English | [简体中文](GETTING_STARTED_cn.md) # Getting Started ## Installation For setting up the running environment, please refer to [installation instructions](INSTALL_cn.md). ## Data preparation - Please refer to [PrepareDataSet](PrepareDataSet.md) for data preparation - Please set the data path for data configuration file in ```configs/datasets``` ## Training & Evaluation & Inference PaddleDetection provides scripts for training, evalution and inference with various features according to different configure. ```bash # training on single-GPU export CUDA_VISIBLE_DEVICES=0 python tools/train.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml # training on multi-GPU export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m paddle.distributed.launch --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml # GPU evaluation export CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/faster_rcnn_r50_fpn_1x_coco.pdparams # Inference python tools/infer.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml --infer_img=demo/000000570688.jpg -o weights=https://paddledet.bj.bcebos.com/models/faster_rcnn_r50_fpn_1x_coco.pdparams ``` ### Other argument list list below can be viewed by `--help` | FLAG | script supported | description | default | remark | | :----------------------: | :------------: | :---------------: | :--------------: | :-----------------: | | -c | ALL | Select config file | None | **required**, such as `-c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml` | | -o | ALL | Set parameters in configure file | None | `-o` has higher priority to file configured by `-c`. Such as `-o use_gpu=False` | | --eval | train | Whether to perform evaluation in training | False | set `--eval` if needed | | -r/--resume_checkpoint | train | Checkpoint path for resuming training | None | such as `-r output/faster_rcnn_r50_1x_coco/10000` | | --slim_config | ALL | Configure file of slim method | None | such as `--slim_config configs/slim/prune/yolov3_prune_l1_norm.yml` | | --use_vdl | train/infer | Whether to record the data with [VisualDL](https://github.com/paddlepaddle/visualdl), so as to display in VisualDL | False | VisualDL requires Python>=3.5 | | --vdl\_log_dir | train/infer | VisualDL logging directory for image | train:`vdl_log_dir/scalar` infer: `vdl_log_dir/image` | VisualDL requires Python>=3.5 | | --output_eval | eval | Directory for storing the evaluation output | None | such as `--output_eval=eval_output`, default is current directory | | --json_eval | eval | Whether to evaluate with already existed bbox.json or mask.json | False | set `--json_eval` if needed and json path is set in `--output_eval` | | --classwise | eval | Whether to eval AP for each class and draw PR curve | False | set `--classwise` if needed | | --output_dir | infer | Directory for storing the output visualization files | `./output` | such as `--output_dir output` | | --draw_threshold | infer | Threshold to reserve the result for visualization | 0.5 | such as `--draw_threshold 0.7` | | --infer_dir | infer | Directory for images to perform inference on | None | One of `infer_dir` and `infer_img` is requied | | --infer_img | infer | Image path | None | One of `infer_dir` and `infer_img` is requied, `infer_img` has higher priority over `infer_dir` | | --save_results | infer | Whether to save detection results to file | False | Optional ## Examples ### Training - Perform evaluation in training ```bash export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m paddle.distributed.launch --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml --eval ``` Perform training and evalution alternatively and evaluate at each end of epoch. Meanwhile, the best model with highest MAP is saved at each epoch which has the same path as `model_final`. If evaluation dataset is large, we suggest modifing `snapshot_epoch` in `configs/runtime.yml` to decrease evaluation times or evaluating after training. - Fine-tune other task When using pre-trained model to fine-tune other task, pretrain\_weights can be used directly. The parameters with different shape will be ignored automatically. For example: ```bash export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 # If the shape of parameters in program is different from pretrain_weights, # then PaddleDetection will not use such parameters. python -m paddle.distributed.launch --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml \ -o pretrain_weights=output/faster_rcnn_r50_1x_coco/model_final \ ``` ##### NOTES - `CUDA_VISIBLE_DEVICES` can specify different gpu numbers. Such as: `export CUDA_VISIBLE_DEVICES=0,1,2,3`. - Dataset will be downloaded automatically and cached in `~/.cache/paddle/dataset` if not be found locally. - Pretrained model is downloaded automatically and cached in `~/.cache/paddle/weights`. - Checkpoints are saved in `output` by default, and can be revised from `save_dir` in `configs/runtime.yml`. ### Evaluation - Evaluate by specified weights path and dataset path ```bash export CUDA_VISIBLE_DEVICES=0 python -u tools/eval.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml \ -o weights=https://paddledet.bj.bcebos.com/models/faster_rcnn_r50_fpn_1x_coco.pdparams ``` The path of model to be evaluted can be both local path and link in [MODEL_ZOO](../MODEL_ZOO_cn.md). - Evaluate with json ```bash export CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml \ --json_eval \ -output_eval evaluation/ ``` The json file must be named bbox.json or mask.json, placed in the `evaluation/` directory. ### Inference - Output specified directory && Set up threshold ```bash export CUDA_VISIBLE_DEVICES=0 python tools/infer.py -c configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.yml \ --infer_img=demo/000000570688.jpg \ --output_dir=infer_output/ \ --draw_threshold=0.5 \ -o weights=output/faster_rcnn_r50_fpn_1x_coco/model_final \ --use_vdl=True ``` `--draw_threshold` is an optional argument. Default is 0.5. Different thresholds will produce different results depending on the calculation of [NMS](https://ieeexplore.ieee.org/document/1699659). ## Deployment Please refer to [depolyment](../../deploy/README_en.md) ## Model Compression Please refer to [slim](../../configs/slim/README_en.md)