README_Classification.md

Image Classification

Training and Testing on the ImageNet

This repository supports training and testing of following models:

• DiCENet
• ESPNetv2 (CVPR'19)
• ShuffleNetv2 (ECCV'18)

Dataset preparation

For dataset preparation, please follow the instuctions given here.

Training

For smaller models (< 80 MFLOPs), we used a scale augmentation of (0.2, 1.0) while for other models, we used a scale augmentation of (0.08, 1.0)

For training, you can use below command

CUDA_VISIBLE_DEVICES=0,1,2,3 python train_classification.py --s $s --model <model> --scheduler <scheduler> --data <data-path> --dataset imagenet --scale <scale>

Example for DiCENet with a network width scaling factor of 0.2 model is shown below.

CUDA_VISIBLE_DEVICES=0,1,2,3 python train_classification.py --s 0.2 --model dicenet --scheduler hybrid --epochs 120 --clr-max 61 --data ./imagenet --scale 0.2 1.0 

Testing

Testing can be done in two ways:

• The first option loads the weights automatically from weight dictionary defined in /model/weight_locations and you can use below command to test the models

CUDA_VISIBLE_DEVICES=0 python test_classification.py --model dicenet --s 0.2 --dataset imagenet --data <imagenet-loc>

• The second option allows you to specify the location of a pretrained weights file, as shown below

CUDA_VISIBLE_DEVICES=0 python test_classification.py --model dicenet --s 0.2 --dataset imagenet --data <imagenet-loc> --weights <weights-loc

Training and Testing on the MS-COCO (multi-label classification)

If you want to use the MS-COCO dataset, change the --dataset imagenet to --dataset coco in above commands. Obviously, you need to provide the dataset location too using the --data flag.