Skip to content

Latest commit

 

History

History
216 lines (175 loc) · 7.12 KB

RELEASE.md

File metadata and controls

216 lines (175 loc) · 7.12 KB

Release Note

0.3.0 (2024/1/17)

Release 0.3.0 is published. We will drop MindSpore 1.x in the future release.

  1. New models:
  2. Features:
    • AsymmetricLoss & JSDCrossEntropy
    • Augmentations Split
    • Customized AMP
  3. Bug fixes:
    • Since the classifier weights are not fully deleted, you may encounter an error passing in the num_classes when creating a pre-trained model.
  4. Refactoring:
    • The names of many models have been refactored for better understanding.
    • Script of VisionTransformer.
    • Script of Mixed(PyNative+jit) mode training.
  5. Documentation:
    • A guide of how to extract multiscale features from backbone.
    • A guide of how to finetune the pre-trained model on a custom dataset.
  6. BREAKING CHANGES:
    • We are going to drop support of MindSpore 1.x for it's EOL.
    • Configuration filter_bias_and_bn will be removed and renamed as weight_decay_filter, due to a prolonged misunderstanding of the MindSpore optimizer. We will migrate the existing training recipes, but the signature change of function a will be incompatible and the old version training recipes will also be incompatible. See PR/752 for details.

0.2.2 (2023/6/16)

  1. New version 0.2.2 is released! We upgrade to support MindSpore v2.0 while maintaining compatibility of v1.8
  2. New models:
  3. New features:
    • Gradient Accumulation
    • DynamicLossScale for customized TrainStep
    • OneCycleLR and CyclicLR learning rate scheduler
    • Refactored Logging
    • Pyramid Feature Extraction
  4. Bug fixes:
    • Serving Deployment Tutorial(mobilenet_v3 doesn't work on ms1.8 when using Ascend backend)
    • Some broken links on our documentation website.

0.2.1

  • 2023/6/2
  1. New version: 0.2.1 is released!
  2. New documents is online!
  • 2023/5/30
  1. New Models:
  2. New Features:
    • 3-Augment, Augmix, TrivialAugmentWide
  3. Bug Fixes:
    • ViT pooling mode
  • 2023/04/28
  1. Add some new models, listed as following
  2. Bug fix:
    • Setting the same random seed for each rank
    • Checking if options from yaml config exist in argument parser
    • Initializing flag variable as Tensor in Optimizer Adan

0.2.0

  • 2023/03/25
  1. Update checkpoints for pretrained ResNet for better accuracy
    • ResNet18 (from 70.09 to 70.31 @Top1 accuracy)
    • ResNet34 (from 73.69 to 74.15 @Top1 accuracy)
    • ResNet50 (from 76.64 to 76.69 @Top1 accuracy)
    • ResNet101 (from 77.63 to 78.24 @Top1 accuracy)
    • ResNet152 (from 78.63 to 78.72 @Top1 accuracy)
  2. Rename checkpoint file name to follow naming rule ({model_scale-sha256sum.ckpt}) and update download URLs.
  • 2023/03/05
  1. Add Lion (EvoLved Sign Momentum) optimizer from paper https://arxiv.org/abs/2302.06675
    • To replace adamw with lion, LR is usually 3-10x smaller, and weight decay is usually 3-10x larger than adamw.
  2. Add 6 new models with training recipes and pretrained weights for
  3. Support gradient clip
  4. Arg name use_ema changed to ema, add ema: True in yaml to enable EMA.

0.1.1

  • 2023/01/10
  1. MindCV v0.1 released! It can be installed via PyPI pip install mindcv now.
  2. Add training recipe and trained weights of googlenet, inception_v3, inception_v4, xception

0.1.0

  • 2022/12/09
  1. Support lr warmup for all lr scheduling algorithms besides cosine decay.
  2. Add repeated augmentation, which can be enabled by setting --aug_repeats to be a value larger than 1 (typically, 3 or 4 is a common choice).
  3. Add EMA.
  4. Improve BCE loss to support mixup/cutmix.
  • 2022/11/21
  1. Add visualization for loss and acc curves
  2. Support epochwise lr warmup cosine decay (previous is stepwise)
  • 2022/11/09
  1. Add 7 pretrained ViT models.
  2. Add RandAugment augmentation.
  3. Fix CutMix efficiency issue and CutMix and Mixup can be used together.
  4. Fix lr plot and scheduling bug.
  • 2022/10/12
  1. Both BCE and CE loss now support class-weight config, label smoothing, and auxiliary logit input (for networks like inception).

0.0.1-beta

  • 2022/09/13
  1. Add Adan optimizer (experimental)

MindSpore Computer Vision 0.0.1

Models

mindcv.models now expose num_classes and in_channels as constructor arguments:

  • Add DenseNet models and pre-trained weights
  • Add GoogLeNet models and pre-trained weights
  • Add Inception V3 models and pre-trained weights
  • Add Inception V4 models and pre-trained weights
  • Add MnasNet models and pre-trained weights
  • Add MobileNet V1 models and pre-trained weights
  • Add MobileNet V2 models and pre-trained weights
  • Add MobileNet V3 models and pre-trained weights
  • Add ResNet models and pre-trained weights
  • Add ShuffleNet V1 models and pre-trained weights
  • Add ShuffleNet V2 models and pre-trained weights
  • Add SqueezeNet models and pre-trained weights
  • Add VGG models and pre-trained weights
  • Add ViT models and pre-trained weights

Dataset

mindcv.data now expose:

  • Add Mnist dataset
  • Add FashionMnist dataset
  • Add Imagenet dataset
  • Add CIFAR10 dataset
  • Add CIFAR100 dataset

Loss

mindcv.loss now expose:

  • Add BCELoss
  • Add CrossEntropyLoss

Optimizer

mindcv.optim now expose:

  • Add SGD optimizer
  • Add Momentum optimizer
  • Add Adam optimizer
  • Add AdamWeightDecay optimizer
  • Add RMSProp optimizer
  • Add Adagrad optimizer
  • Add Lamb optimizer

Learning_Rate Scheduler

mindcv.scheduler now expose:

  • Add WarmupCosineDecay learning rate scheduler
  • Add ExponentialDecayLR learning rate scheduler
  • Add Constant learning rate scheduler

Release

mindcv-0.0.1.apk

mindcv-0.0.1-py3-none-any.whl.sha256

mindcv-0.0.1-py3-none-any.whl