Claude/iterative training plan tbjy5#2
Merged
Conversation
Implements the iterative training plan from the README as a runnable shell script with Pass 1 (full augmentations, shape learning, 200 epochs) and Pass 2 (minimal augmentations, color learning, 100 epochs). Supports --pass1-only and --pass2-only flags for independent execution. Includes k-NN evaluation after both passes for side-by-side comparison. https://claude.ai/code/session_01PxMmQiYRaGMb1ZXsULGVoN
main_dino.py loads ImageFolder directly from the root (no train/val split in the directory structure). eval_knn.py handles the 80/20 split internally via sklearn train_test_split. Removed the /train suffix from --data_path in both torchrun calls in run_dino_iterative.sh and corrected the matching README commands and dataset layout description. https://claude.ai/code/session_01PxMmQiYRaGMb1ZXsULGVoN
…rado/dinoLearning into claude/iterative-training-plan-tbjy5
Added option to load pre-trained weights. This is useful in the current approach doing the training in two stages
- run_dino_iterative.sh: increase batch size from 32 to 64 for both passes; double learning rates accordingly (P1: 0.0000625 -> 0.000125, P2: 0.00000625 -> 0.0000125) to maintain linear LR scaling - run_visualize_attention.sh: point checkpoint to epoch-40 snapshot (checkpoint0040.pth), switch input image to img.png, reduce image size from 1440x1440 to 960x960, set attention threshold to 0.3 - plot_training_metrics.ipynb: update notebook cell outputs/parameters
…tImages Outlines 8 steps to improve k-NN/linear accuracy beyond the current 88.96 (20-NN) baseline, including dataset cleaning, dropping the broken Pass 2 self-sup in favour of a linear probe, hyperparameter fixes (LR, out_dim, teacher_temp), patch size 8, augmentation calibration, self-distillation, and evaluation/infra upgrades.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.