v0.9.1

Jun 7, 2024

1. Development environment -> python 3.8 & pytorch 1.13.1+cu117
2. Clarify the compatible version relationship between Python and PyTorch. See README.md - Installation.
3. Minor adjustments.

v0.9.0

Mar 6, 2024

1. Fix a compatibility bug in accuracy(output, target, topk=(1,)) (see qtcls/utils/misc.py Line 301) when computing the Top-5 Acc.
2. Add simple augmentations for training ImageNet with CNNs.
3. Introduce a new command-line argument --simple_aug which serves as an option of using the simple augmentations for ImageNet.
4. The --note argument is now appended to the pre-defined meta note (see main.py Line 115, 260~263). In previous versions, the --note was directly used without a meta note.
5. Update requirements.txt.
6. Update docs.

v0.8.1

Feb 5, 2024

1. Fix a bug in swin_transformer.py where the input image size cannot be customized.
2. A tuple is also acceptable for _image_size in qtcls/datasets/__init__.py.
3. Add a hint about '-p' when the pretraining path is not found.

v0.8.0

Dec 2, 2023

1. Introduce a command-line argument --flops, supporting model FLOPs computation and display.
2. Introduce a command-line argument --pretrain/-p. To temporarily use a pre-trained weight path, you can specify it by command-line argument --pretrain / -p. For long-term use of a pre-trained weight path, it is preferable to write it in _pretrain_.py. Loading Priority: --pretrain > model_local_paths > model_urls. (Please carefully distinguish between --pretrain/-p and --resume/-r.)
3. Enhance support for loading pre-trained weights on timm models. Loading Priority: --pretrain > model_local_paths > timm.create_model(..., pretrained=True).
4. Introduce a command-line argument --clear_output_dir/-co, with which the specified --output_dir will be cleared first.
5. In qtcls/datasets/__init__.py, num_classes -> _num_classes. Add _image_size where users can optionally specify the dataset's default target image size. During data preprocessing, images will be automatically scaled to the target image size. You can also specify the target image size by --image_size (priority: --image_size > _image_size[dataset_name]).
6. Fix the bug where the image size cannot adapt to changes when using fake data (--dummy) of different datasets such as CIFAR (32×32) and ImageNet (224×224).
7. Redefine imagenet.py and add more guidelines.
8. Rename plot_utils.py to log_plot.py.
9. Minor bug fixes and improvements.
10. Update docs.

v0.7.0

Sep 18, 2023

1. Starting from v0.7.0 (this version), command-line arguments after --config xxx / -c xxx override the config arguments if the name is duplicated. E.g., python main.py -c configs/_demo_.py -d cifar100 -b 64.
2. Support fake datasets. Please use --dummy if you don't have the specified --dataset file on hand but want to test your code. E.g., python main.py -m resnet50 -d imagenet1k --dummy.
3. Support the FashionMNIST dataset.
4. Update schedulers.
5. Rearrange configs/_demo_.py.
6. Standardize the code & Refine the code structure.
7. Update docs.
8. Bug fixes and improvements.

v0.6.1

Jun 29, 2023

1. Standardize the code.
2. Add author information under the Apache-2.0 License.

v0.6.0

Jun 16, 2023

1. Support checking for version updates by executing check_for_updates.py.
2. Support a new dataset: Food101.
3. Fix bugs in qtcls/datasets.
4. Add a description of folder format datasets in README.md and data/README.md.
5. Information adjustments and improvements.

v0.5.1

Jun 8, 2023

Fix compatibility issues.

v0.5.0

May 25, 2023

1. Support new datasets: Flowers102 and Stanford Cars.
2. Support new models: DeiT, PVT, TNT and Twins.
3. Add some practical module wrappers for nn.Module and nn.ModuleList.
4. Update configs/_demo_.py.
5. Rename the default --model_lib from torchvision-ex to default.
6. Remove redundant code.
7. Update docs.

v0.4.0

Apr 24, 2023

1. Add strong data augmentations.
2. Support both single image augmentations and batch augmentations.
3. Support the argument in_chans for all models. That is, you can now specify the number of channels of the input image for the model.
4. Support imagenet21k (also called imagenet22k).