def gru_module(x, num_state, num_node):
'''
Global Reasoning Unit: projection --> graph reasoning --> reverse projection
params:
x: B x C x H x W
num_state: the dimension of each vertex feature
num_node: the number of vertet
output: B x C x H x W
feature trans:
B, C, H, W --> B, N, H, W --> B, N, H*W -->B, N, C1 -->B, C1, N-->B, C1, N-->B, C1, H*W-->B, C, H, W
--> B, C1,H, W -->B, C1,H*W -->B, H*W, C1
'''
# generate B
num_batch, C, H, W = x.shape
with scope('projection'):
B = conv(x,
num_node,
filter_size=1,
bias_attr=True,
name='projection' + '_conv') #num_batch, node, H, W
B = fluid.layers.reshape(
B, shape=[num_batch, num_node,
H * W]) # Projection Matrix: num_batch, node, L=H*W
# reduce dimension
with scope('reduce_channel'):
x_reduce = conv(x,
num_state,
filter_size=1,
bias_attr=True,
name='reduce_channel' +
'_conv') #num_batch, num_state, H, W
x_reduce = fluid.layers.reshape(x_reduce,
shape=[num_batch, num_state, H * W
]) #num_batch, num_state, L
x_reduce = fluid.layers.transpose(x_reduce,
perm=[0, 2,
1]) #num_batch, L, num_state
V = fluid.layers.transpose(fluid.layers.matmul(B, x_reduce),
perm=[0, 2, 1]) #num_batch, num_state, num_node
#L = fluid.layers.fill_constant(shape=[1], value=H*W, dtype='float32')
#V = fluid.layers.elementwise_div(V, L)
new_V = gcn_module('gru' + '_gcn', V, num_node, num_state)
B = fluid.layers.reshape(B, shape=[num_batch, num_node, H * W])
D = fluid.layers.transpose(B, perm=[0, 2, 1])
Y = fluid.layers.matmul(D, fluid.layers.transpose(new_V, perm=[0, 2, 1]))
Y = fluid.layers.transpose(Y, perm=[0, 2, 1])
Y = fluid.layers.reshape(Y, shape=[num_batch, num_state, H, W])
with scope('extend_dim'):
Y = conv(Y,
C,
filter_size=1,
bias_attr=False,
name='extend_dim' + '_conv')
#Y = bn_zero(Y)
Y = bn(Y)
out = fluid.layers.elementwise_add(Y, x)
return out
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)
# params to control your flow
s = self.stride
block_point = self.block_point
output_stride = self.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):
self.short_cuts[block_point] = short_cuts[1]
self.stride = s
self.block_point = block_point
return data
def glore(input, num_classes):
"""
Reference:
Chen, Yunpeng, et al. "Graph-Based Global Reasoning Networks", In CVPR 2019
"""
# Backbone: ResNet
res5, feat_dict = resnet(input)
res4 = feat_dict[91]
# 3x3 Conv. 2048 -> 512
reduce_kernel = 3
if cfg.DATASET.DATASET_NAME == 'cityscapes':
reduce_kernel = 1
with scope('feature'):
feature = conv(res5,
512,
filter_size=reduce_kernel,
bias_attr=False,
name='feature_conv')
feature = bn(feature, act='relu')
# GRU Module
gru_output = gru_module(feature, num_state=128, num_node=64)
dropout = fluid.layers.dropout(gru_output,
dropout_prob=0.1,
name="dropout")
logit = get_logit_interp(dropout, num_classes, input.shape[2:])
if cfg.MODEL.GLORE.AuxHead:
aux_logit = FCNHead(res4, 256, num_classes, input.shape[2:])
return logit, aux_logit
return logit
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 get_logit_interp(input, num_classes, out_shape, name="logit"):
# 1x1_Conv
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 PSPHead(input, out_features, num_classes, output_shape):
# Arch of Pyramid Scene Parsing Module:
#
# |----> Pool_1x1 + Conv_1x1 + BN + ReLU + bilinear_interp-------->|————————|
# | | |
# |----> Pool_2x2 + Conv_1x1 + BN + ReLU + bilinear_interp-------->| |
# x ------>| | concat |----> Conv_3x3 + BN + ReLU -->Dropout --> Conv_1x1
# | |----> Pool_3x3 + Conv_1x1 + BN + ReLU + bilinear_interp-------->| |
# | | | |
# | |----> Pool_6x6 + Conv_1x1 + BN + ReLU + bilinear_interp-------->|________|
# | ^
# |——————————————————————————————————————————————————————————————————————————————|
#
cat_layers = []
sizes = (1, 2, 3, 6)
# 4 parallel pooling branches
for size in sizes:
psp_name = "psp" + str(size)
with scope(psp_name):
