Skip to content

P
PaddleSeg

PaddlePaddle / PaddleSeg

通知 285
Star 8
Fork 1
P
PaddleSeg
提交 78791df2 编写于 作者: C chenguowei01
浏览文件
操作

add lr_mult_list

上级 5d19dfa6
隐藏空白更改
内联 并排
Showing with 140 addition and 123 deletion
pdseg/models/backbone/resnet_vd.py
浏览文件 @ 78791df2
...@@ -40,11 +40,21 @@ train_parameters = { ...@@ -40,11 +40,21 @@ train_parameters = {
class ResNet(): class ResNet():
def __init__(self, layers=50, scale=1.0, stem=None): def __init__(self,
layers=50,
scale=1.0,
stem=None,
lr_mult_list=[1.0, 1.0, 1.0, 1.0, 1.0]):
self.params = train_parameters self.params = train_parameters
self.layers = layers self.layers = layers
self.scale = scale self.scale = scale
self.stem = stem self.stem = stem
self.lr_mult_list = lr_mult_list
assert len(
self.lr_mult_list
) == 5, "lr_mult_list length in ResNet must be 5 but got {}!!".format(
len(self.lr_mult_list))
self.curr_stage = 0
def net(self, def net(self,
input, input,
...@@ -86,42 +96,37 @@ class ResNet(): ...@@ -86,42 +96,37 @@ class ResNet():
num_filters = [64, 128, 256, 512] num_filters = [64, 128, 256, 512]
if self.stem == 'icnet' or self.stem == 'pspnet' or self.stem == 'deeplab': if self.stem == 'icnet' or self.stem == 'pspnet' or self.stem == 'deeplab':
conv = self.conv_bn_layer( conv = self.conv_bn_layer(input=input,
input=input, num_filters=int(32 * self.scale),
num_filters=int(32 * self.scale), filter_size=3,
filter_size=3, stride=2,
stride=2, act='relu',
act='relu', name="conv1_1")
name="conv1_1") conv = self.conv_bn_layer(input=conv,
conv = self.conv_bn_layer( num_filters=int(32 * self.scale),
input=conv, filter_size=3,
num_filters=int(32 * self.scale), stride=1,
filter_size=3, act='relu',
stride=1, name="conv1_2")
act='relu', conv = self.conv_bn_layer(input=conv,
name="conv1_2") num_filters=int(64 * self.scale),
conv = self.conv_bn_layer( filter_size=3,
input=conv, stride=1,
num_filters=int(64 * self.scale), act='relu',
filter_size=3, name="conv1_3")
stride=1,
act='relu',
name="conv1_3")
else: else:
conv = self.conv_bn_layer( conv = self.conv_bn_layer(input=input,
input=input, num_filters=int(64 * self.scale),
num_filters=int(64 * self.scale), filter_size=7,
filter_size=7, stride=2,
stride=2, act='relu',
act='relu', name="conv1")
name="conv1")
conv = fluid.layers.pool2d(input=conv,
conv = fluid.layers.pool2d( pool_size=3,
input=conv, pool_stride=2,
pool_size=3, pool_padding=1,
pool_stride=2, pool_type='max')
pool_padding=1,
pool_type='max')
layer_count = 1 layer_count = 1
if check_points(layer_count, decode_points): if check_points(layer_count, decode_points):
...@@ -132,6 +137,7 @@ class ResNet(): ...@@ -132,6 +137,7 @@ class ResNet():
if layers >= 50: if layers >= 50:
for block in range(len(depth)): for block in range(len(depth)):
self.curr_stage += 1
for i in range(depth[block]): for i in range(depth[block]):
if layers in [101, 152] and block == 2: if layers in [101, 152] and block == 2:
if i == 0: if i == 0:
...@@ -164,8 +170,10 @@ class ResNet(): ...@@ -164,8 +170,10 @@ class ResNet():
np.ceil( np.ceil(
np.array(conv.shape[2:]).astype('int32') / 2)) np.array(conv.shape[2:]).astype('int32') / 2))
pool = fluid.layers.pool2d( pool = fluid.layers.pool2d(input=conv,
input=conv, pool_size=7, pool_type='avg', global_pooling=True) pool_size=7,
pool_type='avg',
global_pooling=True)
stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0) stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
out = fluid.layers.fc( out = fluid.layers.fc(
input=pool, input=pool,
...@@ -174,6 +182,7 @@ class ResNet(): ...@@ -174,6 +182,7 @@ class ResNet():
initializer=fluid.initializer.Uniform(-stdv, stdv))) initializer=fluid.initializer.Uniform(-stdv, stdv)))
