search

lllllllllllll-llll / hyperIQA

Pytorch version of the CVPR 2020 paper: Blindly Assess Image Quality in the Wild Guided by A Self-Adaptive Hyper Network
13 stars 0 forks source link

Replace my own dataset #1

Open ke-s opened 3 years ago

ke-s commented 3 years ago

Hello, how should I do if I want to train the model with a dataset except in the paper? Images in the dataset and their corresponding scores are present in the TXT file and stored as A0248_00_00.bmp,1623.5506 A0248_00_01.bmp,1365.2132 A0248_00_02.bmp,1560.2607 A0248_00_03.bmp,1403.9202 A0248_00_04.bmp,1284.2490 A0248_00_05.bmp,1610.9308 A0248_00_06.bmp,1428.3295 A0248_00_07.bmp,1450.5216 A0248_00_08.bmp,1531.8135 A0248_00_09.bmp,1275.9054 A0248_00_10.bmp,1355.4574 Thanks

lllllllllllll-llll commented 3 years ago

Hello, I suggest you to see this code https://github.com/lllllllllllll-llll/hyperIQA-1. For your question, you can add some codes to read the TXT files, e.g., https://github.com/lllllllllllll-llll/hyperIQA-1/blob/master/folders.py.

lllllllllllll-llll commented 3 years ago

The TXT files and the image names are familiar to me, is that the Large-Scale Quality-annotated dataset proposed by "J. Wu, J. Ma, F. Liang, W. Dong, G. Shi and W. Lin, "End-to-End Blind Image Quality Prediction With Cascaded Deep Neural Network," in IEEE Transactions on Image Processing, vol. 29, pp. 7414-7426, 2020, doi: 10.1109/TIP.2020.3002478."

ke-s commented 3 years ago

Thank you very much. The dataset I want to test is "PIPAL: a Large-Scale Image Quality Assessment Dataset for Perceptual Image Restoration " with more range of images than the live dataset. But the type of distortion and storage seem to be different from the dataset like live, so I want to consult

---Original--- From: @.> Date: Wed, Oct 13, 2021 21:19 PM To: @.>; Cc: @.**@.>; Subject: Re: [lllllllllllll-llll/hyperIQA] Replace my own dataset (#1)

The TXT files and the image names are familiar to me, is that the Large-Scale Quality-annotated dataset proposed by "J. Wu, J. Ma, F. Liang, W. Dong, G. Shi and W. Lin, "End-to-End Blind Image Quality Prediction With Cascaded Deep Neural Network," in IEEE Transactions on Image Processing, vol. 29, pp. 7414-7426, 2020, doi: 10.1109/TIP.2020.3002478."

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

lllllllllllll-llll commented 3 years ago

I have test the code on PIPAL using my own IQA model, I can share the codes for you to load the dataset.

ke-s commented 3 years ago

Thank you very much for your help. Your generous help is really timely for a student with not strong code ability Thanks again for your help

---Original--- From: @.> Date: Thu, Oct 14, 2021 10:49 AM To: @.>; Cc: @.**@.>; Subject: Re: [lllllllllllll-llll/hyperIQA] Replace my own dataset (#1)

I have test the code on PIPAL using my own IQA model, I can share the codes for you to load the dataset.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

lllllllllllll-llll commented 3 years ago

see https://github.com/IIGROUP/RADN/blob/main/main.py

Mishra1995 commented 1 year ago

Hi @lllllllllllll-llll

Can you please suggest how to finetune HyperIQA based model for custom images I have. The two approaches I could think of is the freezing of backbone or updating the backbone with very small lr. Is this thinking correct? Kindly provide your insights?

lllllllllllll-llll commented 1 year ago

Hi @Mishra1995 You are right! If you want to finetune the baseline code, you can just freeze the base network parameters and set a small lr. I will upload the full version code these two days, including croos-database tests