def test_resize_nearest(self):
program = Program()
with program_guard(program):
x = layers.data(name='x', shape=[3, 9, 6], dtype="float32")
output = layers.resize_nearest(x, out_shape=[12, 12])
self.assertIsNotNone(output)
output = layers.resize_nearest(x, scale=3)
self.assertIsNotNone(output)
print(str(program))
def forward(self, inputs):
assert len(inputs) == len(self.in_channels)
outs = []
out = inputs[-1]
outs.append(out)
for i in range(self.num_ins):
out = L.resize_nearest(out, scale=2, align_corners=False)
out = L.pad2d(out, [0, 1, 0, 1])
out = self.conv2x2[i](out)
if i < 4:
out = L.concat([out, inputs[-i - 2]], axis=1)
identity = self.conv1x1[i](out)
out = self.deres_layers[i](out) + identity
outs.append(out)
outs[-1] = L.tanh(outs[-1])
return tuple(outs)
def forward(self, input):
y = layers.resize_nearest(input, scale=self.scale_factor)
return y
def YOLOv3(inputs, initial_filters, num_classes, is_test, trainable=True):
i32 = initial_filters
i64 = i32 * 2
i128 = i32 * 4
i256 = i32 * 8
i512 = i32 * 16
i1024 = i32 * 32
''' darknet53部分，所有卷积层都没有偏移bias_attr=False '''
x = conv2d_unit(inputs,
i32, (3, 3),
stride=1,
padding=1,
name='conv01',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i64, (3, 3),
stride=2,
padding=1,
name='conv02',
is_test=is_test,
trainable=trainable)
x = stack_residual_block(x,
i32,
n=1,
conv_start_idx=3,
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i128, (3, 3),
stride=2,
padding=1,
name='conv05',
is_test=is_test,
trainable=trainable)
x = stack_residual_block(x,
i64,
n=2,
conv_start_idx=6,
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i256, (3, 3),
stride=2,
padding=1,
name='conv10',
is_test=is_test,
trainable=trainable)
act11 = stack_residual_block(x,
i128,
n=8,
conv_start_idx=11,
is_test=is_test,
trainable=trainable)
x = conv2d_unit(act11,
i512, (3, 3),
stride=2,
padding=1,
name='conv27',
is_test=is_test,
trainable=trainable)
act19 = stack_residual_block(x,
i256,
n=8,
conv_start_idx=28,
is_test=is_test,
trainable=trainable)
x = conv2d_unit(act19,
i1024, (3, 3),
stride=2,
padding=1,
name='conv44',
is_test=is_test,
trainable=trainable)
act23 = stack_residual_block(x,
i512,
n=4,
conv_start_idx=45,
is_test=is_test,
trainable=trainable)
''' darknet53部分结束，余下部分不再有残差块stack_residual_block() '''
x = conv2d_unit(act23,
i512, (1, 1),
stride=1,
padding=0,
name='conv53',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i1024, (3, 3),
stride=1,
padding=1,
name='conv54',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i512, (1, 1),
stride=1,
padding=0,
name='conv55',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i1024, (3, 3),
stride=1,
padding=1,
name='conv56',
is_test=is_test,
trainable=trainable)
lkrelu57 = conv2d_unit(x,
i512, (1, 1),
stride=1,
padding=0,
name='conv57',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(lkrelu57,
i1024, (3, 3),
stride=1,
padding=1,
name='conv58',
is_test=is_test,
trainable=trainable)
y1 = P.conv2d(
x,
3 * (num_classes + 5),
filter_size=(1, 1),
param_attr=ParamAttr(initializer=fluid.initializer.Normal(0.0, 0.01),
name="conv59.conv.weights"),
bias_attr=ParamAttr(initializer=fluid.initializer.Constant(0.0),
name="conv59.conv.bias"))
x = conv2d_unit(lkrelu57,
i256, (1, 1),
stride=1,
padding=0,
name='conv60',
is_test=is_test,
trainable=trainable)
x = P.resize_nearest(x, scale=float(2))
x = P.concat([x, act19], axis=1)
x = conv2d_unit(x,
i256, (1, 1),
stride=1,
padding=0,
name='conv61',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i512, (3, 3),
stride=1,
padding=1,
name='conv62',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i256, (1, 1),