else: else:
for block in range(len(depth)): for block in range(len(depth)):
self.curr_stage += 1
for i in range(depth[block]): for i in range(depth[block]):
conv_name = "res" + str(block + 2) + chr(97 + i) conv_name = "res" + str(block + 2) + chr(97 + i)
conv = self.basic_block( conv = self.basic_block(
...@@ -189,8 +198,10 @@ class ResNet(): ...@@ -189,8 +198,10 @@ class ResNet():
if check_points(layer_count, end_points): if check_points(layer_count, end_points):
return conv, decode_ends return conv, decode_ends
pool = fluid.layers.pool2d( pool = fluid.layers.pool2d(input=conv,
input=conv, pool_size=7, pool_type='avg', global_pooling=True) pool_size=7,
pool_type='avg',
global_pooling=True)
stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0) stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
out = fluid.layers.fc( out = fluid.layers.fc(
input=pool, input=pool,
...@@ -217,23 +228,25 @@ class ResNet(): ...@@ -217,23 +228,25 @@ class ResNet():
act=None, act=None,
name=None): name=None):
lr_mult = self.lr_mult_list[self.curr_stage]
if self.stem == 'pspnet': if self.stem == 'pspnet':
bias_attr = ParamAttr(name=name + "_biases") bias_attr = ParamAttr(name=name + "_biases")
else: else:
bias_attr = False bias_attr = False
conv = fluid.layers.conv2d( conv = fluid.layers.conv2d(input=input,
input=input, num_filters=num_filters,
num_filters=num_filters, filter_size=filter_size,
filter_size=filter_size, stride=stride,
stride=stride, padding=(filter_size - 1) //
padding=(filter_size - 1) // 2 if dilation == 1 else 0, 2 if dilation == 1 else 0,
dilation=dilation, dilation=dilation,
groups=groups, groups=groups,
act=None, act=None,
param_attr=ParamAttr(name=name + "_weights"), param_attr=ParamAttr(name=name + "_weights",
bias_attr=bias_attr, learning_rate=lr_mult),
name=name + '.conv2d.output.1') bias_attr=bias_attr,
name=name + '.conv2d.output.1')
if name == "conv1": if name == "conv1":
bn_name = "bn_" + name bn_name = "bn_" + name
...@@ -243,8 +256,9 @@ class ResNet(): ...@@ -243,8 +256,9 @@ class ResNet():
input=conv, input=conv,
act=act, act=act,
name=bn_name + '.output.1', name=bn_name + '.output.1',
param_attr=ParamAttr(name=bn_name + '_scale'), param_attr=ParamAttr(name=bn_name + '_scale',
bias_attr=ParamAttr(bn_name + '_offset'), learning_rate=lr_mult),
bias_attr=ParamAttr(bn_name + '_offset', learning_rate=lr_mult),
moving_mean_name=bn_name + '_mean', moving_mean_name=bn_name + '_mean',
moving_variance_name=bn_name + '_variance', moving_variance_name=bn_name + '_variance',
) )
...@@ -257,24 +271,24 @@ class ResNet(): ...@@ -257,24 +271,24 @@ class ResNet():
groups=1, groups=1,
act=None, act=None,
name=None): name=None):
pool = fluid.layers.pool2d( lr_mult = self.lr_mult_list[self.curr_stage]
input=input, pool = fluid.layers.pool2d(input=input,
pool_size=2, pool_size=2,
pool_stride=2, pool_stride=2,
pool_padding=0, pool_padding=0,
pool_type='avg', pool_type='avg',
ceil_mode=True) ceil_mode=True)
conv = fluid.layers.conv2d( conv = fluid.layers.conv2d(input=pool,
input=pool, num_filters=num_filters,
num_filters=num_filters, filter_size=filter_size,
filter_size=filter_size, stride=1,
stride=1, padding=(filter_size - 1) // 2,
padding=(filter_size - 1) // 2, groups=groups,
groups=groups, act=None,
act=None, param_attr=ParamAttr(name=name + "_weights",
param_attr=ParamAttr(name=name + "_weights"), learning_rate=lr_mult),
bias_attr=False) bias_attr=False)
if name == "conv1": if name == "conv1":
bn_name = "bn_" + name bn_name = "bn_" + name
else: else:
...@@ -282,8 +296,9 @@ class ResNet(): ...@@ -282,8 +296,9 @@ class ResNet():
return fluid.layers.batch_norm( return fluid.layers.batch_norm(
