_channels=64):
x = relu(bn(conv(x, dw_channels1, 3, 2)))
with scope('dsconv1'):
x = separate_conv(x,
dw_channels2,
stride=2,
filter=3,
act=fluid.layers.relu)
with scope('dsconv2'):
x = separate_conv(x,
out_channels,
stride=2,
filter=3,
act=fluid.layers.relu)
return x
def CCF124(sub1_out, sub24_out, input_shape):
tmp = zero_padding(sub24_out, padding=2)
with scope("conv_sub2"):
tmp = conv(tmp, 128, 3, dilation=2)
tmp = bn(tmp)
tmp = tmp + sub1_out
tmp = fluid.layers.relu(tmp)
tmp = interp(tmp, input_shape // 4)
return tmp
def net(self, higher_res_feature, lower_res_feature):
h, w = higher_res_feature.shape[2:]
lower_res_feature = fluid.layers.resize_bilinear(lower_res_feature,
[h, w],
align_mode=0)
with scope('dwconv'):
lower_res_feature = relu(
bn(conv(lower_res_feature, self.out_channels,
1))) #(lower_res_feature)
with scope('conv_lower_res'):
lower_res_feature = bn(
conv(lower_res_feature, self.out_channels, 1, bias_attr=True))
with scope('conv_higher_res'):
higher_res_feature = bn(
conv(higher_res_feature, self.out_channels, 1, bias_attr=True))
out = higher_res_feature + lower_res_feature
return relu(out)
def get_logit(data, num_classes):
# 根据类别数设置最后一个卷积层输出
param_attr = fluid.ParamAttr(
name='weights',
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0),
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.01))
with scope("logit"):
data = conv(
data, num_classes, 3, stride=1, padding=1, param_attr=param_attr)
return data
def psp_module(input, out_features):
# Pyramid Scene Parsing 金字塔池化模块
# 输入：backbone输出的特征
# 输出：对输入进行不同尺度pooling, 卷积操作后插值回原始尺寸，并concat
# 最后进行一个卷积及BN操作
cat_layers = []
sizes = (1, 2, 3, 6)
for size in sizes:
psp_name = "psp" + str(size)
with scope(psp_name):
pool = fluid.layers.adaptive_pool2d(
input,
pool_size=[size, size],
pool_type='avg',
name=psp_name + '_adapool')
data = conv(
pool,
out_features,
filter_size=1,
bias_attr=True,
name=psp_name + '_conv')
data_bn = bn(data, act='relu')
interp = fluid.layers.resize_bilinear(
data_bn, out_shape=input.shape[2:], name=psp_name + '_interp')
cat_layers.append(interp)
cat_layers = [input] + cat_layers[::-1]
cat = fluid.layers.concat(cat_layers, axis=1, name='psp_cat')
psp_end_name = "psp_end"
with scope(psp_end_name):
data = conv(
cat,
out_features,
filter_size=3,
padding=1,
bias_attr=True,
name=psp_end_name)
out = bn(data, act='relu')
return out
def net(self,
input,
output_stride=32,
num_classes=1000,
end_points=None,
decode_points=None):
self.stride = 2
self.block_point = 0
self.output_stride = output_stride
self.decode_points = decode_points
self.short_cuts = dict()
with scope(self.backbone):
# Entry flow
data = self.entry_flow(input)
if check_points(self.block_point, end_points):
return data, self.short_cuts
# Middle flow
data = self.middle_flow(data)
if check_points(self.block_point, end_points):
return data, self.short_cuts
# Exit flow
data = self.exit_flow(data)
if check_points(self.block_point, end_points):
return data, self.short_cuts
data = fluid.layers.reduce_mean(data, [2, 3], keep_dim=True)
data = fluid.layers.dropout(data, 0.5)
stdv = 1.0 / math.sqrt(data.shape[1] * 1.0)
with scope("logit"):
out = fluid.layers.fc(
input=data,
size=num_classes,
act='softmax',
param_attr=fluid.param_attr.ParamAttr(
name='weights',
initializer=fluid.initializer.Uniform(-stdv, stdv)),
bias_attr=fluid.param_attr.ParamAttr(name='bias'))
return out
def decoder(encode_data, decode_shortcut):
# 解码器配置
# encode_data：编码器输出
# decode_shortcut: 从backbone引出的分支, resize后与encode_data concat
# DECODER_USE_SEP_CONV: 默认为真，则concat后连接两个可分离卷积，否则为普通卷积
param_attr = fluid.ParamAttr(name=name_scope + 'weights',
regularizer=None,
initializer=fluid.initializer.TruncatedNormal(
loc=0.0, scale=0.06))
with scope('decoder'):
with scope('concat'):
decode_shortcut = bn_relu(
conv(decode_shortcut,
48,
1,
1,
groups=1,
padding=0,
param_attr=param_attr))
encode_data = fluid.layers.resize_bilinear(
encode_data, decode_shortcut.shape[2:])
encode_data = fluid.layers.concat([encode_data, decode_shortcut],
axis=1)
if cfg.MODEL.DEEPLAB.DECODER_USE_SEP_CONV:
with scope("separable_conv1"):
encode_data = separate_conv(encode_data,
256,
1,
3,
dilation=1,
act=relu)
with scope("separable_conv2"):
encode_data = separate_conv(encode_data,
256,
1,
3,
dilation=1,
act=relu)
else:
with scope("decoder_conv1"):
encode_data = bn_relu(
conv(encode_data,
256,
stride=1,
filter_size=3,
dilation=1,
padding=1,
