Image Aesthetic Quality Assesment (IAQA) Research Repository

contains:

training files for AVA benchmarking datase as binary classificaiton training is from directories in the following format:

   cwd\
       ⌊_train\
       |       ⌊_low\
       |       |     ⌊_id.jpg,...
       |       ⌊_high\
       |              ⌊_id.jpg,... 
       ⌊_test\...
       ⌊_val\...

Data dictionary of IAQA datasets with a file structure correspoinding to classes of each IAQA dataset

A terminal runnable version --including args --parse is to be completed (almost there)

Training Notebook:

binary:

ten bin thresholded classifier:

Evaluation Notebook:

notebooks currently have all code if first cell

DATA:

Dataset	n Images	Size (compressed)
AVA	255508	32 GB batched into 44 .zip
IAD	10k	2GB

Downloads are handled by code however you can obtain files in terminal using gdwon (best for large files):

to get ava:

gdown https://drive.google.com/drive/folders/1uc-jyzGNndFvhdiHaxAvEkj-jP5e6F1f?usp=sharing

to get IAD:

gdown https://drive.google.com/open?id=1Viswtzb77vqqaaICAQz9iuZ8OEYCu6-_

Frida

Training:

• write code for 10 cls
• Select alternative IAQA dataset 28/01
• train all models on second database:

• Resnets {18,50,152} 29/01
• CvT 29/01
• ConViT (T,S,B)
• CaiT
• BeiT

• Evaluate all models

code:

• update notbooks with git clone (this repository and remove code heavy cells)
• write guidance cells for colab nb. 27/01
• terminal runnable code with args parsers feb
• separate python files (resolving import of data augmentation class) feb
• create notebook with public trained models feb

EVAL

• compile results

Mawady Training:

• Train on AVA CaiT locally 28/01-30/01
• Train on AVA BeiT locally 28/01-30/01