|
a |
|
b/landmark_extraction/models/yolo.py |
|
|
1 |
import argparse |
|
|
2 |
import logging |
|
|
3 |
import sys |
|
|
4 |
from copy import deepcopy |
|
|
5 |
|
|
|
6 |
sys.path.append('./') # to run '$ python *.py' files in subdirectories |
|
|
7 |
logger = logging.getLogger(__name__) |
|
|
8 |
import torch |
|
|
9 |
from models.common import * |
|
|
10 |
from models.experimental import * |
|
|
11 |
from utils.autoanchor import check_anchor_order |
|
|
12 |
from utils.general import make_divisible, check_file, set_logging |
|
|
13 |
from utils.torch_utils import time_synchronized, fuse_conv_and_bn, model_info, scale_img, initialize_weights, \ |
|
|
14 |
select_device, copy_attr |
|
|
15 |
from utils.loss import SigmoidBin |
|
|
16 |
|
|
|
17 |
try: |
|
|
18 |
import thop # for FLOPS computation |
|
|
19 |
except ImportError: |
|
|
20 |
thop = None |
|
|
21 |
|
|
|
22 |
|
|
|
23 |
class Detect(nn.Module): |
|
|
24 |
stride = None # strides computed during build |
|
|
25 |
export = False # onnx export |
|
|
26 |
end2end = False |
|
|
27 |
include_nms = False |
|
|
28 |
|
|
|
29 |
def __init__(self, nc=80, anchors=(), ch=()): # detection layer |
|
|
30 |
super(Detect, self).__init__() |
|
|
31 |
self.nc = nc # number of classes |
|
|
32 |
self.no = nc + 5 # number of outputs per anchor |
|
|
33 |
self.nl = len(anchors) # number of detection layers |
|
|
34 |
self.na = len(anchors[0]) // 2 # number of anchors |
|
|
35 |
self.grid = [torch.zeros(1)] * self.nl # init grid |
|
|
36 |
a = torch.tensor(anchors).float().view(self.nl, -1, 2) |
|
|
37 |
self.register_buffer('anchors', a) # shape(nl,na,2) |
|
|
38 |
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) |
|
|
39 |
self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv |
|
|
40 |
|
|
|
41 |
def forward(self, x): |
|
|
42 |
# x = x.copy() # for profiling |
|
|
43 |
z = [] # inference output |
|
|
44 |
self.training |= self.export |
|
|
45 |
for i in range(self.nl): |
|
|
46 |
x[i] = self.m[i](x[i]) # conv |
|
|
47 |
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) |
|
|
48 |
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() |
|
|
49 |
|
|
|
50 |
if not self.training: # inference |
|
|
51 |
if self.grid[i].shape[2:4] != x[i].shape[2:4]: |
|
|
52 |
self.grid[i] = self._make_grid(nx, ny).to(x[i].device) |
|
|
53 |
y = x[i].sigmoid() |
|
|
54 |
if not torch.onnx.is_in_onnx_export(): |
|
|
55 |
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy |
|
|
56 |
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh |
|
|
57 |
else: |
|
|
58 |
xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy |
|
|
59 |
wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh |
|
|
60 |
y = torch.cat((xy, wh, y[..., 4:]), -1) |
|
|
61 |
z.append(y.view(bs, -1, self.no)) |
|
|
62 |
|
|
|
63 |
if self.training: |
|
|
64 |
out = x |
|
|
65 |
elif self.end2end: |
|
|
66 |
out = torch.cat(z, 1) |
|
|
67 |
elif self.include_nms: |
|
|
68 |
z = self.convert(z) |
|
|
69 |
out = (z, ) |
|
|
70 |
else: |
|
|
71 |
out = (torch.cat(z, 1), x) |
|
|
72 |
|
|
|
73 |
return out |
|
|
74 |
|
|
|
75 |
@staticmethod |
|
|
76 |
def _make_grid(nx=20, ny=20): |
|
|
77 |
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) |
|
|
78 |
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() |
|
|
79 |
|
|
|
80 |
def convert(self, z): |
|
|
81 |
z = torch.cat(z, 1) |
|
|
82 |
box = z[:, :, :4] |
|
|
83 |
conf = z[:, :, 4:5] |
|
|
84 |
score = z[:, :, 5:] |
|
|
85 |
score *= conf |
|
|
86 |
convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]], |
|
|
87 |
dtype=torch.float32, |
|
|
88 |
device=z.device) |
|
|
89 |
box @= convert_matrix |
|
|
90 |
return (box, score) |
|
|
91 |
|
|
|
92 |
|
|
|
93 |
class IDetect(nn.Module): |
|
|
94 |
stride = None # strides computed during build |
|
|
95 |
export = False # onnx export |
|
|
96 |
end2end = False |
|
|
97 |
include_nms = False |
|
|
98 |
|
|
|
99 |
def __init__(self, nc=80, anchors=(), ch=()): # detection layer |
|
|
100 |
super(IDetect, self).__init__() |
|
|
101 |
self.nc = nc # number of classes |
|
|
102 |
self.no = nc + 5 # number of outputs per anchor |
|
|
103 |
self.nl = len(anchors) # number of detection layers |
|
|
104 |
self.na = len(anchors[0]) // 2 # number of anchors |
|
|
105 |
self.grid = [torch.zeros(1)] * self.nl # init grid |
|
|
106 |
a = torch.tensor(anchors).float().view(self.nl, -1, 2) |
|
|
107 |
self.register_buffer('anchors', a) # shape(nl,na,2) |
|
|
108 |
