KDD Project 2018

This is the README.md file of github repository github.com/kvas7andy/kdd_project_2018

General info

This project replicates the paper "Cost-Effective Training of Deep CNNs with Active Model Adaptation" [KDD2018 paper page]

Structure

File/Folder	Description
├── datasets	Contains files for ImageNet, DogVsCat datasets (source files as well as images and .txt contents)
/├── c_x_A_B	Stores calculated features from layers VGG, AlexNet models (from pretrained models)
//├── alex
//└── vgg16	For VGG16: the same
/├── dogvscat	DogVSCat folder to store folder train and train.txt, eval.txt files references images from that folder
/└── imagenet	Files related to structured ImageNet dataset from the
├── libs	Src files for ImageNet dataset download
├── src	All source files for the project, except ImageNet downloaders

Prerequirements:

• python3.6
• pytorch
• numpy
• scipy
• pandas
• pillow
• urllib

Tested on Ubuntu 16.04 with 12 Gb GPU

Compilation

No compilation necessary but preliminary datasets download is necessary, go to folders of ImageNet and DogVSCat dataset

Execution

0. Download data: specific to datasets. ImageNet download with downloadutils.py and target datasets, for instance, from files. See README.md respectively for ImageNet and DogVSCat.

1. Firstly change parameters in utils.py:

• Sampling related
  • active_samepl_size (mini batch to check the score composed of distinctiveness and uncertainty)
  • active_batch_size (for the exact labeled batch of images to ask for)
  • eval_batch_size (after how many images to make evaluation)
• Model related
  • model_type - Deep CNNs to use (Alex or VGG16 available)
  • active - whether apply active batching
  • lamb - speed parameter for disctictiveness/uncertainty trade-off
  • transform - whether to use data augmentation
  • alternate - whether to prevent using images from one class consequently for transfer learning with 1 image, only for miaty class (project team improvement proposal)
• For results report:
  • eval - whether perform evaluation on datasets, and save accuracy of the model into the src/historyactive_... file after 100 batches
  • eval_every - evaluation frequency (number of acative batches)
  • report_every - report evaluation performance by printing

2. Then execute src/main.py: python3.6 ./src/main.py
3. Output will be the accuracy change while fine-tuning on evalutaion dataset stored in .npy format in files historyactive_[active][lambda value]

For any questions refer to the authors:

• KVASOV Andrei: akvasov@connect.ust.hk
• LI Shichao: nicholas.li@connect.ust.hk
• LI Ziyue: zlibn@connect.ust.hk