A model conversion tool written in python3, supports caffe2darknet.
This project is adapted based on pytorch-caffe-darknet-convert, hoping to support the model conversion to darknet. We look forward to your forking this project and submitting an improved pr, and improving the details of model conversion so that it can better adapt to darknet.
Activate the conda environment of Python and the corresponding library dependencies
conda create -n py37 python=3.7
conda activate py37
pip install brocolli-caffe
git clone https://github.com/KerwinKai/Caffe2Darknet.git
example
if __name__ == '__main__':
net_path = './ResNet18/resnet18.prototxt'
weight_path = './ResNet18/resnet18.caffemodel'
out_file_path = './ResNet18/resnet18'
c2d = Caffe2Darknet(net = net_path, weight = weight_path, out_file = out_file_path)
c2d.convert()
Loading caffemodel: ./ResNet18/resnet18.caffemodel
0 conv1 Convolution
1 bn1_bn BatchNorm
2 bn1 Scale
3 relu ReLU
4 maxpool Pooling
5 layer1_0_conv1 Convolution
6 layer1_0_bn1_bn BatchNorm
7 layer1_0_bn1 Scale
8 layer1_0_relu ReLU
9 layer1_0_conv2 Convolution
10 layer1_0_bn2_bn BatchNorm
11 layer1_0_bn2 Scale
12 add Eltwise
14 layer1_1_conv1 Convolution
15 layer1_1_bn1_bn BatchNorm
16 layer1_1_bn1 Scale
17 layer1_1_relu ReLU
18 layer1_1_conv2 Convolution
19 layer1_1_bn2_bn BatchNorm
20 layer1_1_bn2 Scale
21 add_1 Eltwise
23 layer2_0_conv1 Convolution
24 layer2_0_bn1_bn BatchNorm
25 layer2_0_bn1 Scale
26 layer2_0_relu ReLU
27 layer2_0_conv2 Convolution
28 layer2_0_bn2_bn BatchNorm
29 layer2_0_bn2 Scale
30 layer2_0_downsample_0 Convolution
31 layer2_0_downsample_1_bn BatchNorm
32 layer2_0_downsample_1 Scale
33 add_2 Eltwise
35 layer2_1_conv1 Convolution
36 layer2_1_bn1_bn BatchNorm
37 layer2_1_bn1 Scale
38 layer2_1_relu ReLU
39 layer2_1_conv2 Convolution
40 layer2_1_bn2_bn BatchNorm
41 layer2_1_bn2 Scale
42 add_3 Eltwise
44 layer3_0_conv1 Convolution
45 layer3_0_bn1_bn BatchNorm
46 layer3_0_bn1 Scale
47 layer3_0_relu ReLU
48 layer3_0_conv2 Convolution
49 layer3_0_bn2_bn BatchNorm
50 layer3_0_bn2 Scale
51 layer3_0_downsample_0 Convolution
52 layer3_0_downsample_1_bn BatchNorm
53 layer3_0_downsample_1 Scale
54 add_4 Eltwise
56 layer3_1_conv1 Convolution
57 layer3_1_bn1_bn BatchNorm
58 layer3_1_bn1 Scale
59 layer3_1_relu ReLU
60 layer3_1_conv2 Convolution
61 layer3_1_bn2_bn BatchNorm
62 layer3_1_bn2 Scale
63 add_5 Eltwise
65 layer4_0_conv1 Convolution
66 layer4_0_bn1_bn BatchNorm
67 layer4_0_bn1 Scale
68 layer4_0_relu ReLU
69 layer4_0_conv2 Convolution
70 layer4_0_bn2_bn BatchNorm
71 layer4_0_bn2 Scale
72 layer4_0_downsample_0 Convolution
73 layer4_0_downsample_1_bn BatchNorm
74 layer4_0_downsample_1 Scale
75 add_6 Eltwise
77 layer4_1_conv1 Convolution
78 layer4_1_bn1_bn BatchNorm
79 layer4_1_bn1 Scale
80 layer4_1_relu ReLU
81 layer4_1_conv2 Convolution
82 layer4_1_bn2_bn BatchNorm
83 layer4_1_bn2 Scale
84 add_7 Eltwise
86 avgpool Pooling
87 flatten Flatten
unknown type Flatten
88 fc InnerProduct
done
Save to ./ResNet18/resnet18.weights
[net]
batch=1
channels=3
height=224
width=224
[convolutional]
filters=64
size=7
stride=2
pad=1
batch_normalize=1
activation=relu
[maxpool]
size=3
stride=2
pad=1
[convolutional]
filters=64
size=3
stride=1
pad=1
batch_normalize=1
activation=relu
[convolutional]
filters=64
size=3
stride=1
pad=1
batch_normalize=1
activation=linear
[shortcut]
from=-3
activation=relu
[convolutional]
filters=64
size=3
stride=1
pad=1