pool_feat = fluid.layers.adaptive_pool2d(input,
pool_size=[size, size],
pool_type='avg',
name=psp_name +
'_adapool')
conv_feat = conv(pool_feat,
out_features,
filter_size=1,
bias_attr=True,
name=psp_name + '_conv')
bn_feat = bn(conv_feat, act='relu')
interp = fluid.layers.resize_bilinear(bn_feat,
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')
# Conv_3x3 + BN + ReLU
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')
# Dropout
dropout_out = fluid.layers.dropout(out, dropout_prob=0.1, name="dropout")
# Conv_1x1 + bilinear_upsample
seg_name = "logit"
with scope(seg_name):
param_attr = fluid.ParamAttr(
name=seg_name + '_weights',
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.0),
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.01))
logit = conv(dropout_out,
num_classes,
filter_size=1,
param_attr=param_attr,
bias_attr=True,
name=seg_name + '_conv')
logit_interp = fluid.layers.resize_bilinear(logit,
out_shape=output_shape,
name=seg_name + '_interp')
return logit_interp
def ASPPHead(input, mid_channel, num_classes, output_shape):
# Arch of Atorus Spatial Pyramid Pooling Module:
#
# |----> ImagePool + Conv_1x1 + BN + ReLU + bilinear_interp-------->|————————|
# | | |
# |----> Conv_1x1 + BN + ReLU -------->| |
# | | |
# x----->|----> AtrousConv_3x3 + BN + ReLU -------->| concat |----> Conv_1x1 + BN + ReLU -->Dropout --> Conv_1x1
# | | |
# |----> AtrousConv_3x3 + BN + ReLU -------->| |
# | | |
# |----> AtorusConv_3x3 + BN + ReLU -------->|________|
#
#
if cfg.MODEL.BACKBONE_OUTPUT_STRIDE == 16:
aspp_ratios = [6, 12, 18]
elif cfg.MODEL.BACKBONE_OUTPUT_STRIDE == 8:
aspp_ratios = [12, 24, 36]
else:
raise Exception("deeplab only support stride 8 or 16")
param_attr = fluid.ParamAttr(name=name_scope + 'weights', regularizer=None,
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.06))
with scope('ASPPHead'):
with scope("image_pool"):
image_avg = fluid.layers.reduce_mean( input, [2, 3], keep_dim=True)
image_avg = bn_relu( conv( image_avg, mid_channel, 1, 1, groups=1, padding=0, param_attr=param_attr))
image_avg = fluid.layers.resize_bilinear(image_avg, input.shape[2:])
with scope("aspp0"):
aspp0 = bn_relu( conv( input, mid_channel, 1, 1, groups=1, padding=0, param_attr=param_attr))
with scope("aspp1"):
if cfg.MODEL.DEEPLAB.ASPP_WITH_SEP_CONV:
aspp1 = separate_conv( input, mid_channel, 1, 3, dilation=aspp_ratios[0], act=relu)
else:
aspp1 = bn_relu( conv( input, mid_channel, stride=1, filter_size=3, dilation=aspp_ratios[0],
padding=aspp_ratios[0], param_attr=param_attr))
with scope("aspp2"):
if cfg.MODEL.DEEPLAB.ASPP_WITH_SEP_CONV:
aspp2 = separate_conv( input, mid_channel, 1, 3, dilation=aspp_ratios[1], act=relu)
else:
aspp2 = bn_relu( conv( input, mid_channel, stride=1, filter_size=3, dilation=aspp_ratios[1],
padding=aspp_ratios[1], param_attr=param_attr))
with scope("aspp3"):
if cfg.MODEL.DEEPLAB.ASPP_WITH_SEP_CONV:
aspp3 = separate_conv( input, mid_channel, 1, 3, dilation=aspp_ratios[2], act=relu)
else:
aspp3 = bn_relu( conv( input, mid_channel, stride=1, filter_size=3, dilation=aspp_ratios[2],
padding=aspp_ratios[2], param_attr=param_attr))
with scope("concat"):
feat = fluid.layers.concat([image_avg, aspp0, aspp1, aspp2, aspp3], axis=1)
feat = bn_relu( conv( feat, 2*mid_channel, 1, 1, groups=1, padding=0, param_attr=param_attr))
feat = fluid.layers.dropout(feat, 0.1)
# Conv_1x1 + bilinear_upsample
seg_name = "logit"
with scope(seg_name):
param_attr = fluid.ParamAttr( name= seg_name+'_weights',
regularizer=fluid.regularizer.L2DecayRegularizer(regularization_coeff=0.0),
initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.01))
logit = conv(feat, num_classes, filter_size=1, param_attr=param_attr, bias_attr=True, name=seg_name+'_conv')
logit_interp = fluid.layers.resize_bilinear(logit, out_shape=output_shape, name=seg_name+'_interp')
return logit_interp