Add high-resolution fine-tuning stage (448px tiles, magnification-aware)

[dat-rohit]

Summary

Implements Phase 2 high-resolution post-training from the Midnight paper (Section 2). Fine-tunes a Phase 1 checkpoint on 448px tiles at halved magnifications, preserving physical tissue regions while doubling pixel resolution. The ViT processes 392px global crops (28x28 = 784 tokens vs 16x16 = 256).

Key design decisions (validated against paper, codebase, and actual TCGA SVS data):

• 448px tiles (not 512): OpenMidnight Phase 1 uses 224px tiles, so 2x = 448. This preserves the physical-area invariant (224 * mpp = 448 * (mpp/2)). Midnight uses 256→512 (same 2x ratio).
• Magnification-aware sampling: Targets specific µm/px values [1.0, 0.5, 0.25, 0.125] by computing read_size at the best SVS level and resizing. SVS levels are 4x apart, so intermediate magnifications are synthesized.
• Loads backbone + heads: Loads DINO and iBOT heads from checkpoint (not just backbone), preserving Phase 1 training. Interpolates pos_embed from 16x16 to 28x28.
• LR scaling fix: Accounts for gradient accumulation in effective batch size calculation.

Changes

• prepatching_scripts/create_sample_dataset_txt_highres.py — Magnification-aware sample list generator with multiprocessing. Outputs path x y level read_size format.
• dinov2/data/datasets/slide_dataset.py — Parse optional read_size field, resize to patch_size_pixels
• dinov2/train/train.py — _load_from_teacher_checkpoint() with pos_embed interpolation, gradient accumulation loop, eval at cfg.crops.global_crops_size
• dinov2/train/ssl_meta_arch.py — loss_scale parameter for gradient accumulation
• dinov2/utils/config.py — LR scaling includes accumulation steps
• dinov2/configs/train/vitg14_reg4_highres.yaml — Phase 2 config
• run_highres_finetune.sh — Launch script

How to use

# 1. Generate sample list (magnification-aware, 448px tiles)
uv run python3 prepatching_scripts/create_sample_dataset_txt_highres.py \
  --target_patches 25000000 --workers 16

# 2. Update config with checkpoint path and sample list path

# 3. Launch
bash run_highres_finetune.sh

Test plan

• Sample list generation verified — read_sizes cluster correctly around expected values for each target µm/px
• Physical area invariant confirmed: [448, 224, 112, 56] µm matches Phase 1's [448, 224, 112, 56] µm
• pos_embed interpolation verified: [1, 257, 1536] → [1, 785, 1536]
• Dry-run training on 1 GPU (in progress)
• Full 120k iteration run
• Eva benchmark at 392px

🤖 Generated with Claude Code