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) |
|
|
109 |
self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv |
|
|
110 |
|
|
|
111 |
self.ia = nn.ModuleList(ImplicitA(x) for x in ch) |
|
|
112 |
self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch) |
|
|
113 |
|
|
|
114 |
def forward(self, x): |
|
|
115 |
# x = x.copy() # for profiling |
|
|
116 |
z = [] # inference output |
|
|
117 |
self.training |= self.export |
|
|
118 |
for i in range(self.nl): |
|
|
119 |
x[i] = self.m[i](self.ia[i](x[i])) # conv |
|
|
120 |
x[i] = self.im[i](x[i]) |
|
|
121 |
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) |
|
|
122 |
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() |
|
|
123 |
|
|
|
124 |
if not self.training: # inference |
|
|
125 |
if self.grid[i].shape[2:4] != x[i].shape[2:4]: |
|
|
126 |
self.grid[i] = self._make_grid(nx, ny).to(x[i].device) |
|
|
127 |
|
|
|
128 |
y = x[i].sigmoid() |
|
|
129 |
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy |
|
|
130 |
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh |
|
|
131 |
z.append(y.view(bs, -1, self.no)) |
|
|
132 |
|
|
|
133 |
return x if self.training else (torch.cat(z, 1), x) |
|
|
134 |
|
|
|
135 |
def fuseforward(self, x): |
|
|
136 |
# x = x.copy() # for profiling |
|
|
137 |
z = [] # inference output |
|
|
138 |
self.training |= self.export |
|
|
139 |
for i in range(self.nl): |
|
|
140 |
x[i] = self.m[i](x[i]) # conv |
|
|
141 |
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) |
|
|
142 |
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() |
|
|
143 |
|
|
|
144 |
if not self.training: # inference |
|
|
145 |
if self.grid[i].shape[2:4] != x[i].shape[2:4]: |
|
|
146 |
self.grid[i] = self._make_grid(nx, ny).to(x[i].device) |
|
|
147 |
|
|
|
148 |
y = x[i].sigmoid() |
|
|
149 |
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy |
|
|
150 |
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh |
|
|
151 |
z.append(y.view(bs, -1, self.no)) |
|
|
152 |
|
|
|
153 |
if self.training: |
|
|
154 |
out = x |
|
|
155 |
elif self.end2end: |
|
|
156 |
out = torch.cat(z, 1) |
|
|
157 |
elif self.include_nms: |
|
|
158 |
z = self.convert(z) |
|
|
159 |
out = (z, ) |
|
|
160 |
else: |
|
|
161 |
out = (torch.cat(z, 1), x) |
|
|
162 |
|
|
|
163 |
return out |
|
|
164 |
|
|
|
165 |
def fuse(self): |
|
|
166 |
print("IDetect.fuse") |
|
|
167 |
# fuse ImplicitA and Convolution |
|
|
168 |
for i in range(len(self.m)): |
|
|
169 |
c1,c2,_,_ = self.m[i].weight.shape |
|
|
170 |
c1_,c2_, _,_ = self.ia[i].implicit.shape |
|
|
171 |
self.m[i].bias += torch.matmul(self.m[i].weight.reshape(c1,c2),self.ia[i].implicit.reshape(c2_,c1_)).squeeze(1) |
|
|
172 |
|
|
|
173 |
# fuse ImplicitM and Convolution |
|
|
174 |
for i in range(len(self.m)): |
|
|
175 |
c1,c2, _,_ = self.im[i].implicit.shape |
|
|
176 |
self.m[i].bias *= self.im[i].implicit.reshape(c2) |
|
|
177 |
self.m[i].weight *= self.im[i].implicit.transpose(0,1) |
|
|
178 |
|
|
|
179 |
@staticmethod |
|
|
180 |
def _make_grid(nx=20, ny=20): |
|
|
181 |
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) |
|
|
182 |
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() |
|
|
183 |
|
|
|
184 |
def convert(self, z): |
|
|
185 |
z = torch.cat(z, 1) |
|
|
186 |
box = z[:, :, :4] |
|
|
187 |
conf = z[:, :, 4:5] |
|
|
188 |
score = z[:, :, 5:] |
|
|
189 |
score *= conf |
|
|
190 |
convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]], |
|
|
191 |
dtype=torch.float32, |
|
|
192 |
device=z.device) |
|
|
193 |
box @= convert_matrix |
|
|
194 |
return (box, score) |
|
|
195 |
|
|
|
196 |
|
|
|
197 |
class IKeypoint(nn.Module): |
|
|
198 |
stride = None # strides computed during build |
|
|
199 |
export = False # onnx export |
|
|
200 |
|
|
|
201 |
def __init__(self, nc=80, anchors=(), nkpt=17, ch=(), inplace=True, dw_conv_kpt=False): # detection layer |
|
|
202 |
super(IKeypoint, self).__init__() |
|
|
203 |
self.nc = nc # number of classes |
|
|
204 |
self.nkpt = nkpt |
|
|
205 |
self.dw_conv_kpt = dw_conv_kpt |
|
|
206 |
self.no_det=(nc + 5) # number of outputs per anchor for box and class |
|
|
207 |
self.no_kpt = 3*self.nkpt ## number of outputs per anchor for keypoints |
|
|
208 |
self.no = self.no_det+self.no_kpt |
|
|
209 |
self.nl = len(anchors) # number of detection layers |
|
|
210 |
self.na = len(anchors[0]) // 2 # number of anchors |
|
|
211 |
self.grid = [torch.zeros(1)] * self.nl # init grid |
|
|
212 |
self.flip_test = False |
|
|
213 |
a = torch.tensor(anchors).float().view(self.nl, -1, 2) |
