def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False):
# ch_in, ch_out, kernel, stride, groups, expansion, shortcut
super(CrossConv, self).__init__()
c_ = int(c2 * e) # hidden channels
self.cv1 = Conv(c1, c_, (1, k), (1, s))
self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g)
self.add = shortcut and c1 == c2
def __init__(self,
c1,
c2,
k=1,
s=1,
g=1,
act=True): # ch_in, ch_out, kernel, stride, groups
super(GhostConv, self).__init__()
c_ = c2 // 2 # hidden channels
self.cv1 = Conv(c1, c_, k, s, None, g, act)
self.cv2 = Conv(c_, c_, 5, 1, None, c_, act)
def __init__(self, c1, c2, k=3, s=1): # ch_in, ch_out, kernel, stride
super(GhostBottleneck, self).__init__()
c_ = c2 // 2
self.conv = nn.Sequential(
GhostConv(c1, c_, 1, 1), # pw
DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(), # dw
GhostConv(c_, c2, 1, 1, act=False)) # pw-linear
self.shortcut = nn.Sequential(DWConv(
c1, c1, k, s, act=False), Conv(
c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity()
def __init__(self, P3_size=512, P4_size=256, P5_size=512, inner_p3=256, inner_p4=512, inner_p5=1024, version='L'):
super(PAN, self).__init__()
self.version = str(version)
gains = {'s': {'gd': 0.33, 'gw': 0.5},
'm': {'gd': 0.67, 'gw': 0.75},
'l': {'gd': 1, 'gw': 1},
'x': {'gd': 1.33, 'gw': 1.25}}
if self.version.lower() in gains:
# only for yolov5
self.gd = gains[self.version.lower()]['gd'] # depth gain
self.gw = gains[self.version.lower()]['gw'] # width gain
else:
self.gd = 1
self.gw = 1
self.channels_out = {
'inner_p3': inner_p3,
'inner_p4': inner_p4,
'inner_p5': inner_p5
}
self.re_channels_out()
self.P3_size = P3_size
self.P4_size = P4_size
self.P5_size = P5_size
self.inner_p3 = self.channels_out['inner_p3']
self.inner_p4 = self.channels_out['inner_p4']
self.inner_p5 = self.channels_out['inner_p5']
self.P3 = C3(self.P3_size, self.inner_p3, self.get_depth(3), False)
self.convP3 = Conv(self.inner_p3, self.inner_p3, 3, 2)
self.P4 = C3(self.P4_size + self.inner_p3, self.inner_p4, self.get_depth(3), False)
self.convP4 = Conv(self.inner_p4, self.inner_p4, 3, 2)
self.P5 = C3(self.inner_p4 + P5_size, self.inner_p5, self.get_depth(3), False)
self.concat = Concat()
self.out_shape = (self.inner_p3, self.inner_p4, self.inner_p5)
print("PAN input channel size: P3 {}, P4 {}, P5 {}".format(self.P3_size, self.P4_size, self.P5_size))
print("PAN output channel size: PP3 {}, PP4 {}, PP5 {}".format(self.inner_p3, self.inner_p4, self.inner_p5))
def __init__(self, C3_size=256, C4_size=512, C5_size=512, inner_p4=512, outer_p4=256, version='S'):
super(PyramidFeatures, self).__init__()
self.C3_size = C3_size
self.C4_size = C4_size
self.C5_size = C5_size
self.version = version
gains = {'s': {'gd': 0.33, 'gw': 0.5},
'm': {'gd': 0.67, 'gw': 0.75},
'l': {'gd': 1, 'gw': 1},
'x': {'gd': 1.33, 'gw': 1.25}}
if self.version.lower() in gains:
# only for yolov5
self.gd = gains[self.version.lower()]['gd'] # depth gain
self.gw = gains[self.version.lower()]['gw'] # width gain
else:
self.gd = 0.33
