def forward(self, x):
# x_1 = x[:, :self.inp, :, :]
# x_2 = x[:, self.inp:, :, :]
b, c, h, w = x.value
x_1 = flops_counter.TensorSize([b, self.inp, h, w])
x_2 = flops_counter.TensorSize([b, c - self.inp, h, w])
a_1 = self.conv1x1_1(x_1)
# g_1 = F.sigmoid(self.gate_1(x_1))
g_1 = self.sigmoid1(self.gate_1(x_1))
a_2 = self.conv1x1_2(x_2)
# g_2 = F.sigmoid(self.gate_2(x_2))
g_2 = self.sigmoid2(self.gate_2(x_2))
ret = flops_counter.cat(
(self.eltmul1(a_1, g_1), self.eltmul2(a_2, g_2)), 1)
return ret
def _calc_flops(self, img, scale=1, flip=False):
x = img
h, w = x.shape[:2]
scale_h, scale_w = scale, scale
if self.max_downsample > 1:
h_new = int(
np.ceil(scale * h / self.max_downsample) * self.max_downsample)
w_new = int(
np.ceil(scale * w / self.max_downsample) * self.max_downsample)
scale_h = h_new / h
scale_w = w_new / w
x = cv2.resize(img,
None,
None,
fx=scale_w,
fy=scale_h,
interpolation=cv2.INTER_LINEAR)
elif scale != 1:
x = cv2.resize(img,
None,
None,
fx=scale_w,
fy=scale_h,
interpolation=cv2.INTER_LINEAR)
h, w, c = x.shape
if 'hr' in self.model.name.lower():
if scale > 1 and (h > 5000 or w > 5000):
return 0
x = flops_counter.TensorSize([1, c, h, w])
y = self.model(x)
flops = self.model.flops * 2 if flip else self.model.flops
self.model.set_flops_zero()
return flops
import sys
sys.path.append('.')
import flops_counter as fc
from models import YuFaceDetectNet
model = YuFaceDetectNet()
x = fc.TensorSize([1, 3, 224, 210])
y = model(x)
print(y, model.flops)
# print(model)
def forward(self, x):
conv1, pool1 = self.layer1(x)
conv2, pool2 = self.layer2(pool1)
conv3, pool3 = self.layer3(pool2)
conv4, pool4 = self.layer4(pool3)
conv5, pool5 = self.layer5(pool4)
conv6 = self.layer6(pool5)
conv7 = self.layer7(conv6)
conv8 = self.layer8(conv7)
lfpn2_on_conv5 = self.lfpn2_on_conv5(conv6, conv5)
lfpn1_on_conv4 = self.lfpn1_on_conv4(lfpn2_on_conv5, conv4)
lfpn0_on_conv3 = self.lfpn0_on_conv3(lfpn1_on_conv4, conv3)
ssh_conv3 = self.ssh_conv3(lfpn0_on_conv3)
ssh_conv4 = self.ssh_conv4(lfpn1_on_conv4)
ssh_conv5 = self.ssh_conv5(lfpn2_on_conv5)
ssh_conv6 = self.ssh_conv6(conv6)
ssh_conv7 = self.ssh_conv7(conv7)
ssh_conv8 = self.ssh_conv8(conv8)
ssh_conv3_l2norm = self.ssh_conv3_l2norm(ssh_conv3)
ssh_conv4_l2norm = self.ssh_conv4_l2norm(ssh_conv4)
ssh_conv5_l2norm = self.ssh_conv5_l2norm(ssh_conv5)
inputs = [
ssh_conv3_l2norm, ssh_conv4_l2norm, ssh_conv5_l2norm, ssh_conv6,
ssh_conv7, ssh_conv8
]
face_confs = []
head_confs = []
for i, feat in enumerate(inputs):
mbox_loc = self.mbox_loc[i](feat)
# print(mbox_loc)
if i == 0:
temp_conf = self.mbox_conf[i](feat)
# face_conf3 = temp_conf[:, 0:3, :, :]
face_conf3 = [x for x in temp_conf.value]
face_conf3[1] = 3
face_conf3 = flops_counter.TensorSize(face_conf3)
# face_conf1 = temp_conf[:, 3:4, :, :]
face_conf1 = [x for x in temp_conf.value]