|
|
214 |
self.register_buffer('anchors', a) # shape(nl,na,2) |
|
|
215 |
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) |
|
|
216 |
self.m = nn.ModuleList(nn.Conv2d(x, self.no_det * self.na, 1) for x in ch) # output conv |
|
|
217 |
|
|
|
218 |
self.ia = nn.ModuleList(ImplicitA(x) for x in ch) |
|
|
219 |
self.im = nn.ModuleList(ImplicitM(self.no_det * self.na) for _ in ch) |
|
|
220 |
|
|
|
221 |
if self.nkpt is not None: |
|
|
222 |
if self.dw_conv_kpt: #keypoint head is slightly more complex |
|
|
223 |
self.m_kpt = nn.ModuleList( |
|
|
224 |
nn.Sequential(DWConv(x, x, k=3), Conv(x,x), |
|
|
225 |
DWConv(x, x, k=3), Conv(x, x), |
|
|
226 |
DWConv(x, x, k=3), Conv(x,x), |
|
|
227 |
DWConv(x, x, k=3), Conv(x, x), |
|
|
228 |
DWConv(x, x, k=3), Conv(x, x), |
|
|
229 |
DWConv(x, x, k=3), nn.Conv2d(x, self.no_kpt * self.na, 1)) for x in ch) |
|
|
230 |
else: #keypoint head is a single convolution |
|
|
231 |
self.m_kpt = nn.ModuleList(nn.Conv2d(x, self.no_kpt * self.na, 1) for x in ch) |
|
|
232 |
|
|
|
233 |
self.inplace = inplace # use in-place ops (e.g. slice assignment) |
|
|
234 |
|
|
|
235 |
def forward(self, x): |
|
|
236 |
# x = x.copy() # for profiling |
|
|
237 |
z = [] # inference output |
|
|
238 |
self.training |= self.export |
|
|
239 |
for i in range(self.nl): |
|
|
240 |
if self.nkpt is None or self.nkpt==0: |
|
|
241 |
x[i] = self.im[i](self.m[i](self.ia[i](x[i]))) # conv |
|
|
242 |
else : |
|
|
243 |
x[i] = torch.cat((self.im[i](self.m[i](self.ia[i](x[i]))), self.m_kpt[i](x[i])), axis=1) |
|
|
244 |
|
|
|
245 |
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) |
|
|
246 |
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() |
|
|
247 |
x_det = x[i][..., :6] |
|
|
248 |
x_kpt = x[i][..., 6:] |
|
|
249 |
|
|
|
250 |
if not self.training: # inference |
|
|
251 |
if self.grid[i].shape[2:4] != x[i].shape[2:4]: |
|
|
252 |
self.grid[i] = self._make_grid(nx, ny).to(x[i].device) |
|
|
253 |
kpt_grid_x = self.grid[i][..., 0:1] |
|
|
254 |
kpt_grid_y = self.grid[i][..., 1:2] |
|
|
255 |
|
|
|
256 |
if self.nkpt == 0: |
|
|
257 |
y = x[i].sigmoid() |
|
|
258 |
else: |
|
|
259 |
y = x_det.sigmoid() |
|
|
260 |
|
|
|
261 |
if self.inplace: |
|
|
262 |
xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy |
|
|
263 |
wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].view(1, self.na, 1, 1, 2) # wh |
|
|
264 |
if self.nkpt != 0: |
|
|
265 |
x_kpt[..., 0::3] = (x_kpt[..., ::3] * 2. - 0.5 + kpt_grid_x.repeat(1,1,1,1,17)) * self.stride[i] # xy |
|
|
266 |
x_kpt[..., 1::3] = (x_kpt[..., 1::3] * 2. - 0.5 + kpt_grid_y.repeat(1,1,1,1,17)) * self.stride[i] # xy |
|
|
267 |
#x_kpt[..., 0::3] = (x_kpt[..., ::3] + kpt_grid_x.repeat(1,1,1,1,17)) * self.stride[i] # xy |
|
|
268 |
#x_kpt[..., 1::3] = (x_kpt[..., 1::3] + kpt_grid_y.repeat(1,1,1,1,17)) * self.stride[i] # xy |
|
|
269 |
#print('=============') |
|
|
270 |
#print(self.anchor_grid[i].shape) |
|
|
271 |
#print(self.anchor_grid[i][...,0].unsqueeze(4).shape) |
|
|
272 |
#print(x_kpt[..., 0::3].shape) |
|
|
273 |
#x_kpt[..., 0::3] = ((x_kpt[..., 0::3].tanh() * 2.) ** 3 * self.anchor_grid[i][...,0].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_x.repeat(1,1,1,1,17) * self.stride[i] # xy |
|
|
274 |
#x_kpt[..., 1::3] = ((x_kpt[..., 1::3].tanh() * 2.) ** 3 * self.anchor_grid[i][...,1].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_y.repeat(1,1,1,1,17) * self.stride[i] # xy |
|
|
275 |
#x_kpt[..., 0::3] = (((x_kpt[..., 0::3].sigmoid() * 4.) ** 2 - 8.) * self.anchor_grid[i][...,0].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_x.repeat(1,1,1,1,17) * self.stride[i] # xy |
|
|
276 |
#x_kpt[..., 1::3] = (((x_kpt[..., 1::3].sigmoid() * 4.) ** 2 - 8.) * self.anchor_grid[i][...,1].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_y.repeat(1,1,1,1,17) * self.stride[i] # xy |
|
|
277 |
x_kpt[..., 2::3] = x_kpt[..., 2::3].sigmoid() |
|
|
278 |
|
|
|
279 |
y = torch.cat((xy, wh, y[..., 4:], x_kpt), dim = -1) |
|
|
280 |
|
|
|
281 |
else: # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953 |
|
|
282 |
xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy |
|
|
283 |
wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh |
|
|
284 |
if self.nkpt != 0: |
|
|
285 |
y[..., 6:] = (y[..., 6:] * 2. - 0.5 + self.grid[i].repeat((1,1,1,1,self.nkpt))) * self.stride[i] # xy |
|
|
286 |
y = torch.cat((xy, wh, y[..., 4:]), -1) |
|
|
287 |
|
|
|
288 |
z.append(y.view(bs, -1, self.no)) |
|
|
