search

narthchin / DEIQT

Checkpoints, logs and source code for AAAI-23 paper 'Data-Efficient Image Quality Assessment with Attention-Panel Decoder'
32 stars 3 forks source link

Regarding the strange performance of the model, as well as consultation on parameter configurations for the TID2013 dataset #9

Open easeAutumn opened 5 months ago

easeAutumn commented 5 months ago

Thank you for this amazing work.However, I've run into some challenges.I downloaded the code, but did not receive any configuration files for the TID2013 dataset. I conducted experiments on the TID2013 dataset by using the configuration parameters from the LIVEC dataset, and initially, the model indeed produced impressive results. However, during subsequent training, the model's performance was difficult to improve, and even exhibited a downward trend. Additionally, it seems that the model always achieves the best performance after roughly the same number of steps. I am so confused with these phenomena. Here are the experimental results. 1715342168804

narthchin commented 5 months ago

Sorry that we forgot to upload config for TID XD. It will be uploaded these days.

narthchin commented 5 months ago

Try this for your training:

DATA:
  DATASET: tid2013
  BATCH_SIZE: 48
  PATCH_NUM: 10
  CROP_SIZE: (224,224)
SET:
  COUNT: 25
easeAutumn commented 5 months ago

Thank you for your reply.I will try this configuration.