Written by Alex Choi, Oct. 8, 2020.

Basic Environment

Creating CONDA Environment

Run a command line tool, i.e. Anaconda Prompt and type the following command:
```
conda create -n centernet python=3.7
```
NOTE: Python 3.8 not tested yet. But you can try if you want.
Activate centernet Anaconda environment:
```
conda activate centernet
```

Install necessary python packages:

python -m pip install opencv-python Cython numba progress matplotlib easydict scipy

Clone cocoapi tools git repository to any path where you want
```
git clone https://github.com/cocodataset/cocoapi.git
```

Open cocoapi/PythonAPI/setup.py using a text editor and modify the following line:

extra_compile_args=['-Wno-cpp', '-Wno-unused-function', '-std=c99']

to

extra_compile_args={'gcc': ['/Qstd=c99']},

In the command line tool move to the following path:
```
cd cocoapi/PythonAPI
```
Build cocoapi tools:
```
python setup.py build_ext install
```

Using a text editor open up C:/ProgramData/Anaconda3/envs/centernet/Lib/site-packages/torch/utils/cpp_extension.py and modify the following line:

match = re.search(r'(\d+)\.(\d+)\.(\d+)', compiler_info.decode().strip())

to

match = re.search(r'(\d+)\.(\d+)\.(\d+)', compiler_info.decode("utf8", "ignore").strip())

Copy "CenterNet.zip" file to any path you want (maybe COI project folder).
Unzip the file using BandiZip or any unzipping software you want. NOTE: You can find more information from the PyTorch CenterNet official repo.

'nms' is short for "Non-Maximum Suppression." CenterNet does not usually use Non-Maximum Suppression, but it is sometimes useful.
In order to avoid build error comment out the following line after opening CenterNet/src/lib/external/setup.py up using a text editor.
```
#extra_compile_args=["-Wno-cpp", "-Wno-unused-function"]
```
NOTE: The line above may be already commented out actually.
Build nms with the following command line:
```
python setup.py  build_ext --inplace
```
NOTE: You can find more information about building nms here.

Move to the following path in your CenterNet path:
```
cd CenterNet/src/lib/models/networks/DCNv2/
```
Build DCNv2 using the following command line (this may take some time to finish building):
```
python setup.py build develop
```
NOTE: You can find more information about DCNv2 here.

Unzip "ImageNet-Weights.zip" and copy all the contained files (ImageNet pre-trained models) to C:\Users\{WINDOWS_USER_ACCOUNT_NAME}\.cache\torch\checkpoints\.
Unzip "Centernet-Models.zip" and copy all the contained files to CenterNet/models/.

Finally, test with some COI images using the following command line:

python demo.py ctdet --demo ../data/COI/images/image_0001.jpg --load_model ../exp/ctdet/COI/model_best.pth

python demo.py ctdet --demo ../data/COI/images/image_0002.jpg --load_model ../exp/ctdet/COI/model_best.pth

python demo.py ctdet --demo ../data/COI/images/image_0003.jpg --load_model ../exp/ctdet/COI/model_best.pth

If you are using CenterNet.zip you can find the training information for COI from CenterNet/exp/ctdet/COI/.

There are 2 options for training:

One is training from scratch (random initialization).

python main.py ctdet --exp_id COI --batch_size 16 --lr 1.25e-4 --gpus 0 --load_model ../models/ctdet_coco_resdcn18.pth --resume False

The other is resume training from pre-trained model.

python main.py ctdet --exp_id COI --batch_size 16 --lr 1.25e-4 --gpus 0 --load_model ../models/ctdet_coco_resdcn18.pth --resume True

You can check your training information such as loss history, accuracy history and so on using TensorBoard.
You can find your training information for COI from the path, CenterNet/exp/ctdet/COI/
Install tensorboard using the following command line:
```
python -m pip install tensorboard
```