289 |
|
|
|
290 |
return x if self.training else (torch.cat(z, 1), x) |
|
|
291 |
|
|
|
292 |
@staticmethod |
|
|
293 |
def _make_grid(nx=20, ny=20): |
|
|
294 |
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) |
|
|
295 |
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() |
|
|
296 |
|
|
|
297 |
|
|
|
298 |
class IAuxDetect(nn.Module): |
|
|
299 |
stride = None # strides computed during build |
|
|
300 |
export = False # onnx export |
|
|
301 |
|
|
|
302 |
def __init__(self, nc=80, anchors=(), ch=()): # detection layer |
|
|
303 |
super(IAuxDetect, self).__init__() |
|
|
304 |
self.nc = nc # number of classes |
|
|
305 |
self.no = nc + 5 # number of outputs per anchor |
|
|
306 |
self.nl = len(anchors) # number of detection layers |
|
|
307 |
self.na = len(anchors[0]) // 2 # number of anchors |
|
|
308 |
self.grid = [torch.zeros(1)] * self.nl # init grid |
|
|
309 |
a = torch.tensor(anchors).float().view(self.nl, -1, 2) |
|
|
310 |
self.register_buffer('anchors', a) # shape(nl,na,2) |
|
|
311 |
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) |
|
|
312 |
self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch[:self.nl]) # output conv |
|
|
313 |
self.m2 = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch[self.nl:]) # output conv |
|
|
314 |
|
|
|
315 |
self.ia = nn.ModuleList(ImplicitA(x) for x in ch[:self.nl]) |
|
|
316 |
self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch[:self.nl]) |
|
|
317 |
|
|
|
318 |
def forward(self, x): |
|
|
319 |
# x = x.copy() # for profiling |
|
|
320 |
z = [] # inference output |
|
|
321 |
self.training |= self.export |
|
|
322 |
for i in range(self.nl): |
|
|
323 |
x[i] = self.m[i](self.ia[i](x[i])) # conv |
|
|
324 |
x[i] = self.im[i](x[i]) |
|
|
325 |
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) |
|
|
326 |
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() |
|
|
327 |
|
|
|
328 |
x[i+self.nl] = self.m2[i](x[i+self.nl]) |
|
|
329 |
x[i+self.nl] = x[i+self.nl].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() |
|
|
330 |
|
|
|
331 |
if not self.training: # inference |
|
|
332 |
if self.grid[i].shape[2:4] != x[i].shape[2:4]: |
|
|
333 |
self.grid[i] = self._make_grid(nx, ny).to(x[i].device) |
|
|
334 |
|
|
|
335 |
y = x[i].sigmoid() |
|
|
336 |
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy |
|
|
337 |
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh |
|
|
338 |
z.append(y.view(bs, -1, self.no)) |
|
|
339 |
|
|
|
340 |
return x if self.training else (torch.cat(z, 1), x[:self.nl]) |
|
|
341 |
|
|
|
342 |
@staticmethod |
|
|
343 |
def _make_grid(nx=20, ny=20): |
|
|
344 |
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) |
|
|
345 |
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() |
|
|
346 |
|
|
|
347 |
|
|
|
348 |
class IBin(nn.Module): |
|
|
349 |
stride = None # strides computed during build |
|
|
350 |
export = False # onnx export |
|
|
351 |
|
|
|
352 |
def __init__(self, nc=80, anchors=(), ch=(), bin_count=21): # detection layer |
|
|
353 |
super(IBin, self).__init__() |
|
|
354 |
self.nc = nc # number of classes |
|
|
355 |
self.bin_count = bin_count |
|
|
356 |
|
|
|
357 |
self.w_bin_sigmoid = SigmoidBin(bin_count=self.bin_count, min=0.0, max=4.0) |
|
|
358 |
self.h_bin_sigmoid = SigmoidBin(bin_count=self.bin_count, min=0.0, max=4.0) |
|
|
359 |
# classes, x,y,obj |
|
|
360 |
self.no = nc + 3 + \ |
|
|
361 |
self.w_bin_sigmoid.get_length() + self.h_bin_sigmoid.get_length() # w-bce, h-bce |
|
|
362 |
# + self.x_bin_sigmoid.get_length() + self.y_bin_sigmoid.get_length() |
|
|
363 |
|
|
|
364 |
self.nl = len(anchors) # number of detection layers |
|
|
365 |
self.na = len(anchors[0]) // 2 # number of anchors |
|
|
366 |
self.grid = [torch.zeros(1)] * self.nl # init grid |
|
|
367 |
a = torch.tensor(anchors).float().view(self.nl, -1, 2) |
|
|
368 |
self.register_buffer('anchors', a) # shape(nl,na,2) |
|
|
369 |
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) |
|
|
370 |
self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv |
|
|
371 |
|
|
|
372 |
self.ia = nn.ModuleList(ImplicitA(x) for x in ch) |
|
|
373 |
self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch) |
|
|
374 |
|
|
|
375 |
def forward(self, x): |
|
|
376 |
|
|
|
377 |
#self.x_bin_sigmoid.use_fw_regression = True |
|
|
378 |
#self.y_bin_sigmoid.use_fw_regression = True |
|
|
379 |
self.w_bin_sigmoid.use_fw_regression = True |
|
|
380 |
self.h_bin_sigmoid.use_fw_regression = True |
|
|
381 |
|
|
|