input=conv, input=conv,
act=act, act=act,
param_attr=ParamAttr(name=bn_name + '_scale'), param_attr=ParamAttr(name=bn_name + '_scale',
bias_attr=ParamAttr(bn_name + '_offset'), learning_rate=lr_mult),
bias_attr=ParamAttr(bn_name + '_offset', learning_rate=lr_mult),
moving_mean_name=bn_name + '_mean', moving_mean_name=bn_name + '_mean',
moving_variance_name=bn_name + '_variance') moving_variance_name=bn_name + '_variance')
...@@ -294,8 +309,11 @@ class ResNet(): ...@@ -294,8 +309,11 @@ class ResNet():
if is_first or stride == 1: if is_first or stride == 1:
return self.conv_bn_layer(input, ch_out, 1, stride, name=name) return self.conv_bn_layer(input, ch_out, 1, stride, name=name)
else: else:
return self.conv_bn_layer_new( return self.conv_bn_layer_new(input,
input, ch_out, 1, stride, name=name) ch_out,
1,
stride,
name=name)
elif is_first: elif is_first:
return self.conv_bn_layer(input, ch_out, 1, stride, name=name) return self.conv_bn_layer(input, ch_out, 1, stride, name=name)
else: else:
...@@ -308,60 +326,59 @@ class ResNet(): ...@@ -308,60 +326,59 @@ class ResNet():
name, name,
is_first=False, is_first=False,
dilation=1): dilation=1):
conv0 = self.conv_bn_layer( conv0 = self.conv_bn_layer(input=input,
input=input, num_filters=num_filters,
num_filters=num_filters, filter_size=1,
filter_size=1, dilation=1,
dilation=1, stride=1,
stride=1, act='relu',
act='relu', name=name + "_branch2a")
name=name + "_branch2a")
if dilation > 1: if dilation > 1:
conv0 = self.zero_padding(conv0, dilation) conv0 = self.zero_padding(conv0, dilation)
conv1 = self.conv_bn_layer( conv1 = self.conv_bn_layer(input=conv0,
input=conv0, num_filters=num_filters,
num_filters=num_filters, filter_size=3,
filter_size=3, dilation=dilation,
dilation=dilation, stride=stride,
stride=stride, act='relu',
act='relu', name=name + "_branch2b")
name=name + "_branch2b") conv2 = self.conv_bn_layer(input=conv1,
conv2 = self.conv_bn_layer( num_filters=num_filters * 4,
input=conv1, dilation=1,
num_filters=num_filters * 4, filter_size=1,
dilation=1, act=None,
filter_size=1, name=name + "_branch2c")
act=None,
name=name + "_branch2c") short = self.shortcut(input,
num_filters * 4,
short = self.shortcut( stride,
input, is_first=is_first,
num_filters * 4, name=name + "_branch1")
stride,
is_first=is_first,
name=name + "_branch1")
print(input.shape, short.shape, conv2.shape) print(input.shape, short.shape, conv2.shape)
print(stride) print(stride)
return fluid.layers.elementwise_add( return fluid.layers.elementwise_add(x=short,
x=short, y=conv2, act='relu', name=name + ".add.output.5") y=conv2,
act='relu',
name=name + ".add.output.5")
def basic_block(self, input, num_filters, stride, is_first, name): def basic_block(self, input, num_filters, stride, is_first, name):
conv0 = self.conv_bn_layer( conv0 = self.conv_bn_layer(input=input,
input=input, num_filters=num_filters,
num_filters=num_filters, filter_size=3,
filter_size=3, act='relu',
act='relu', stride=stride,
stride=stride, name=name + "_branch2a")
name=name + "_branch2a") conv1 = self.conv_bn_layer(input=conv0,
conv1 = self.conv_bn_layer( num_filters=num_filters,
input=conv0, filter_size=3,
num_filters=num_filters, act=None,
filter_size=3, name=name + "_branch2b")
act=None, short = self.shortcut(input,
name=name + "_branch2b") num_filters,
short = self.shortcut( stride,
input, num_filters, stride, is_first, name=name + "_branch1") is_first,
name=name + "_branch1")
return fluid.layers.elementwise_add(x=short, y=conv1, act='relu') return fluid.layers.elementwise_add(x=short, y=conv1, act='relu')
......
Markdown is supported
0% 或 .
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 或