Skip to content

Latest commit

 

History

History
executable file
·
86 lines (69 loc) · 4.89 KB

README.md

File metadata and controls

executable file
·
86 lines (69 loc) · 4.89 KB
Raw

FastClassification

FastClassification is a tensorflow toolbox for class classification. It provides a training module with various backbones and training tricks towards state-of-the-art class classification.

features
backbone ResNet50,ResNet101,ResNet152, EfficientNetB0,EfficientNetB1,..., EfficientNetB7
augment
  • mixup
  • cutmix
  • rand_augment
  • auto_augment
tricks
  • Sharpness-Aware Minimization(SAM)
  • progressive-resizing
  • lr_finder
  • mixed-precision
  • warmup
  • concat-max-and-average-pool
  • label-smoothing
  • dataset-sample-ratio
other
  • confusion_matrix
  • data sanity check
  • class imbalance check
  • Precision
  • recall
  • show wrong prediction images
  • tensorboard

Update Log

[2021-06-26]

  • Add classic models and training tricks.

chess_p5_detection.png chess_p5_detection.png chess_p5_detection.png chess_p5_detection.png chess_p5_detection.png

Installation

1. Clone project

git clone https://github.com/wangermeng2021/FastClassification.git
cd FastClassification

2. Install environment

  • Install tesnorflow (skip this step if it's already installed,test environment:tensorflow 2.4.0)
  • Install dependencies: pip install -r requirements.txt

Training:

  • Set the learning rate manually: 
    python train.py --init-lr 1e-3 --progressive-resizing 224 224 --dataset-dir dataset/catdog --epochs 10 --batch-size 32 --augment cutmix --weights imagenet
  • For training with Sharpness-Aware Minimization optimizer(support SAM-SGD,SAM-Adam): 
    python train.py --optimizer SAM-SGD --augment baseline --init-lr 1e-3 --progressive-resizing 224 224 --dataset-dir dataset/catdog --epochs 10 --batch-size 32  --weights imagenet
  • For training with lr-finder: 
    python train.py --init-lr 0 --progressive-resizing 224 224 --dataset-dir dataset/catdog --epochs 10 --batch-size 32 --augment cutmix --weights imagenet
  • For training with progressive-resizing:[(128,128),[224,224]] 
    python train.py --progressive-resizing 128 128 224 224  --dataset-dir dataset/catdog  --epochs 10 --batch-size 32 --augment cutmix --weights imagenet
  • For training with mixed-precision: 
    python train.py --mixed-precision True --init-lr 1e-3 --progressive-resizing 224 224 --dataset-dir dataset/catdog --epochs 10 --batch-size 32 --augment cutmix --weights imagenet
  • For training with rand_augment: 
    python train.py --augment rand_augment --init-lr 1e-3 --progressive-resizing 224 224 --dataset-dir dataset/catdog --epochs 10 --batch-size 32  --weights imagenet

Tensorboard visualization:

  • Navigate to http://0.0.0.0:6006: you need to manually enable: "Setting"-->"Reload data" on tensorboard home page to automatically update data

Evaluation results on toy catdog:

(1000 pictures,train:valid=0.7:0.3,EfficientB0,epochs=10,batchsize=32)

model val_acc
progressive-resizing(224x224)+SGD+baseline 0.987
progressive-resizing(224x224)+SAM-SGD+baseline_augment 0.990
progressive-resizing(224x224)+SAM-SGD+mixup 0.947
progressive-resizing(224x224)+SAM-SGD+cutmix 0.927
progressive-resizing(224x224)+SAM-SGD+auto_augment 0.980
progressive-resizing(224x224)+SAM-SGD+rand_augment 0.960
progressive-resizing(224x224)+SAM-SGD+baseline+mixed-precision 0.980
progressive-resizing(224x224)+SAM-SGD+baseline+lr-finder 0.907
progressive-resizing(128x128,224x224)+SAM-SGD+baseline 0.963
progressive-resizing(128x128,224x224)+SAM-SGD+mixup 0.936

References