382 |
# x = x.copy() # for profiling |
|
|
383 |
z = [] # inference output |
|
|
384 |
self.training |= self.export |
|
|
385 |
for i in range(self.nl): |
|
|
386 |
x[i] = self.m[i](self.ia[i](x[i])) # conv |
|
|
387 |
x[i] = self.im[i](x[i]) |
|
|
388 |
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) |
|
|
389 |
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() |
|
|
390 |
|
|
|
391 |
if not self.training: # inference |
|
|
392 |
if self.grid[i].shape[2:4] != x[i].shape[2:4]: |
|
|
393 |
self.grid[i] = self._make_grid(nx, ny).to(x[i].device) |
|
|
394 |
|
|
|
395 |
y = x[i].sigmoid() |
|
|
396 |
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy |
|
|
397 |
#y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh |
|
|
398 |
|
|
|
399 |
|
|
|
400 |
#px = (self.x_bin_sigmoid.forward(y[..., 0:12]) + self.grid[i][..., 0]) * self.stride[i] |
|
|
401 |
#py = (self.y_bin_sigmoid.forward(y[..., 12:24]) + self.grid[i][..., 1]) * self.stride[i] |
|
|
402 |
|
|
|
403 |
pw = self.w_bin_sigmoid.forward(y[..., 2:24]) * self.anchor_grid[i][..., 0] |
|
|
404 |
ph = self.h_bin_sigmoid.forward(y[..., 24:46]) * self.anchor_grid[i][..., 1] |
|
|
405 |
|
|
|
406 |
#y[..., 0] = px |
|
|
407 |
#y[..., 1] = py |
|
|
408 |
y[..., 2] = pw |
|
|
409 |
y[..., 3] = ph |
|
|
410 |
|
|
|
411 |
y = torch.cat((y[..., 0:4], y[..., 46:]), dim=-1) |
|
|
412 |
|
|
|
413 |
z.append(y.view(bs, -1, y.shape[-1])) |
|
|
414 |
|
|
|
415 |
return x if self.training else (torch.cat(z, 1), x) |
|
|
416 |
|
|
|
417 |
@staticmethod |
|
|
418 |
def _make_grid(nx=20, ny=20): |
|
|
419 |
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) |
|
|
420 |
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() |
|
|
421 |
|
|
|
422 |
|
|
|
423 |
class Model(nn.Module): |
|
|
424 |
def __init__(self, cfg='yolor-csp-c.yaml', ch=3, nc=None, anchors=None): # model, input channels, number of classes |
|
|
425 |
super(Model, self).__init__() |
|
|
426 |
self.traced = False |
|
|
427 |
if isinstance(cfg, dict): |
|
|
428 |
self.yaml = cfg # model dict |
|
|
429 |
else: # is *.yaml |
|
|
430 |
import yaml # for torch hub |
|
|
431 |
self.yaml_file = Path(cfg).name |
|
|
432 |
with open(cfg) as f: |
|
|
433 |
self.yaml = yaml.load(f, Loader=yaml.SafeLoader) # model dict |
|
|
434 |
|
|
|
435 |
# Define model |
|
|
436 |
ch = self.yaml['ch'] = self.yaml.get('ch', ch) # input channels |
|
|
437 |
if nc and nc != self.yaml['nc']: |
|
|
438 |
logger.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}") |
|
|
439 |
self.yaml['nc'] = nc # override yaml value |
|
|
440 |
if anchors: |
|
|
441 |
logger.info(f'Overriding model.yaml anchors with anchors={anchors}') |
|
|
442 |
self.yaml['anchors'] = round(anchors) # override yaml value |
|
|
443 |
self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch]) # model, savelist |
|
|
444 |
self.names = [str(i) for i in range(self.yaml['nc'])] # default names |
|
|
445 |
# print([x.shape for x in self.forward(torch.zeros(1, ch, 64, 64))]) |
|
|
446 |
|
|
|
447 |
# Build strides, anchors |
|
|
448 |
m = self.model[-1] # Detect() |
|
|
449 |
if isinstance(m, Detect): |
|
|
450 |
s = 256 # 2x min stride |
|
|
451 |
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward |
|
|
452 |
m.anchors /= m.stride.view(-1, 1, 1) |
|
|
453 |
check_anchor_order(m) |
|
|
454 |
self.stride = m.stride |
|
|
455 |
self._initialize_biases() # only run once |
|
|
456 |
# print('Strides: %s' % m.stride.tolist()) |
|
|
457 |
if isinstance(m, IDetect): |
|
|
458 |
s = 256 # 2x min stride |
|
|
459 |
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward |
|
|
460 |
m.anchors /= m.stride.view(-1, 1, 1) |
|
|
461 |
check_anchor_order(m) |
|
|
462 |
self.stride = m.stride |
|
|
463 |
self._initialize_biases() # only run once |
|
|
464 |
# print('Strides: %s' % m.stride.tolist()) |
|
|
465 |
if isinstance(m, IAuxDetect): |
|
|
466 |
s = 256 # 2x min stride |
|
|
467 |
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))[:4]]) # forward |
|
|
468 |
#print(m.stride) |
|
|
469 |
m.anchors /= m.stride.view(-1, 1, 1) |
|
|
470 |
check_anchor_order(m) |
|
|
471 |
self.stride = m.stride |
|
|
472 |
self._initialize_aux_biases() # only run once |
|
|
473 |
# print('Strides: %s' % m.stride.tolist()) |
|
|
474 |
if isinstance(m, IBin): |
|
|
475 |
s = 256 # 2x min stride |
|
|
476 |
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward |
|
|
477 |
m.anchors /= m.stride.view(-1, 1, 1) |
|
|
478 |
check_anchor_order(m) |
|
|
