Skip to content
This repository has been archived by the owner on Dec 15, 2024. It is now read-only.

Latest commit

 

History

History
167 lines (133 loc) · 7.91 KB

CHANGE.md

File metadata and controls

167 lines (133 loc) · 7.91 KB
Raw

CHANGE

v0.5.3

  • New features
    1. perf(sampler): when in distribued, set train_sampler be shuffle=True. 80f35daee50e78b
    2. perf(train.py): add device setting for single GPU training. 932e466f5
    3. perf(model): update DDP(model) use. c25110795c
  • Bug fixes
  • Breaking changes.

v0.5.2

  • New features
    1. build(python): update requirements.txt. aa196099a
  • Bug fixes
  • Breaking changes.

v0.5.1

  • New features
    1. build(python): update setup.py INSTALL_REQUIRES. 3ce08a61b1
    2. fix(setup.py): fix INSTALL_REQUIRES settings. 03050ebc
  • Bug fixes
  • Breaking changes.

v0.5.0

  • New features
    1. perf(train): convert nvidia/apex to torch.cuda.apex. d165dbc23ad
  • Bug fixes
    1. fix(tools): update train.sh and eval.sh. 7ca74fc5d86
  • Breaking changes.

v0.4.5

  • New features
    1. fix(transform): fix Normalize mean/std with max_value use. 51fc80d2
  • Bug fixes
  • Breaking changes.

v0.4.4

  • New features
    1. fix(cccf.py): fix class_path load. 3ef83e8a7f
  • Bug fixes
  • Breaking changes.

v0.4.3

  • New features
    1. perf(dataset): update torchvision.dataset.ImageFolder and build.py use. e95f4991
  • Bug fixes
  • Breaking changes.

v0.4.2

  • New features
    1. feat(transform): add max_value param in _C.TRANSFORM.NORMALIZE. f51f304
  • Bug fixes
    1. fix(criterion): fix SoftTargetCrossEntropy use. 04cd13c
  • Breaking changes.

v0.4.1

  • New features
    1. feat(config): add _C.TRAIN.CALCULATE_ACCURACY. 97f3c791
  • Bug fixes
    1. fix(trainer.py): fix mixup usage. 1140fe47
  • Breaking changes.

v0.4.0

  • New features
    1. perf(misc.py): update resume() to misc.py. 4f417e194
    2. perf(parser.py): remove --resume settings. 710791e
    3. feat(tools): add eval.sh. 545e4e4
    4. feat(optimizer): add Adam. 7cd0a3a2
  • Bug fixes
    1. fix(train.py): make mixup_fn to train() and update resume() usage. b5be2ed3
  • Breaking changes.

v0.3.0

  • New features
    1. style(models): update all use. 0467e383
  • Bug fixes
    1. fix(soft_target_cross_entropy_loss.py): set targets to one-hot code. 08352474
  • Breaking changes
    1. refactor(zcls2): refactor criterion/optimizer/lr_scheduler use. c740025

v0.2.3

  • New features
    1. feat(model): add EfficientNetLite support. 84172a15
    2. feat(transform): add Mixup + Cutmix support. 902a102e
    3. feat(criterion): add label_smoothing support. 6b4eb4eb
    4. feat(model): add EfficientNet. aaebc5b6
    5. perf(multi_step_lr.py): add cfg.LR_SCHEDULER.MULTISTEP_LR.STEP_SIZE. 2494defb
    6. feat(optimizer): add RMSPRop optimizer. b3669030
  • Bug fixes
  • Breaking changes

v0.2.2

  • New features
    1. build(python): update timm version and change >= to ~=. 80cd4a38
    2. perf(train.sh): make CUDA_VISIBLE_DEVICES and master_port configurable. c5385bf7
  • Bug fixes
    1. fix(model): fix ghostnet replacing the forward function failed. 48794a1d
    2. fix(cccf): convert gray img to three-channel. 96fb0fe3
    3. fix(infer): fix wrong use param i. 820b4e2d
    4. fix(trainer): fix wrong use param i. d503e227
  • Breaking changes

v0.2.1

  • New features
    1. feat(dataset): new Dataset CCCF. d02b2f66
      1. The CCCF is a custom mixed classification dataset
      2. Including CIFAR100/CUB-200-2011/Caltech-101/Food-101
    2. feat(configs): add cfg.TRAIN.TOP_K supports and reformat prec@k log. 7877607f
    3. chore(benchmarks): add mobilenetv2 for cifar10/cifar100/fashionmnist. ef753ed5
  • Bug fixes
  • Breaking changes

v0.2.0

  • New features
    1. Update benchmarks (Apex vs. ZCls2)
  • Bug fixes
  • Breaking changes
    1. Change model output type. 899c725

v0.1.1

  • New features
    1. Adjust zcls2 python version constraints (python >= 3.9 to python >= 3.8)
  • Bug fixes
  • Breaking changes.

v0.1.0

  • New features
    1. New-built training module, default supported (derived from nvidia/apex)
      1. Distributed training
      2. Mixed-precision training
      3. Linear warmup
      4. Data prefetcher
    2. New-built model module, support resnet(torchvision)/mobilenet(torchvision)/ghostnet(timm)
    3. New-built criterion module, support CrossEntroyLoss/LargeMarginSoftmaxLoss
    4. New-built optimizer module, support SGD
    5. New-built lr_scheduler module, support MultiStepLR/CosineAnnealingLR
    6. New-built dataset module, support
      1. CIFAR10/CIFAR100
      2. GeneralDataset/GeneralDatasetV2
      3. MPDataset
    7. New-built transform module, support (derived from torchvision)
      1. Normal transform (ConvertImageDtype/Normalize/ToPILImage/ToTensor)
      2. Color transform (ColorJitter/Grayscale/RandomAutocontrast/RandomAutocontrast/RandomAdjustSharpness/RandomErasing/RandomPosterize)
      3. Geometric transform (CenterCrop/RandomCrop/RandomHorizontalFlip/RandomVerticalFlip/RandomRotation/RandomResizedCrop/Resize)
      4. Augment (AutoAugment/RandAugment)
      5. Custom (OpenCVResize/SquarePad)
    8. New-built config module (derived from ZJCV/ZCls)
    9. New-built logging module (derived from ZJCV/ZCls)
  • Bug fixes
  • Breaking changes.