face_conf1[1] = 1
face_conf1 = flops_counter.TensorSize(face_conf1)
# head_conf3 = temp_conf[:, 4:7, :, :]
head_conf3 = [x for x in temp_conf.value]
head_conf3[1] = 3
head_conf3 = flops_counter.TensorSize(head_conf3)
# head_conf1 = temp_conf[:, 7:, :, :]
head_conf1 = [x for x in temp_conf.value]
head_conf1[1] = 1
head_conf1 = flops_counter.TensorSize(head_conf1)
# face conf
face_conf3_maxin = face_conf3.max(1, keepdim=True)
face_confs.append(
flops_counter.cat([face_conf3_maxin, face_conf1],
1).permute(0, 2, 3, 1))
# head conf
head_conf3_maxin = head_conf3.max(1, keepdim=True)
head_confs.append(
flops_counter.cat([head_conf3_maxin, head_conf1],
1).permute(0, 2, 3, 1))
else:
temp_conf = self.mbox_conf[i](feat)
# face_conf1 = temp_conf[:, 0:1, :, :]
face_conf1 = [x for x in temp_conf.value]
face_conf1[1] = 1
face_conf1 = flops_counter.TensorSize(face_conf1)
# face_conf3 = temp_conf[:, 1:4, :, :]
face_conf3 = [x for x in temp_conf.value]
face_conf3[1] = 3
face_conf3 = flops_counter.TensorSize(face_conf3)
# head_conf = temp_conf[:, 4:, :, :]
head_conf1 = [x for x in temp_conf.value]
head_conf1[1] = 4
head_conf1 = flops_counter.TensorSize(head_conf1)
# face conf
face_conf3_maxin = face_conf3.max(1, keepdim=True)
face_confs.append(
flops_counter.cat([face_conf1, face_conf3_maxin],
1).permute(0, 2, 3, 1))
# head conf
head_confs.append(head_conf1.permute(0, 2, 3, 1))
# print(temp_conf)
face_conf = flops_counter.cat(
[o.view(o.value[0], -1) for o in face_confs], 1)
head_conf = flops_counter.cat(
[o.view(o.value[0], -1) for o in head_confs], 1)
face_conf_softmax = self.softmax(
face_conf.view(face_conf.value[0], -1, 2))
return face_conf_softmax
def forward(self, x):
loc = list()
conf = list()
######
# Backbone
######
conv3_3_x = self.layer1(x)
conv4_3_x = self.layer2(conv3_3_x)
conv5_3_x = self.layer3(conv4_3_x)
fc7_x = self.layer4(conv5_3_x)
conv6_2_x = self.layer5(fc7_x)
conv7_2_x = self.layer6(conv6_2_x)
######
# dsfd specific layers
######
# fpn
lfpn3_fc7_x_up = self.upsample(self.latlayer3(fc7_x))
lfpn3_conv5_3_x = self.smooth3(conv5_3_x)
if lfpn3_fc7_x_up.value[2] != lfpn3_conv5_3_x.value[
2] or lfpn3_fc7_x_up.value[3] != lfpn3_conv5_3_x.value[3]:
pad = (0, lfpn3_conv5_3_x.value[3] - lfpn3_fc7_x_up.value[3], 0,
lfpn3_conv5_3_x.value[2] - lfpn3_fc7_x_up.value[2])
lfpn3_fc7_x_up = F.pad(lfpn3_fc7_x_up, pad)
lfpn3 = self.eltmul(lfpn3_fc7_x_up, lfpn3_conv5_3_x)
lfpn2_lfpn3_up = self.upsample(self.latlayer2(lfpn3))
lfpn2_conv4_3_x = self.smooth2(conv4_3_x)
if lfpn2_lfpn3_up.value[2] != lfpn2_conv4_3_x.value[
2] or lfpn2_lfpn3_up.value[3] != lfpn2_conv4_3_x.value[3]:
pad = (0, lfpn2_conv4_3_x.value[3] - lfpn2_lfpn3_up.value[3], 0,
lfpn2_conv4_3_x.value[2] - lfpn2_lfpn3_up.value[2])
lfpn2_lfpn3_up = F.pad(lfpn2_lfpn3_up, pad)
