/ap-net

The source code for "Image Recognition with Augmentation Pathways "

Primary LanguagePythonApache License 2.0Apache-2.0

ap-net

The pytorch implementation for "Image Recognition with Augmentation Pathways".

Requirements

  1. Python 3.7
  2. PyTorch 1.3.0
  3. pytorch-randaugment

Dataset Prepare

  1. Download ImageNet dataset to folder datasets/
  2. Data Organization:
    • All images are supposed to be in datasets/imagenet/all
    • Train/Valid annotations should be in datasets/imagenet/ and named as (train|valid).txt
    • Annotation file format:
    name_of_image.jpg label_num_from_zero\n

Training

For ImageNet dataset, you can train AP-IResNet-50 from scratch with following script:

python train.py -M 9 -N 2 --desc AP-IResNet-50

Results

Model #Params MACs Augmentation Acc
iResNet-50 25.6M 4.15G Baseline 77.59/93.55
iResNet-50 25.6M 4.15G RandAugment 77.20/93.52
AP-iResNet-50 21.8M 3.95G RandAugment 78.20/93.95