-
Notifications
You must be signed in to change notification settings - Fork 136
/
convert_FPGA.py
193 lines (178 loc) · 9.99 KB
/
convert_FPGA.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
import argparse
import struct
from models import * # set ONNX_EXPORT in models.py
from utils.utils import *
def convert():
img_size = (320, 192) if ONNX_EXPORT else opt.img_size # (320, 192) or (416, 256) or (608, 352) for (height, width)
weights = opt.weights
# Initialize
device = torch_utils.select_device(device='cpu' if ONNX_EXPORT else opt.device)
# Initialize model
model = Darknet(opt.cfg, img_size, quantized=opt.quantized, a_bit=opt.a_bit, w_bit=opt.w_bit,
FPGA=opt.FPGA, is_gray_scale=opt.gray_scale)
# Load weights
attempt_download(weights)
if weights.endswith('.pt'): # pytorch format
model.load_state_dict(torch.load(weights, map_location=device)['model'])
else: # darknet format
_ = load_darknet_weights(model, weights, FPGA=opt.FPGA)
if opt.quantized == 0:
save_weights(model, path='weights/' + opt.cfg.split('/')[-1].replace('.cfg', '') + '-best.weights')
else:
w_file = open('weights/' + opt.cfg.split('/')[-1].replace('.cfg', '') + '_weights.bin', 'wb')
b_file = open('weights/' + opt.cfg.split('/')[-1].replace('.cfg', '') + '_bias.bin', 'wb')
if opt.quantized == 1:
w_scale = open('weights/' + opt.cfg.split('/')[-1].replace('.cfg', '') + '_w_scale.bin', 'wb')
a_scale = open('weights/' + opt.cfg.split('/')[-1].replace('.cfg', '') + '_a_scale.bin', 'wb')
b_scale = open('weights/' + opt.cfg.split('/')[-1].replace('.cfg', '') + '_b_scale.bin', 'wb')
s_scale = open('weights/' + opt.cfg.split('/')[-1].replace('.cfg', '') + '_s_scale.bin', 'wb')
if opt.w_bit == 16:
a = struct.pack('<i', 14)
if opt.w_bit == 8:
a = struct.pack('<i', 7)
a_scale.write(a)
for _, (mdef, module) in enumerate(zip(model.module_defs, model.module_list)):
print(mdef)
if mdef['type'] == 'convolutional':
conv_layer = module[0]
# 使用BN训练中量化，融合BN参数
weight, bias = conv_layer.BN_fuse()
if opt.quantized == 1:
# 得到缩放因子
activate_scale = -math.log(conv_layer.activation_quantizer.scale.cpu().data.numpy()[0], 2)
weight_scale = -math.log(conv_layer.weight_quantizer.scale.cpu().data.numpy()[0], 2)
a = struct.pack('<i', int(activate_scale))
a_scale.write(a)
a = struct.pack('<i', int(weight_scale))
w_scale.write(a)
# 处理weights
para = conv_layer.weight_quantizer.get_quantize_value(weight)
if opt.reorder:
# 重排序参数
print("use reorder!")
shape_output = para.shape[0]
shape_input = para.shape[1]
num_TN = int(shape_input / opt.TN)
remainder_TN = shape_input % opt.TN
num_TM = int(shape_output / opt.TM)
remainder_TM = shape_output % opt.TM
first = True
for j in range(num_TM):
for k in range(num_TN):
temp = para[j * opt.TM:(j + 1) * opt.TM, k * opt.TN:(k + 1) * opt.TN, :, :]
temp = temp.view(temp.shape[0], temp.shape[1], temp.shape[2] * temp.shape[3])
temp = temp.permute(2, 0, 1).contiguous().view(-1)
if first:
reorder_para = temp.clone().cpu().data.numpy()
first = False
else:
reorder_para = np.append(reorder_para, temp.cpu().data.numpy())
temp = para[j * opt.TM:(j + 1) * opt.TM, num_TN * opt.TN:num_TN * opt.TN + remainder_TN, :, :]
temp = temp.view(temp.shape[0], temp.shape[1], temp.shape[2] * temp.shape[3])
temp = temp.permute(2, 0, 1).contiguous().view(-1)
if first:
reorder_para = temp.clone().cpu().data.numpy()
first = False
else:
reorder_para = np.append(reorder_para, temp.cpu().data.numpy())
for k in range(num_TN):
temp = para[num_TM * opt.TM:num_TM * opt.TM + remainder_TM, k * opt.TN:(k + 1) * opt.TN, :, :]
temp = temp.view(temp.shape[0], temp.shape[1], temp.shape[2] * temp.shape[3])
