Competition Rules Data and Evaluation Leaderboard FAQs
Data Description Using SOPHGO TPU Submission Requirements Evaluation

Data Description

  • Test set A

Obtained low resolution images with 4 reduction factors from 2k/8k high-definition high-resolution images collected from the network, and was open to contestants in the A-list stage. Players can use this dataset for testing within a limited time to generate A-list scores and rankings.

 

  • Test set B

Obtained low resolution images with 4 reduction factors from 2k/8k high-definition high-resolution images collected from the network. In the stage of calculating A-list scores, release test set B. Players can use this dataset for testing within a limited time to generate B-list scores and rankings.

Submission Requirements

  • Preliminary stage

 

JSON result submission:

Participants will submit their results in a single JSON file to the big data competition platform, which will conduct online scoring and real-time ranking to determine the team that is qualified for the final round based on the deadline ranking;

Participants are required to submit a result file named "test. json" that infers the training set on the TPU platform. The inference results need to include model size, average inference time (in seconds) for generating each image, average NIQE value for each image, name of each image, inference time for generating each image, and corresponding NIQE value for each image. Details:

[test. json file]

算能1.png

 

  • Recurrence stage

    The top 5 teams in the preliminary B-list have entered the reproduction stage and need to submit reproduction materials as required. After the reproduction is completed, the list of teams shortlisted for the finals will be announced.

    Participants are required to submit a BModel model and corresponding test code that can run on the BM1684x platform.

  • Details:
  1. The test code should be saved in test.py format
  2. Original model weights and bmodel models (bmodel models should be saved in out. bmodel format)
  3. The test code and model should be placed in the same folder, compressed into a. zip file, and then submitted

Evaluation

  1. Evaluate the model accuracy by averaging the Natural Image Quality Evaluator (NIQE) metrics for each image;
  2. Reasoning time i through the model_ Time evaluation model performance, i_ Time should be the average time for image inference in the dataset, in seconds;
  3. The final scoring formula is: score=sqrt (7-niqe_score)/i_ Time * 200