lfpn2 = self.eltmul(lfpn2_lfpn3_up, lfpn2_conv4_3_x)
lfpn1_lfpn2_up = self.upsample(self.latlayer1(lfpn2))
lfpn1_conv3_3_x = self.smooth1(conv3_3_x)
if lfpn1_lfpn2_up.value[2] != lfpn1_conv3_3_x.value[
2] or lfpn1_lfpn2_up.value[3] != lfpn1_conv3_3_x.value[3]:
pad = (0, lfpn1_conv3_3_x.value[3] - lfpn1_lfpn2_up.value[3], 0,
lfpn1_conv3_3_x.value[2] - lfpn1_lfpn2_up.value[2])
lfpn1_lfpn2_up = F.pad(lfpn1_lfpn2_up, pad)
lfpn1 = self.eltmul(lfpn1_lfpn2_up, lfpn1_conv3_3_x)
conv5_3_x = lfpn3
conv4_3_x = lfpn2
conv3_3_x = lfpn1
# fem
sources = [
conv3_3_x, conv4_3_x, conv5_3_x, fc7_x, conv6_2_x, conv7_2_x
]
sources[0] = self.cpm3_3(sources[0])
sources[1] = self.cpm4_3(sources[1])
sources[2] = self.cpm5_3(sources[2])
sources[3] = self.cpm7(sources[3])
sources[4] = self.cpm6_2(sources[4])
sources[5] = self.cpm7_2(sources[5])
# apply multibox head to source layers
loc = list()
conf = list()
for x, l, c in zip(sources, self.loc, self.conf):
# l(x)
loc.append(l(x).permute(0, 2, 3, 1))
# mio: max_in_out
conf.append(c(x).permute(0, 2, 3, 1))
# face_conf = flops_counter.cat([flops_counter.view([o[1], o[2], 2], (1, -1)) for o in conf], 1)
# output = self.softmax(flops_counter.view(face_conf, (1, -1, 2)))
face_confs = list()
for o in conf:
dst = [i for i in o.value]
dst[-1] = 2
face_confs.append(flops_counter.TensorSize(dst))
face_conf = flops_counter.cat(
[o.view(o.value[0], -1) for o in face_confs], 1)
output = self.softmax(face_conf.view(face_conf.value[0], -1, 2))
return output
def forward(self, x):
out = x
# get conv3_3
for k in range(16):
out = self.vgg16[k](out)
conv3_3 = out # channels = 256
conv3_3_norm = self.l2norm_conv3_3(conv3_3)
# get conv4_3
for k in range(16, 23):
out = self.vgg16[k](out)
conv4_3 = out # channels = 512
conv4_3_norm = self.l2norm_conv4_3(conv4_3)
# get conv5_3
for k in range(23, 30):
out = self.vgg16[k](out)
conv5_3 = out # channels = 512
conv5_3_norm = self.l2norm_conv5_3(conv5_3)
out = self.vgg16[30](out)
# get conv_fc7
out = self.conv_fc6(out)
out = self.relu_fc6(out)
out = self.conv_fc7(out)
out = self.relu_fc7(out)
conv_fc7 = out
# get conv6_2
out = self.conv6_1(out)
out = self.relu_conv6_1(out)
out = self.conv6_2(out)
out = self.relu_conv6_2(out)
conv6_2 = out
# get conv7_2
out = self.conv7_1(out)
out = self.relu_conv7_1(out)
out = self.conv7_2(out)
out = self.relu_conv7_2(out)
conv7_2 = out
# Detection Head - mbox_loc
mbox_loc_inputs = [
self.mbox_loc_conv3_3_norm(conv3_3_norm),
self.mbox_loc_conv4_3_norm(conv4_3_norm),
self.mbox_loc_conv5_3_norm(conv5_3_norm),
self.mbox_loc_conv_fc7(conv_fc7),
self.mbox_loc_conv6_2(conv6_2),
self.mbox_loc_conv7_2(conv7_2)
]
mbox_loc = flops_counter.cat(
[o.permute(0, 2, 3, 1).view(1, -1, 4) for o in mbox_loc_inputs], 1)
