Date format: YY/MM/DD
Major changes
- Add caching in LaSOT dataset
- Add multi-mode in siamfcpp visualization debugging tools
- demo/main/debug/visualize_siamfcpp_training_data.py
- Seperate config file for debugging: experiments/siamfcpp/data/siamfcpp_data-trn.yaml
Major changes
- Update all training configuration
- the cuda device number is defined by num_processes rather than devices in training phase
- the settings between DP and DDP are unified now
- Adopt loguru to replace logging for better logger
- Update code for distributed training
- Update code for model/pipeline/trainer
Major changes
- Update all training configuration
- learning rate multiplied by 2 in order to be compatible with PyTorch==1.4.0 & CUDA==10.1
- Update requirements.txt
- Add tinyconv training configuration
- experiments/siamfcpp/train/siamfcpp_tinyconv-trn.yaml
- Add dataloader visualization tools for SiamFC++ (SOT)
- demo/main/debug/visualize_siamfcpp_training_data.py
Major changes
- Add webcam demo
- demo/main/video/sot_video.py
Major changes
- Add training with PyTorch Distributed Data Parallel (DDP)
- main/dist_trian.py
- .../trainer_impl/distributed_regular_trainer.py
Major changes
- Add TensorboardLogger in monitor
- Recursively put engine_data as scalar into Tensorboard
Major changes
- Caching mechanism added to got10k dataset
- Build cache at root_dir/subset.pkl (default), subset=[train, val, test]
- Cache created/loaded in GOT10k.data_dict and will be queried every time __getitem__ is called
- Remove num_iterations item in training .yaml files
- replaced by value calculated with nr_image_per_epoch and minibatch
- num_iterations = nr_image_per_epoch // minibatch
- all training configs has been updated.
- replaced by value calculated with nr_image_per_epoch and minibatch
Minor changes
- pipeline.builder's builder method has been changed
- build_pipeline -> build
- Add contrib module's template
- docs/TEMPLATES/contrib_module
- Add one-shot detection demo
- API (see docs/PIPELINE_API.md
- Runnable demo
- Complete ShuffleNetV2x1.0 experiment
- Support Training and Test on GOT-10k for Single Object Tracking
- SiamFC++-AlexNet
- SiamFC++-GoogLeNet
- SiamFC++-ShuffleNetV2
- Training Components
- Dataset helper, data sampler, transformer (data augmentation), etc.
- Optimization: learing rate scheduler, dynamic freezing, etc.
- Trainer & Monitor
- Support SiamFC++ test on VOT benchmark
- SiamFC++-AlexNet
- SiamFC++-GoogLeNet