stride=1,
padding=0,
name='conv63',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i512, (3, 3),
stride=1,
padding=1,
name='conv64',
is_test=is_test,
trainable=trainable)
lkrelu64 = conv2d_unit(x,
i256, (1, 1),
stride=1,
padding=0,
name='conv65',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(lkrelu64,
i512, (3, 3),
stride=1,
padding=1,
name='conv66',
is_test=is_test,
trainable=trainable)
y2 = P.conv2d(
x,
3 * (num_classes + 5),
filter_size=(1, 1),
param_attr=ParamAttr(initializer=fluid.initializer.Normal(0.0, 0.01),
name="conv67.conv.weights"),
bias_attr=ParamAttr(initializer=fluid.initializer.Constant(0.0),
name="conv67.conv.bias"))
x = conv2d_unit(lkrelu64,
i128, (1, 1),
stride=1,
padding=0,
name='conv68',
is_test=is_test,
trainable=trainable)
x = P.resize_nearest(x, scale=float(2))
x = P.concat([x, act11], axis=1)
x = conv2d_unit(x,
i128, (1, 1),
stride=1,
padding=0,
name='conv69',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i256, (3, 3),
stride=1,
padding=1,
name='conv70',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i128, (1, 1),
stride=1,
padding=0,
name='conv71',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i256, (3, 3),
stride=1,
padding=1,
name='conv72',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i128, (1, 1),
stride=1,
padding=0,
name='conv73',
is_test=is_test,
trainable=trainable)
x = conv2d_unit(x,
i256, (3, 3),
stride=1,
padding=1,
name='conv74',
is_test=is_test,
trainable=trainable)
y3 = P.conv2d(
x,
3 * (num_classes + 5),
filter_size=(1, 1),
param_attr=ParamAttr(initializer=fluid.initializer.Normal(0.0, 0.01),
name="conv75.conv.weights"),
bias_attr=ParamAttr(initializer=fluid.initializer.Constant(0.0),
name="conv75.conv.bias"))
# 相当于numpy的transpose()，交换下标
y1 = P.transpose(y1, perm=[0, 2, 3, 1])
y2 = P.transpose(y2, perm=[0, 2, 3, 1])
y3 = P.transpose(y3, perm=[0, 2, 3, 1])
return y1, y2, y3
def __call__(self, body_feats):
'''
:param body_feats: [s8, s16, s32]
:return:
bs32
|
卷积
|
bs16 [fs32]
| |
卷积 上采样
| |
lateral topdown
\ /
相加
|
bs8 [fs16]
| |
卷积 上采样
| |
lateral topdown
\ /
相加
|
[fs8]
fpn_inner_output = [fs32, fs16, fs8]
然后 fs32, fs16, fs8 分别再接一个卷积得到 p5, p4, p3 ；
p5 接一个卷积得到 p6， p6 接一个卷积得到 p7。
'''
reverse_body_feats = body_feats[::-1] # [s32, s16, s8]
num_backbone_stages = len(reverse_body_feats) # 3
# fpn内部的输出
fpn_inner_output = [None for _ in range(num_backbone_stages)]
body_input = reverse_body_feats[0] # 骨干网络的s32。先接一个卷积
fpn_inner_output[0] = self.convs[0](body_input) # fpn的s32
for i in range(1, num_backbone_stages):
body_input = reverse_body_feats[i] # 骨干网络的s16
top_output = fpn_inner_output[i - 1] # fpn的s32
# 骨干网络的s16卷积，fpn的s32上采样，再融合，融合方式为逐元素相加
lateral = self.convs[i](body_input) # 骨干网络的s16卷积，stride=16
topdown = L.resize_nearest(top_output, scale=float(2)) # fpn的s32上采样，stride=16
fpn_inner_single = lateral + topdown # fpn的s16
fpn_inner_output[i] = fpn_inner_single # fpn的s16
# 第二次卷积
fpn_outputs = [None for _ in range(num_backbone_stages)]
for i in range(num_backbone_stages):
fpn_input = fpn_inner_output[i] # fpn的s32
fpn_output = self.second_convs[i](fpn_input) # fpn的s32
fpn_outputs[i] = fpn_output
# p6p7
if self.use_p6p7:
p6_input = fpn_outputs[0] # p5
p6 = self.p6_conv(p6_input)
p7 = self.p7_conv(p6)
outs = [p7, p6] + fpn_outputs # [p7, p6, p5, p4, p3]
spatial_scale = [1. / 128., 1. / 64., 1. / 32., 1. / 16., 1. / 8.]
return outs, spatial_scale
else:
outs = fpn_outputs # [p5, p4, p3]
spatial_scale = [1. / 32., 1. / 16., 1. / 8.]
return outs, spatial_scale
def forward(self, x):
return L.resize_nearest(x, scale=self.scale)