param_attr=param_attr))
with scope("decoder_conv2"):
encode_data = bn_relu(
conv(encode_data,
256,
stride=1,
filter_size=3,
dilation=1,
padding=1,
param_attr=param_attr))
return encode_data
def net(self, x):
x, _ = inverted_blocks(x, self.in_channels, self.t,
self.block_channels[0], self.num_blocks[0], 2,
'inverted_block_1')
x, _ = inverted_blocks(x, self.block_channels[0], self.t,
self.block_channels[1], self.num_blocks[1], 2,
'inverted_block_2')
x, _ = inverted_blocks(x, self.block_channels[1], self.t,
self.block_channels[2], self.num_blocks[2], 1,
'inverted_block_3')
x = psp_module(x, self.block_channels[2] // 4)
with scope('out'):
x = relu(bn(conv(x, self.out_channels, 1)))
return x
def decoder(encode_data, decode_shortcut):
# 解码器配置
# encode_data：编码器输出
# decode_shortcut: 从backbone引出的分支, resize后与encode_data concat
# DECODER_USE_SEP_CONV: 默认为真，则concat后连接两个可分离卷积，否则为普通卷积
param_attr = fluid.ParamAttr(
name=name_scope + 'weights',
regularizer=None,
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.06))
with scope('decoder'):
#if cfg.MODEL.DEEPLAB.DECODER.USE_SUM_MERGE:
# return _decoder_with_sum_merge(encode_data, decode_shortcut,
# param_attr)
return _decoder_with_concat(encode_data, decode_shortcut, param_attr)
def get_logit(data, num_classes, name="logit"):
param_attr = fluid.ParamAttr(
name=name + 'weights',
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0),
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.01))
with scope(name):
data = conv(data,
num_classes,
1,
stride=1,
padding=0,
param_attr=param_attr,
bias_attr=True)
return data
def get_logit_interp(input, num_classes, out_shape, name="logit"):
# 根据类别数决定最后一层卷积输出, 并插值回原始尺寸
param_attr = fluid.ParamAttr(
name=name + 'weights',
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0),
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.01))
with scope(name):
logit = conv(
input,
num_classes,
filter_size=1,
param_attr=param_attr,
bias_attr=True,
name=name + '_conv')
logit_interp = fluid.layers.resize_bilinear(
logit, out_shape=out_shape, name=name + '_interp')
return logit_interp
def up(data, short_cut, out_ch):
# 上采样：data上采样(resize或deconv), 并与short_cut concat
param_attr = fluid.ParamAttr(
name='weights',
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0),
initializer=fluid.initializer.XavierInitializer(),
)
with scope("up"):
if cfg.MODEL.UNET.UPSAMPLE_MODE == 'bilinear':
data = fluid.layers.resize_bilinear(data, short_cut.shape[2:])
else:
data = deconv(data,
out_ch // 2,
filter_size=2,
stride=2,
padding=0,
param_attr=param_attr)
data = fluid.layers.concat([data, short_cut], axis=1)
data = double_conv(data, out_ch)
return data
def fast_scnn(img, num_classes):
size = img.shape[2:]
classifier = Classifier(128, num_classes)
global_feature_extractor = GlobalFeatureExtractor(64, [64, 96, 128], 128,
6, [3, 3, 3])
feature_fusion = FeatureFusionModule(64, 128, 128)
with scope('learning_to_downsample'):
higher_res_features = learning_to_downsample(img, 32, 48, 64)
with scope('global_feature_extractor'):
lower_res_feature = global_feature_extractor.net(higher_res_features)
with scope('feature_fusion'):
x = feature_fusion.net(higher_res_features, lower_res_feature)
with scope('classifier'):
logit = classifier.net(x)
logit = fluid.layers.resize_bilinear(logit, size, align_mode=0)
if len(cfg.MODEL.MULTI_LOSS_WEIGHT) == 3:
with scope('aux_layer_higher'):
higher_logit = aux_layer(higher_res_features, num_classes)
higher_logit = fluid.layers.resize_bilinear(higher_logit,
size,
align_mode=0)
with scope('aux_layer_lower'):
lower_logit = aux_layer(lower_res_feature, num_classes)
lower_logit = fluid.layers.resize_bilinear(lower_logit,
size,
align_mode=0)
return logit, higher_logit, lower_logit
elif len(cfg.MODEL.MULTI_LOSS_WEIGHT) == 2:
with scope('aux_layer_higher'):
higher_logit = aux_layer(higher_res_features, num_classes)
higher_logit = fluid.layers.resize_bilinear(higher_logit,
size,
align_mode=0)
return logit, higher_logit
return logit
def encode(data):
# 编码器设置
short_cuts = []
with scope("encode"):
with scope("block1"):
data = double_conv(data, 64)
short_cuts.append(data)
with scope("block2"):
data = down(data, 128)
short_cuts.append(data)
with scope("block3"):
data = down(data, 256)
short_cuts.append(data)
with scope("block4"):
data = down(data, 512)
short_cuts.append(data)
with scope("block5"):
data = down(data, 512)
return data, short_cuts