self.gw = 0.5
self.channels_out = {
'inner_p4': inner_p4,
'outer_p4': outer_p4,
}
self.re_channels_out()
self.concat = Concat()
self.P5_upsampled = nn.Upsample(scale_factor=2, mode='nearest')
self.P4_1 = C3(self.C5_size + self.C4_size, self.channels_out['inner_p4'], self.get_depth(3), False)
self.P4_upsampled = nn.Upsample(scale_factor=2, mode='nearest')
self.P4_2 = Conv(self.channels_out['inner_p4'], self.channels_out['outer_p4'], 1, 1)
self.out_shape = {'P3_size': self.C3_size + self.channels_out['outer_p4'],
'P4_size': self.channels_out['outer_p4'],
'P5_size': self.C5_size}
print("FPN input channel size: C3 {}, C4 {}, C5 {}".format(self.C3_size, self.C4_size, self.C5_size))
print("FPN output channel size: P3 {}, P4 {}, P5 {}".format(self.C3_size + self.channels_out['outer_p4'], self.channels_out['outer_p4'],
self.C5_size))
def __init__(self, focus=True, version='L', with_C3TR=False):
super(YOLOv5, self).__init__()
self.version = version
self.with_focus = focus
self.with_c3tr = with_C3TR
gains = {
's': {
'gd': 0.33,
'gw': 0.5
},
'm': {
'gd': 0.67,
'gw': 0.75
},
'l': {
'gd': 1,
'gw': 1
},
'x': {
'gd': 1.33,
'gw': 1.25
}
}
self.gd = gains[self.version.lower()]['gd'] # depth gain
self.gw = gains[self.version.lower()]['gw'] # width gain
self.channels_out = {
'stage1': 64,
'stage2_1': 128,
'stage2_2': 128,
'stage3_1': 256,
'stage3_2': 256,
'stage4_1': 512,
'stage4_2': 512,
'stage5': 1024,
'spp': 1024,
'csp1': 1024,
'conv1': 512
}
self.re_channels_out()
if self.with_focus:
self.stage1 = Focus(3, self.channels_out['stage1'])
else:
self.stage1 = Conv(3, self.channels_out['stage1'], 3, 2)
# for latest yolov5, you can change BottleneckCSP to C3
self.stage2_1 = Conv(self.channels_out['stage1'],
self.channels_out['stage2_1'],
k=3,
s=2)
self.stage2_2 = C3(self.channels_out['stage2_1'],
self.channels_out['stage2_2'], self.get_depth(3))
self.stage3_1 = Conv(self.channels_out['stage2_2'],
self.channels_out['stage3_1'], 3, 2)
self.stage3_2 = C3(self.channels_out['stage3_1'],
self.channels_out['stage3_2'], self.get_depth(9))
self.stage4_1 = Conv(self.channels_out['stage3_2'],
self.channels_out['stage4_1'], 3, 2)
self.stage4_2 = C3(self.channels_out['stage4_1'],
self.channels_out['stage4_2'], self.get_depth(9))
self.stage5 = Conv(self.channels_out['stage4_2'],
self.channels_out['stage5'], 3, 2)
self.spp = SPP(self.channels_out['stage5'], self.channels_out['spp'],
[5, 9, 13])
if self.with_c3tr:
self.c3tr = C3TR(self.channels_out['spp'],
self.channels_out['csp1'], self.get_depth(3),
False)
else:
self.csp1 = C3(self.channels_out['spp'], self.channels_out['csp1'],
self.get_depth(3), False)
self.conv1 = Conv(self.channels_out['csp1'],
self.channels_out['conv1'], 1, 1)
self.out_shape = {
'C3_size': self.channels_out['stage3_2'],
'C4_size': self.channels_out['stage4_2'],
'C5_size': self.channels_out['conv1']
}
print("backbone output channel: C3 {}, C4 {}, C5 {}".format(
self.channels_out['stage3_2'], self.channels_out['stage4_2'],
self.channels_out['conv1']))