ENOTDataSphere
This is a repository with ENOT framework usage examples in Yandex DataSphere.
Install
- Go to "install" folder;
- Open Install.ipynb ;
- Execute all the cells one by one;
- Now you can use all Jupyter Notebooks from this repository;
To reproduce results from the webinar:
- Create your own project in Yandex DataSphere - https://cloud.yandex.ru/services/datasphere
- Clone this repository into your DataSphere project (Git -> Clone);
- Download models from the link below, unzip file(Snippets -> Extract ZIP file.py)
and put it in the root directory of the project(so the path to models should be './ENOTDataSphere/models/base_model.pth'): https://yadi.sk/d/nbaV1N1tQMSPpg - Download test dataset from the link below, unzip it(Snippets -> Extract ZIP file.py)
and create "data" folder in "mmdet_tools" folder and copy unziped folder('masks') in './ENOTDataSphere/mmdet_tools/data/'(so the path to your data should be './ENOTDataSphere/mmdet_tools/data/'): https://yadi.sk/d/MwN9o5LmLi5Cvg - Download test video from this link, create folder "video" and put downloaded video there: https://yadi.sk/i/hE05IF9-OEwvKg
Now you can use all notebooks in this repository
- demo_video.ipynb - demo of masks detection on video. You can download your own video, change path to the video and detect masks;
- pretrain_search.ipynb - in this notebook you can find examples of enot_pretrain, enot_search phases and get test metrics on the dataset;
- train_baseline.ipynb - example of train baseline model with MMDetection;
Train detector on custom dataset with MMDET+ENOT
To train your custom dataset for detection and use optimization framework ENOT you should:
1 - Prepare your dataset to COCO annotation format.
- You should have 3 .json files, like . train.json, test.json, val.json . and 3 folders(train, test, val) with images. Make a folder(like 'my_dataset_name') and copy all these files into it. About COCO annotation format you can read here - https://mmdetection.readthedocs.io/en/latest/2_new_data_model.html
- In 'mmdet_tools' make a folder 'data' and copy your directory with dataset('my_dataset_name') to 'data'.
2 - Prepare configs for ENOT_Pretrain, ENOT_search and tune phases.
- In 'mmdet_tools/configs/base/datasets' in files 'mask_face.py' and 'mask_face_pretrain.py' set to variable 'data_root' path to your dataset. Set 'ann_file' parameter for train, test and val dictionaries with paths to yours .json files(train.json, test.json, val.json)
- If you want to change searchspace you should do it in 'mmdet_tools/configs/base/models/search_space_ssd_masks.py'
- For enot_pretrain phase you should change paths to your dataset in 'mmdet_tools/configs/wider_face/search_space_ssd_masks.py', also you can set here augmentations you want.
- For enot_search phase you should change paths to your dataset in 'mmdet_tools/configs/wider_face/search_space_ssd_masks_search.py', dataset and augmentations should be the same like in pretrain phase!
Detailed information about MMDetection config system you can find here - https://mmdetection.readthedocs.io/en/latest/tutorials/config.html
3 - run enot_pretrain, enot_search phases
To start enot_pretrain you should:
- Prepare config(see 2);
- In 'mmdet_tools/run_enot_pretrain.py' in pretrain_cfg dictionary you should change: 'epochs' - number of epochs, 'mmdet_config_path' - path to 'mmdet_tools/configs/base/models/search_space_ssd_masks.py', 'experiment_dir' - path to save checkpoints. In 'enot_pretrain' function you can change type of optimizer, learning rate, set scheduler, and batch size;
- When all configs, learning procedure and paths in 'mmdet_tools/run_enot_pretrain.py' are ready, from jupyter notebook just call 'run_enot_pretrain' function from 'mmdet_tools/run_enot_pretrain.py';
- Prepare config(see 2);
- Choose best checkpoint from pretrain phase;
- In 'mmdet_tools/run_enot_search.py' in pretrain_cfg dictionary you should change: 'epochs' - number of epochs, 'mmdet_config_path' - path to 'mmdet_tools/configs/base/models/search_space_ssd_masks.py', 'experiment_dir' - path to save checkpoints, 'pretrain_checkpoint_path' - path to best checkpoint from pretrain phase. In 'enot_search' function you can change type of optimizer, learning rate, set scheduler, and batch size;
- When all configs, learning procedure and paths in 'mmdet_tools/run_enot_search.py' are ready, from jupyter notebook just call 'run_enot_search' function from 'mmdet_tools/run_enot_search.py'. If you can set up parameter 'latency_loss_weight' to vary complexity of model to find, bigger 'latency_loss_weight' - more lightweight model you will find;
4 - run tune found model
- Prepare config 'mmdet_tools/configs/base/models/mobilenet_from_search_space_mask.py', set indices of best models from search phase. And 'mmdet_tools/configs/wider_face/mobilenet_from_search_space_mask.py' here you should set path to dataset, augmentations, optimizer, learning rate;
- Export weights(.pth file) of the best model. Example you can see in 'mmdet_tools/build_searched_arch_ckpt.py';
- Call train from 'mmdet_tools/tools/train_baseline.py'. Set parameter 'config' - path to 'mmdet_tools/configs/wider_face/mobilenet_from_search_space_mask.py', 'full_pretrained_model' - path to exported .pth file;