batch_normalize=1
activation=relu
[convolutional]
filters=64
size=3
stride=1
pad=1
batch_normalize=1
activation=linear
[shortcut]
from=-3
activation=relu
[convolutional]
filters=128
size=3
stride=2
pad=1
batch_normalize=1
activation=relu
[convolutional]
filters=128
size=3
stride=1
pad=1
batch_normalize=1
activation=linear
[route]
layers=-3
[convolutional]
filters=128
size=1
stride=2
pad=1
batch_normalize=1
activation=linear
[shortcut]
from=-3
activation=relu
[convolutional]
filters=128
size=3
stride=1
pad=1
batch_normalize=1
activation=relu
[convolutional]
filters=128
size=3
stride=1
pad=1
batch_normalize=1
activation=linear
[shortcut]
from=-3
activation=relu
[convolutional]
filters=256
size=3
stride=2
pad=1
batch_normalize=1
activation=relu
[convolutional]
filters=256
size=3
stride=1
pad=1
batch_normalize=1
activation=linear
[route]
layers=-3
[convolutional]
filters=256
size=1
stride=2
pad=1
batch_normalize=1
activation=linear
[shortcut]
from=-3
activation=relu
[convolutional]
filters=256
size=3
stride=1
pad=1
batch_normalize=1
activation=relu
[convolutional]
filters=256
size=3
stride=1
pad=1
batch_normalize=1
activation=linear
[shortcut]
from=-3
activation=relu
[convolutional]
filters=512
size=3
stride=2
pad=1
batch_normalize=1
activation=relu
[convolutional]
filters=512
size=3
stride=1
pad=1
batch_normalize=1
activation=linear
[route]
layers=-3
[convolutional]
filters=512
size=1
stride=2
pad=1
batch_normalize=1
activation=linear
[shortcut]
from=-3
activation=relu
[convolutional]
filters=512
size=3
stride=1
pad=1
batch_normalize=1
activation=relu
[convolutional]
filters=512
size=3
stride=1
pad=1
batch_normalize=1
activation=linear
[shortcut]
from=-3
activation=relu
[avgpool]
[Flatten]
[connected]
output=1000
activation=linear
layer filters size input output
0 conv 64 7 x 7 / 2 224 x 224 x 3 -> 112 x 112 x 64
1 max 3 x 3 / 2 112 x 112 x 64 -> 56 x 56 x 64
2 conv 64 3 x 3 / 1 56 x 56 x 64 -> 56 x 56 x 64
3 conv 64 3 x 3 / 1 56 x 56 x 64 -> 56 x 56 x 64
4 shortcut 1
5 conv 64 3 x 3 / 1 56 x 56 x 64 -> 56 x 56 x 64
6 conv 64 3 x 3 / 1 56 x 56 x 64 -> 56 x 56 x 64
7 shortcut 4
8 conv 128 3 x 3 / 2 56 x 56 x 64 -> 28 x 28 x 128
9 conv 128 3 x 3 / 1 28 x 28 x 128 -> 28 x 28 x 128
10 route 7
11 conv 128 1 x 1 / 2 56 x 56 x 64 -> 28 x 28 x 128
12 shortcut 9
13 conv 128 3 x 3 / 1 28 x 28 x 128 -> 28 x 28 x 128
14 conv 128 3 x 3 / 1 28 x 28 x 128 -> 28 x 28 x 128
15 shortcut 12
16 conv 256 3 x 3 / 2 28 x 28 x 128 -> 14 x 14 x 256
17 conv 256 3 x 3 / 1 14 x 14 x 256 -> 14 x 14 x 256
18 route 15
19 conv 256 1 x 1 / 2 28 x 28 x 128 -> 14 x 14 x 256
20 shortcut 17
21 conv 256 3 x 3 / 1 14 x 14 x 256 -> 14 x 14 x 256
22 conv 256 3 x 3 / 1 14 x 14 x 256 -> 14 x 14 x 256
23 shortcut 20
24 conv 512 3 x 3 / 2 14 x 14 x 256 -> 7 x 7 x 512
25 conv 512 3 x 3 / 1 7 x 7 x 512 -> 7 x 7 x 512
26 route 23
27 conv 512 1 x 1 / 2 14 x 14 x 256 -> 7 x 7 x 512
28 shortcut 25
29 conv 512 3 x 3 / 1 7 x 7 x 512 -> 7 x 7 x 512
30 conv 512 3 x 3 / 1 7 x 7 x 512 -> 7 x 7 x 512
31 shortcut 28
32 avg 7 x 7 x 512 -> 512
unknown type Flatten
34 connected 512 -> 1000
Hash of Darknet model has been published: fc7ccc2dc0e8966994cc28ac0841cddd
True
In the local development environment, we use pre-commit to check the code style to ensure the uniformity of the code style. Before submitting code, you can install pre-commit first, or you can submit directly, I will maintain the code style regularly.