简体中文 | English
-
Linux (current code has not been tested in Windows environment)
-
python3.7 + (python2 is not supported)
-
PyTorch 1.5 or higher
-
CUDA 10.0 or higher
-
NCCL 2
-
GCC(G++) 4.9 or above
The following versions of operating system and software have been tested:
-
Operating system: Ubuntu 16.04/18.04
-
CUDA: 9.2/10.0
-
NCCL: 2.1.15/2.2.13/2.3.7/2.4.2
-
GCC (G + +): 4.9/5.3/7.3
a. Create a conda virtual environment and activate it.
conda create -n yolodet python=3.7 -y
conda activate yolodet
b. Follow official instructions to install PyTorch stable or nightly and torchvision, for example:
conda install pytorch torchvision -c pytorch
c. Clone the yolodet-pytorch library.
git clone https://github.com/wuzhihao7788/yolodet-pytorch.git
cd yolodet-pytorch
d. Install yolodet (other dependencies will be installed automatically).
-
- If you need to install, compile and install when using DCN, you can install it in the following way
python setup.py develop # or "pip install -v -e ."
-
- If you do not use DCN, you can install dependencies in the following ways
pip install -r requirements.txt
It is recommended to connect the root directory of the data set to $YOLODET/data
.
If your folder structure is different, you may need to change the corresponding path in the configuration file.
yolodet-pytorch
├── yolodet
├── tools
├── cfg
├── data
│ ├── your data root #Your data set root directory
│ │ ├── annotations #label storage location
│ │ │ ├── train.txt #Training data set label file. Data format: [picture name x1,y1,x2,y2,label] For example: 5979.jpg 253,420,406,744,0 25,40,46,44,1
│ │ │ ├── val.txt #Verify the data set label file. The data format is the same as above
│ │ │ ├── test.txt #Test data set label file. The data format is the same as above
│ │ ├── images #Picture storage location
│ │ ├── label.names #label name storage location, press label index, store by row