479 |
self.stride = m.stride |
|
|
480 |
self._initialize_biases_bin() # only run once |
|
|
481 |
# print('Strides: %s' % m.stride.tolist()) |
|
|
482 |
if isinstance(m, IKeypoint): |
|
|
483 |
s = 256 # 2x min stride |
|
|
484 |
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward |
|
|
485 |
m.anchors /= m.stride.view(-1, 1, 1) |
|
|
486 |
check_anchor_order(m) |
|
|
487 |
self.stride = m.stride |
|
|
488 |
self._initialize_biases_kpt() # only run once |
|
|
489 |
# print('Strides: %s' % m.stride.tolist()) |
|
|
490 |
|
|
|
491 |
# Init weights, biases |
|
|
492 |
initialize_weights(self) |
|
|
493 |
self.info() |
|
|
494 |
logger.info('') |
|
|
495 |
|
|
|
496 |
def forward(self, x, augment=False, profile=False): |
|
|
497 |
if augment: |
|
|
498 |
img_size = x.shape[-2:] # height, width |
|
|
499 |
s = [1, 0.83, 0.67] # scales |
|
|
500 |
f = [None, 3, None] # flips (2-ud, 3-lr) |
|
|
501 |
y = [] # outputs |
|
|
502 |
for si, fi in zip(s, f): |
|
|
503 |
xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max())) |
|
|
504 |
yi = self.forward_once(xi)[0] # forward |
|
|
505 |
# cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1]) # save |
|
|
506 |
yi[..., :4] /= si # de-scale |
|
|
507 |
if fi == 2: |
|
|
508 |
yi[..., 1] = img_size[0] - yi[..., 1] # de-flip ud |
|
|
509 |
elif fi == 3: |
|
|
510 |
yi[..., 0] = img_size[1] - yi[..., 0] # de-flip lr |
|
|
511 |
y.append(yi) |
|
|
512 |
return torch.cat(y, 1), None # augmented inference, train |
|
|
513 |
else: |
|
|
514 |
return self.forward_once(x, profile) # single-scale inference, train |
|
|
515 |
|
|
|
516 |
def forward_once(self, x, profile=False): |
|
|
517 |
y, dt = [], [] # outputs |
|
|
518 |
for m in self.model: |
|
|
519 |
if m.f != -1: # if not from previous layer |
|
|
520 |
x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers |
|
|
521 |
|
|
|
522 |
if not hasattr(self, 'traced'): |
|
|
523 |
self.traced=False |
|
|
524 |
|
|
|
525 |
if self.traced: |
|
|
526 |
if isinstance(m, Detect) or isinstance(m, IDetect) or isinstance(m, IAuxDetect) or isinstance(m, IKeypoint): |
|
|
527 |
break |
|
|
528 |
|
|
|
529 |
if profile: |
|
|
530 |
c = isinstance(m, (Detect, IDetect, IAuxDetect, IBin)) |
|
|
531 |
o = thop.profile(m, inputs=(x.copy() if c else x,), verbose=False)[0] / 1E9 * 2 if thop else 0 # FLOPS |
|
|
532 |
for _ in range(10): |
|
|
533 |
m(x.copy() if c else x) |
|
|
534 |
t = time_synchronized() |
|
|
535 |
for _ in range(10): |
|
|
536 |
m(x.copy() if c else x) |
|
|
537 |
dt.append((time_synchronized() - t) * 100) |
|
|
538 |
print('%10.1f%10.0f%10.1fms %-40s' % (o, m.np, dt[-1], m.type)) |
|
|
539 |
|
|
|
540 |
x = m(x) # run |
|
|
541 |
|
|
|
542 |
y.append(x if m.i in self.save else None) # save output |
|
|
543 |
|
|
|
544 |
if profile: |
|
|
545 |
print('%.1fms total' % sum(dt)) |
|
|
546 |
return x |
|
|
547 |
|
|
|
548 |
def _initialize_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency |
|
|
549 |
# https://arxiv.org/abs/1708.02002 section 3.3 |
|
|
550 |
# cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. |
|
|
551 |
m = self.model[-1] # Detect() module |
|
|
552 |
for mi, s in zip(m.m, m.stride): # from |
|
|
553 |
b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85) |
|
|
554 |
b.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) |
|
|
555 |
b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls |
|
|
556 |
mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) |
|
|
557 |
|
|
|
558 |
def _initialize_aux_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency |
|
|
559 |
# https://arxiv.org/abs/1708.02002 section 3.3 |
|
|
560 |
# cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. |
|
|
561 |
m = self.model[-1] # Detect() module |
|
|
562 |
for mi, mi2, s in zip(m.m, m.m2, m.stride): # from |
|
|
563 |
b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85) |
|
|
564 |
b.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) |
|
|
565 |
b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls |
|
|
566 |
mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) |
|
|
567 |
b2 = mi2.bias.view(m.na, -1) # conv.bias(255) to (3,85) |
|
|
568 |
b2.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) |
|
|
569 |
b2.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls |
|
|
570 |
mi2.bias = torch.nn.Parameter(b2.view(-1), requires_grad=True) |
|
|
571 |
|
|
|
572 |
def _initialize_biases_bin(self, cf=None): # initialize biases into Detect(), cf is class frequency |
