v0.3.0

Apr 22, 2023

1. Highlight the usage of configs in README.md. We recommend writing arguments into a config file.
2. Fix: Support automatic mixed precision (amp) during evaluation.
3. Support log plotting. See qtcls/utils/plot_utils.py.
4. Support using args.image_size to set the input image size.
5. Support using args.model_kwargs to set the model arguments in dict format. Here is an example in configs/_demo_.py. args.model_kwargs will be passed in qtcls/models/__init__.py, line 57.
6. Minor adjustments.

v0.2.0

Apr 10, 2023

1. Enrich the architectures of Swin Transformer and Vision Transformer (timm).
2. Support new models: Swin Transformer V2, CaiT, LeViT, MLP-Mixer and PoolFormer.
3. Optimize cosine lr scheduler. Now, the learning rate can be changed not only epoch by epoch, but also step by step.
4. Starting from this version (v0.2.0), if you want to use your config file, set --config / -c to your config file path,
   like configs/_demo_.py. Note that --config supports any file system path. Please see configs for more details.
5. Give hints for resolving some exceptions/errors.
6. Minor bug fixes and improvements.

v0.1.0

Apr 3, 2023

1. Support almost all popular datasets and models.
2. Support single-gpu and multi-gpu (DDP) training.
3. Support automatic mixed precision training.
4. Fully functional workflow.
5. Highly customizable pipeline.