temp = temp.permute(2, 0, 1).contiguous().view(-1)
if first:
reorder_para = temp.clone().cpu().data.numpy()
first = False
else:
reorder_para = np.append(reorder_para, temp.cpu().data.numpy())
temp = para[num_TM * opt.TM:num_TM * opt.TM + remainder_TM,
num_TN * opt.TN:num_TN * opt.TN + remainder_TN, :, :]
temp = temp.view(temp.shape[0], temp.shape[1], temp.shape[2] * temp.shape[3])
temp = temp.permute(2, 0, 1).contiguous().view(-1)
if first:
reorder_para = temp.clone().cpu().data.numpy()
first = False
else:
reorder_para = np.append(reorder_para, temp.cpu().data.numpy())
para_flatten = reorder_para
else:
para_flatten = para.cpu().data.numpy().flatten() # 展开
# 存储weights
for i in para_flatten:
if opt.w_bit == 16:
# Dorefa量化为非对称量化 Google量化为对称量化
if opt.quantized == 1:
a = struct.pack('<h', int(i))
if opt.quantized == 2:
a = struct.pack('<H', int(i))
elif opt.w_bit == 8:
# Dorefa量化为非对称量化 Google量化为对称量化
if opt.quantized == 1:
a = struct.pack('b', int(i))
if opt.quantized == 2:
a = struct.pack('B', int(i))
else:
a = struct.pack('<f', i)
w_file.write(a)
# 处理bias
if bias != None:
# 生成量化后的参数
para = conv_layer.bias_quantizer.get_quantize_value(bias)
if opt.quantized == 1:
bias_scale = -math.log(conv_layer.bias_quantizer.scale.cpu().data.numpy()[0], 2)
a = struct.pack('<i', int(bias_scale))
b_scale.write(a)
# print(para.shape)
para_flatten = para.cpu().data.numpy().flatten() # 展开
# 存储bias
for i in para_flatten:
if opt.w_bit == 16:
# Dorefa量化为非对称量化 Google量化为对称量化
if opt.quantized == 1:
a = struct.pack('<h', int(i))
if opt.quantized == 2:
a = struct.pack('<H', int(i))
elif opt.w_bit == 8:
# Dorefa量化为非对称量化 Google量化为对称量化
if opt.quantized == 1:
a = struct.pack('b', int(i))
if opt.quantized == 2:
a = struct.pack('B', int(i))
else:
a = struct.pack('<f', i)
b_file.write(a)
if mdef['type'] == 'shortcut':
shortcut_scale = -math.log(module.scale.cpu().data.numpy()[0], 2)
a = struct.pack('<i', int(shortcut_scale))
s_scale.write(a)
if opt.quantized == 1:
w_scale.close()
a_scale.close()
b_scale.close()
s_scale.close()
w_file.close()
b_file.close()
# Eval mode
model.to(device).eval()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--cfg', type=str, default='cfg/yolov3.cfg', help='cfg file path')
parser.add_argument('--data', type=str, default='data/coco2017.data', help='coco.data file path')
parser.add_argument('--weights', type=str, default='weights/yolov3.weights', help='path to weights file')
parser.add_argument('--source', type=str, default='data/samples', help='source') # input file/folder, 0 for webcam
parser.add_argument('--output', type=str, default='output', help='output folder') # output folder
parser.add_argument('--img_size', type=int, default=416, help='inference size (pixels)')
parser.add_argument('--conf-thres', type=float, default=0.6, help='object confidence threshold')
parser.add_argument('--nms-thres', type=float, default=0.8, help='iou threshold for non-maximum suppression')
parser.add_argument('--device', default='', help='device id (i.e. 0 or 0,1) or cpu')
parser.add_argument('--quantized', type=int, default=-1, help='quantization way')
parser.add_argument('--a-bit', type=int, default=8, help='a-bit')
parser.add_argument('--w-bit', type=int, default=8, help='w-bit')
parser.add_argument('--FPGA', action='store_true', help='FPGA')
parser.add_argument('--reorder', action='store_true', help='reorder')
parser.add_argument('--TN', type=int, default=8, help='TN')
parser.add_argument('--TM', type=int, default=64, help='TN')
parser.add_argument('--gray-scale', action='store_true', help='gray scale trainning')
opt = parser.parse_args()
print(opt)
with torch.no_grad():
convert()