Image Classification

This repository is copied from Image Classification Examples of MXNet. And I made some modifications to these examples, including removing some files that I don't need (eg. R programs) and adding some useful files. And I simplfied the README.md, if you want a more detailed description, you should read the original README.md.

Contents

1. How to prepare datasets
2. How to train
3. How to score
4. How to predict an image
5. How to fine-tune

Prepare Datasets

Assume all images are stored as individual image files such as .png or .jpg, and images belonging to the same class are placed in the same directory. All these class directories are then in the same root data directory.

• We first prepare two .lst files, which consist of the labels and image paths can be used for generating rec files.

  python im2rec.py rec/img data/ --list=True --train-ratio=0.8 --recursive=True

• Then we generate the .rec files. We resize the images such that the short edge is at least 256px.

  python im2rec.py rec/img_train data/ --resize=256
  python im2rec.py rec/img_val data/ --resize=256

• Use python im2rec.py --help to see more options.

Train

Python training programs is provided (R training programs is not provided here). Use train_*.py to train a network on a particular dataset. For example:

• Train the datasets prepared before. There is a rich set of options, one can list them by passing --help.

  python train.py --data-train rec/img_train.rec --data-val rec/img_val.rec

• When the training process is finished, we generate a training log in log/train.log and use parse_log.py to parse it into a markdown table.

   python parse_log.py log/train.log

Score

We can use score.py to score a model on a dataset.

python score.py --model models/img_cls --epoch 30 --data-val rec/img_val.rec

Predict

We can use predict.py to predict an image.

python predict.py img.jpg --model models/img_cls --epoch 30

Fine-tune

TODO