Checklist NOAI China 2024 Round 2 (China Stage) · Task 2
Real or Fake Image Recognition Task
English title: 真假图像识别任务
Train a small CNN in PyTorch to tell real CIFAR-10 images from diffusion-generated ones, rewarding simpler networks.
The task
The real images are 5,000 training and 1,000 test images (3x32x32) taken from CIFAR-10; the fake images are 6,000 images generated by a diffusion model trained on CIFAR-10 (5,000 for training, 1,000 for testing). Training images are in train/cifar (real, label 0) and train/uvit (fake, label 1); the test set is hidden.
The model class must be named MyModel and contain at least 2 nn.Conv2d and 2 nn.MaxPool2d layers and at most 2 nn.Linear layers, built directly without nn.Sequential. Activation functions must be chosen from nn.ReLU, nn.Sigmoid, nn.Tanh, nn.ELU, nn.LeakyReLU and nn.PreLU [sic]; loss, optimiser and learning rate are free.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
- 10,000 training PNG images (5,000 real, 5,000 fake) from the Bohrium datasets tab; 2,000 hidden test images.
- You submit
submission.zipcontainingsubmission_model.pyandsubmission_dic.pth.- Scoring
- If the constraints are met: Network_Simplicity_Score = 1 / (Num_Linear + Num_Conv + 1); Score = (Network_Simplicity_Score + Accuracy) * 3/4; otherwise 0. Leaderboard B is final.
- Rules
- Class name MyModel.
- >= 2 Conv2d and >= 2 MaxPool2d layers; <= 2 Linear layers; no nn.Sequential.
- Activation functions only from the permitted list.
- Format
- NOAI 2024 Round 2 (China Stage), a one-day practical round in Beijing on 10 June 2024 (date and place from a third-party overview page). Republished on Bohrium as the 'NOAI2025 teaching test (NOAI2024 real problems)' and, in English, as the 'APOAI2025 Mock Competition' (10 Nov 2024 - 31 May 2025). Leaderboard A uses 50% of the test set during the contest; leaderboard B (the remaining 50%) is final. Question 2.