analysis_report.md

# Sweep Analysis Report

Combined checkpoints: 129

- best lpips: scratch 27500 (step 27500, output `/apdcephfs_zwfy2/share_303944931/connorxian/Code/medUni/wt/zr/gen_joint/outputs_ckp_sweep/scratch/ckpt_0027500`)
- best ssim: scratch 22000 (step 22000, output `/apdcephfs_zwfy2/share_303944931/connorxian/Code/medUni/wt/zr/gen_joint/outputs_ckp_sweep/scratch/ckpt_0022000`)
- best psnr: merge 7500 (step 7500, output `/apdcephfs_zwfy2/share_303944931/connorxian/Code/medUni/wt/zr/gen_joint/outputs_ckp_sweep/merge/ckpt_0007500`)
- best fid: scratch 16000 (step 16000, output `/apdcephfs_zwfy2/share_303944931/connorxian/Code/medUni/wt/zr/gen_joint/outputs_ckp_sweep/scratch/ckpt_0016000`)
- best reference score: scratch 30500 (step 30500, output `/apdcephfs_zwfy2/share_303944931/connorxian/Code/medUni/wt/zr/gen_joint/outputs_ckp_sweep/scratch/ckpt_0030500`)

## Limitations
- sample_mask_metrics.csv files were found, but current test_joint_mask.py outputs only aggregate image metrics. Sample-level image-mask scatter plots were skipped because all_sample_metrics.csv/sample_image_metrics.csv were absent.
- Reference score is a review aid, not an automatic final selection rule.