Pytorch Implementation of CLIC.

The official code of our paper:

CLIC: Contrastive Learning Framework for Unsupervised Image Complexity Representation

CLIC Framework and Results

t-SNE Visualization and Activation Map

Usage of CLIC

1. Requirements

Note that the version is not required, just in our experiment.

python==3.7.6
torch==1.12.0+cu116
torchaudio==0.12.0+cu116
torchvision==0.13.0+cu116

2. Data Preparation

Download ImageNet and Flickr.

Fine-tuning dataset is IC9600. You can see their github page.

Flickr parser scripts is ./dada/get_flickr.py. ImageNet parser please see ImageNet.

Images collection scripts is ./data/uniform_sample.py. Then you can get the clic dataset and folder architecture is below:

clic_dataset
    |—— images
        |- 001.jpg
        |- 002.jpg

3. Unsupervised Training

To do unsupervised pre-training run:

python train.py 
# you can modify args in this scripts.

4. Fine-tuning

To do fine-tuning on IC9600 run:

python fine_tuning.py
# you can modify args in this scripts.

Acknowledgment

• MoCo: Official PyTorch implementation of the MoCo.
• ICNet&IC9600:IC9600: A Benchmark Dataset for Automatic Image Complexity Assessment.
• ImageNet: ImageNet: An large-scale image dataset.
• Flickr-5B: Flickr: Approximately 5 billion images from Flickr.