|
|
573 |
# https://arxiv.org/abs/1708.02002 section 3.3 |
|
|
574 |
# cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. |
|
|
575 |
m = self.model[-1] # Bin() module |
|
|
576 |
bc = m.bin_count |
|
|
577 |
for mi, s in zip(m.m, m.stride): # from |
|
|
578 |
b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85) |
|
|
579 |
old = b[:, (0,1,2,bc+3)].data |
|
|
580 |
obj_idx = 2*bc+4 |
|
|
581 |
b[:, :obj_idx].data += math.log(0.6 / (bc + 1 - 0.99)) |
|
|
582 |
b[:, obj_idx].data += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) |
|
|
583 |
b[:, (obj_idx+1):].data += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls |
|
|
584 |
b[:, (0,1,2,bc+3)].data = old |
|
|
585 |
mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) |
|
|
586 |
|
|
|
587 |
def _initialize_biases_kpt(self, cf=None): # initialize biases into Detect(), cf is class frequency |
|
|
588 |
# https://arxiv.org/abs/1708.02002 section 3.3 |
|
|
589 |
# cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. |
|
|
590 |
m = self.model[-1] # Detect() module |
|
|
591 |
for mi, s in zip(m.m, m.stride): # from |
|
|
592 |
b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85) |
|
|
593 |
b.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) |
|
|
594 |
b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls |
|
|
595 |
mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) |
|
|
596 |
|
|
|
597 |
def _print_biases(self): |
|
|
598 |
m = self.model[-1] # Detect() module |
|
|
599 |
for mi in m.m: # from |
|
|
600 |
b = mi.bias.detach().view(m.na, -1).T # conv.bias(255) to (3,85) |
|
|
601 |
print(('%6g Conv2d.bias:' + '%10.3g' * 6) % (mi.weight.shape[1], *b[:5].mean(1).tolist(), b[5:].mean())) |
|
|
602 |
|
|
|
603 |
# def _print_weights(self): |
|
|
604 |
# for m in self.model.modules(): |
|
|
605 |
# if type(m) is Bottleneck: |
|
|
606 |
# print('%10.3g' % (m.w.detach().sigmoid() * 2)) # shortcut weights |
|
|
607 |
|
|
|
608 |
def fuse(self): # fuse model Conv2d() + BatchNorm2d() layers |
|
|
609 |
print('Fusing layers... ') |
|
|
610 |
for m in self.model.modules(): |
|
|
611 |
if isinstance(m, RepConv): |
|
|
612 |
#print(f" fuse_repvgg_block") |
|
|
613 |
m.fuse_repvgg_block() |
|
|
614 |
elif isinstance(m, RepConv_OREPA): |
|
|
615 |
#print(f" switch_to_deploy") |
|
|
616 |
m.switch_to_deploy() |
|
|
617 |
elif type(m) is Conv and hasattr(m, 'bn'): |
|
|
618 |
m.conv = fuse_conv_and_bn(m.conv, m.bn) # update conv |
|
|
619 |
delattr(m, 'bn') # remove batchnorm |
|
|
620 |
m.forward = m.fuseforward # update forward |
|
|
621 |
elif isinstance(m, IDetect): |
|
|
622 |
m.fuse() |
|
|
623 |
m.forward = m.fuseforward |
|
|
624 |
self.info() |
|
|
625 |
return self |
|
|
626 |
|
|
|
627 |
def nms(self, mode=True): # add or remove NMS module |
|
|
628 |
present = type(self.model[-1]) is NMS # last layer is NMS |
|
|
629 |
if mode and not present: |
|
|
630 |
print('Adding NMS... ') |
|
|
631 |
m = NMS() # module |
|
|
632 |
m.f = -1 # from |
|
|
633 |
m.i = self.model[-1].i + 1 # index |
|
|
634 |
self.model.add_module(name='%s' % m.i, module=m) # add |
|
|
635 |
self.eval() |
|
|
636 |
elif not mode and present: |
|
|
637 |
print('Removing NMS... ') |
|
|
638 |
self.model = self.model[:-1] # remove |
|
|
639 |
return self |
|
|
640 |
|
|
|
641 |
def autoshape(self): # add autoShape module |
|
|
642 |
print('Adding autoShape... ') |
|
|
643 |
m = autoShape(self) # wrap model |
|
|
644 |
copy_attr(m, self, include=('yaml', 'nc', 'hyp', 'names', 'stride'), exclude=()) # copy attributes |
|
|
645 |
return m |
|
|
646 |
|
|
|
647 |
def info(self, verbose=False, img_size=640): # print model information |
|
|
648 |
model_info(self, verbose, img_size) |
|
|
649 |
|
|
|
650 |
|
|
|
651 |
def parse_model(d, ch): # model_dict, input_channels(3) |
|
|
652 |
logger.info('\n%3s%18s%3s%10s %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments')) |
|
|
653 |
anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple'] |
|
|
654 |
na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors # number of anchors |
|
|
655 |
no = na * (nc + 5) # number of outputs = anchors * (classes + 5) |
|
|
656 |
|
|
|
657 |
layers, save, c2 = [], [], ch[-1] # layers, savelist, ch out |
|
|
658 |
for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']): # from, number, module, args |
|
|
659 |
m = eval(m) if isinstance(m, str) else m # eval strings |
|
|
660 |
for j, a in enumerate(args): |
|
|
661 |
try: |
|
|
662 |