# Detection Head - mbox_conf
mbox_conf_conv3_3_norm = self.mbox_conf_conv3_3_norm(conv3_3_norm)
conf1 = [i for i in mbox_conf_conv3_3_norm.value]
conf1[1] = 1
conf1 = flops_counter.TensorSize(conf1)
conf234 = [i for i in mbox_conf_conv3_3_norm.value]
conf234[1] = 3
conf234 = flops_counter.TensorSize(conf234)
conf234 = conf234.max(1, keepdim=True)
mbox_conf_conv3_3_norm = flops_counter.cat([conf1, conf234], 1)
mbox_conf_inputs = [
mbox_conf_conv3_3_norm,
self.mbox_conf_conv4_3_norm(conv4_3_norm),
self.mbox_conf_conv5_3_norm(conv5_3_norm),
self.mbox_conf_conv_fc7(conv_fc7),
self.mbox_conf_conv6_2(conv6_2),
self.mbox_conf_conv7_2(conv7_2)
]
mbox_conf = flops_counter.cat(
[o.permute(0, 2, 3, 1).view(1, -1, 2) for o in mbox_conf_inputs],
1)
mbox_conf = self.softmax(mbox_conf)
return mbox_loc, mbox_conf
def forward(self, x):
sources = list()
loc = list()
conf = list()
end = 8 if self.model_type == 32 else 6
for k in range(end):
x = self.base[k](x)
s1 = x # (640, 640) -> (320, 320)
for k in range(2, end):
x = self.base[k](x)
s2 = x # (160, 160) -> (80, 80)
for k in range(2, end):
x = self.base[k](x)
s3 = x # (80, 80) -> (40, 40)
for k in range(2, end):
x = self.base[k](x)
s4 = x # (40, 40) -> (20, 20)
for k in range(2, end):
x = self.base[k](x)
s5 = x # (20, 20) -> (10, 10)
for k in range(2, end):
x = self.base[k](x)
s6 = x # (10, 10) -> (5, 5)
sources.append(s6)
# def upsample_add(seq, source, target, up_handle, add_handle):
u1 = upsample_add(self.upfeat[0], s6, s5, self.upsample, self.eltadd)
sources.append(u1)
u2 = upsample_add(self.upfeat[0], u1, s4, self.upsample, self.eltadd)
sources.append(u2)
u3 = upsample_add(self.upfeat[0], u2, s3, self.upsample, self.eltadd)
sources.append(u3)
u4 = upsample_add(self.upfeat[0], u3, s2, self.upsample, self.eltadd)
sources.append(u4)
u5 = upsample_add(self.upfeat[0], u4, s1, self.upsample, self.eltadd)
sources.append(u5)
sources = sources[::-1]
loc_x = self.loc[0](sources[0])
conf_x = self.conf[0](sources[0])
conf_x_b, conf_x_c, conf_x_h, conf_x_w = conf_x.value
conf_x = flops_counter.TensorSize([conf_x_b, 2, conf_x_h, conf_x_w])
loc.append(loc_x.permute(0, 2, 3, 1))
conf.append(conf_x.permute(0, 2, 3, 1))
for i in range(1, len(sources)):
x = sources[i]
conf.append(self.conf[i](x).permute(0, 2, 3, 1))
loc.append(self.loc[i](x).permute(0, 2, 3, 1))
loc = flops_counter.cat([o.view(o.size(0), -1) for o in loc], 1)
conf = flops_counter.cat([o.view(o.size(0), -1) for o in conf], 1)
return loc.view(loc.size(0), -1,
4), self.softmax(conf.view(conf.size(0), -1, 2))
import sys
sys.path.append('.')
import flops_counter as fc
from models import CSP
model = CSP()
# input size either height or width must be dividable by 16,
# this is due to the model design
x = fc.TensorSize([1, 3, 224, 112])
y = model(x)
print(y, model.flops)
# print(model)
