Skip to content

Latest commit

 

History

History
133 lines (97 loc) · 3.97 KB

get_started.md

File metadata and controls

133 lines (97 loc) · 3.97 KB
Raw

Getting started

We are glad that you use FCOSR. This guide mainly provides the basic usage of FCOSR. Our code has been modified based on mmdetection, so it is slightly different from the way you are used to. Please read it carefully.
For installation instructions, please see install.md.

Prepare DOTA dataset.

It is recommended to symlink the dataset root to FCOSR/data.

Here, we give an example for single scale data preparation of DOTA1.0.

First, make sure your initial data are in the following structure.

data/dota10
├── train
│   ├──images
│   └── labelTxt
├── val
│   ├── images
│   └── labelTxt
└── test
    └── images

Split the original images and create COCO format json.

python DOTA_devkit/prepare_dota.py path_to_dota path_to_split

Then you will get data in the following structure

dota10_1024
├── test1024
│   ├── DOTA_test1024.json
│   └── images
└── trainval1024
     ├── DOTA_trainval1024.json
     └── images

The arguments of prepare_dota.py in the list.

Arguments type default discribe
srcpath str None source path
dstpath str None DOTA save path.
-s, --subsize int 1024 patch size
-g, --gaps int 200 gap
-t, --type str "dota10" dota version
-p, --processor-num int 16 multiprocessor number
-d, --drop-threshold float 0.5 Drop samples threshold
--scales list[float] None None means single scale

Inference with pretrained models

Test a dataset

  • specified GPU testing (unify single GPU and multi GPU)

You can use the following commands to test a dataset.

python analyse_result.py ${CONFIG_FILE} ${CHECKPOINT_FILE}

Note: This command will generate a result file (.pkl) and DOTA detection result in a folder.

The arguments of analyse_result.py in the list.

Arguments type default discribe
config str None config file path
checkpoint str None checkpoint path
-d, --draw None False Draw result switch
-n, --draw-num str or int 'all' Draw number
-s, --source-dir str None Source origin image dir
-p, --processor-num int 16 multiprocessor number
-D, --draw-dir str None Draw result folder path
-g, --gpus str "0" GPUs, use common to separate multi devices.

Train a model

mmdetection implements distributed training and non-distributed training, which uses MMDistributedDataParallel and MMDataParallel respectively.

All outputs (log files and checkpoints) will be saved to the working directory, which is specified by work_dir in the config file.

*Important*: The default learning rate in config files is for 8 GPUs. If you use less or more than 8 GPUs, you need to set the learning rate proportional to the GPU num, e.g., 0.01 for 4 GPUs and 0.04 for 16 GPUs.

Train with a single GPU

python tools/train.py ${CONFIG_FILE}

If you want to specify the working directory in the command, you can add an argument --work_dir ${YOUR_WORK_DIR}.

Train with multiple GPUs

# GPU_order is a string, just like "0,1,2,3". You can specify the device.
./tools/dist_train.sh ${CONFIG_FILE} ${GPU_order} [optional arguments]

Optional arguments are:

  • --work_dir ${WORK_DIR}: Override the working directory specified in the config file.
  • --resume_from ${CHECKPOINT_FILE}: Resume from a previous checkpoint file.

How to use benchmark

Speed

# modify and run shell script file analyse_inference_speed.sh
./analyse_inference_speed.sh

Transfer to onnx and TensorRT

transfer path: checkpoint(.pth) -> onnx format(.onnx) -> tensorrt engine(.trt)

# checkpoint(.pth) -> onnx format(.onnx)
python tools/deployment/pytorch2onnx.py config_file checkpoint_file --output-file output_path --shape 1024 1024

onnx -> tensorrt (onnx-tensorrt 8.0)

The TensorRT inference implement can be found in my other repository