args[j] = eval(a) if isinstance(a, str) else a # eval strings |
|
|
663 |
except: |
|
|
664 |
pass |
|
|
665 |
|
|
|
666 |
n = max(round(n * gd), 1) if n > 1 else n # depth gain |
|
|
667 |
if m in [nn.Conv2d, Conv, RobustConv, RobustConv2, DWConv, GhostConv, RepConv, RepConv_OREPA, DownC, |
|
|
668 |
SPP, SPPF, SPPCSPC, GhostSPPCSPC, MixConv2d, Focus, Stem, GhostStem, CrossConv, |
|
|
669 |
Bottleneck, BottleneckCSPA, BottleneckCSPB, BottleneckCSPC, |
|
|
670 |
RepBottleneck, RepBottleneckCSPA, RepBottleneckCSPB, RepBottleneckCSPC, |
|
|
671 |
Res, ResCSPA, ResCSPB, ResCSPC, |
|
|
672 |
RepRes, RepResCSPA, RepResCSPB, RepResCSPC, |
|
|
673 |
ResX, ResXCSPA, ResXCSPB, ResXCSPC, |
|
|
674 |
RepResX, RepResXCSPA, RepResXCSPB, RepResXCSPC, |
|
|
675 |
Ghost, GhostCSPA, GhostCSPB, GhostCSPC, |
|
|
676 |
SwinTransformerBlock, STCSPA, STCSPB, STCSPC, |
|
|
677 |
SwinTransformer2Block, ST2CSPA, ST2CSPB, ST2CSPC]: |
|
|
678 |
c1, c2 = ch[f], args[0] |
|
|
679 |
if c2 != no: # if not output |
|
|
680 |
c2 = make_divisible(c2 * gw, 8) |
|
|
681 |
|
|
|
682 |
args = [c1, c2, *args[1:]] |
|
|
683 |
if m in [DownC, SPPCSPC, GhostSPPCSPC, |
|
|
684 |
BottleneckCSPA, BottleneckCSPB, BottleneckCSPC, |
|
|
685 |
RepBottleneckCSPA, RepBottleneckCSPB, RepBottleneckCSPC, |
|
|
686 |
ResCSPA, ResCSPB, ResCSPC, |
|
|
687 |
RepResCSPA, RepResCSPB, RepResCSPC, |
|
|
688 |
ResXCSPA, ResXCSPB, ResXCSPC, |
|
|
689 |
RepResXCSPA, RepResXCSPB, RepResXCSPC, |
|
|
690 |
GhostCSPA, GhostCSPB, GhostCSPC, |
|
|
691 |
STCSPA, STCSPB, STCSPC, |
|
|
692 |
ST2CSPA, ST2CSPB, ST2CSPC]: |
|
|
693 |
args.insert(2, n) # number of repeats |
|
|
694 |
n = 1 |
|
|
695 |
elif m is nn.BatchNorm2d: |
|
|
696 |
args = [ch[f]] |
|
|
697 |
elif m is Concat: |
|
|
698 |
c2 = sum([ch[x] for x in f]) |
|
|
699 |
elif m is Chuncat: |
|
|
700 |
c2 = sum([ch[x] for x in f]) |
|
|
701 |
elif m is Shortcut: |
|
|
702 |
c2 = ch[f[0]] |
|
|
703 |
elif m is Foldcut: |
|
|
704 |
c2 = ch[f] // 2 |
|
|
705 |
elif m in [Detect, IDetect, IAuxDetect, IBin, IKeypoint]: |
|
|
706 |
args.append([ch[x] for x in f]) |
|
|
707 |
if isinstance(args[1], int): # number of anchors |
|
|
708 |
args[1] = [list(range(args[1] * 2))] * len(f) |
|
|
709 |
elif m is ReOrg: |
|
|
710 |
c2 = ch[f] * 4 |
|
|
711 |
elif m is Contract: |
|
|
712 |
c2 = ch[f] * args[0] ** 2 |
|
|
713 |
elif m is Expand: |
|
|
714 |
c2 = ch[f] // args[0] ** 2 |
|
|
715 |
else: |
|
|
716 |
c2 = ch[f] |
|
|
717 |
|
|
|
718 |
m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args) # module |
|
|
719 |
t = str(m)[8:-2].replace('__main__.', '') # module type |
|
|
720 |
np = sum([x.numel() for x in m_.parameters()]) # number params |
|
|
721 |
m_.i, m_.f, m_.type, m_.np = i, f, t, np # attach index, 'from' index, type, number params |
|
|
722 |
logger.info('%3s%18s%3s%10.0f %-40s%-30s' % (i, f, n, np, t, args)) # print |
|
|
723 |
save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist |
|
|
724 |
layers.append(m_) |
|
|
725 |
if i == 0: |
|
|
726 |
ch = [] |
|
|
727 |
ch.append(c2) |
|
|
728 |
return nn.Sequential(*layers), sorted(save) |
|
|
729 |
|
|
|
730 |
|
|
|
731 |
if __name__ == '__main__': |
|
|
732 |
parser = argparse.ArgumentParser() |
|
|
733 |
parser.add_argument('--cfg', type=str, default='yolor-csp-c.yaml', help='model.yaml') |
|
|
734 |
parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') |
|
|
735 |
parser.add_argument('--profile', action='store_true', help='profile model speed') |
|
|
736 |
opt = parser.parse_args() |
|
|
737 |
opt.cfg = check_file(opt.cfg) # check file |
|
|
738 |
set_logging() |
|
|
739 |
device = select_device(opt.device) |
|
|
740 |
|
|
|
741 |
# Create model |
|
|
742 |
model = Model(opt.cfg).to(device) |
|
|
743 |
model.train() |
|
|
744 |
|
|
|
745 |
if opt.profile: |
|
|
746 |
img = torch.rand(1, 3, 640, 640).to(device) |
|
|
747 |
y = model(img, profile=True) |
|
|
748 |
|
|
|
749 |
# Profile |
|
|
750 |
# img = torch.rand(8 if torch.cuda.is_available() else 1, 3, 640, 640).to(device) |
|
|
751 |
# y = model(img, profile=True) |
|
|
752 |
|
|
|
753 |
# Tensorboard |
|
|
754 |
# from torch.utils.tensorboard import SummaryWriter |
|
|
755 |
# tb_writer = SummaryWriter() |
|
|
756 |
# print("Run 'tensorboard --logdir=models/runs' to view tensorboard at http://localhost:6006/") |
|
|
757 |
# tb_writer.add_graph(model.model, img) # add model to tensorboard |
|
|
758 |
# tb_writer.add_image('test', img[0], dataformats='CWH') # add model to tensorboard |