Datasets:
Add files using upload-large-folder tool
Browse files- STAGING_FILELIST_DEPTH2.tsv +13 -0
- model_cache_code_step8000/CODE_USAGE.md +55 -0
- model_cache_code_step8000/DATA_ASSETS.md +64 -0
- model_cache_code_step8000/HF_UPLOAD.md +52 -0
- model_cache_code_step8000/PACKAGE_COMPLETE +0 -0
- model_cache_code_step8000/README.md +71 -0
- model_cache_code_step8000/SHA256SUMS.txt +44 -0
- model_cache_code_step8000/UPLOAD_NOTES.md +8 -0
- raw_pants_train_test/PACKAGE_COMPLETE +0 -0
- raw_pants_train_test/README.md +25 -0
- raw_pants_train_test/SHA256SUMS.txt +0 -0
- raw_pants_train_test/metadata/package_du.txt +1 -0
- raw_pants_train_test/metadata/package_filelist.tsv +563 -0
- raw_pants_train_test/metadata/package_tree_depth3.txt +29 -0
- raw_pants_train_test/metadata/pants-captions-ldm/LICENSE +51 -0
- raw_pants_train_test/metadata/pants-captions-ldm/MANIFEST.sha256 +521 -0
- raw_pants_train_test/metadata/pants-captions-ldm/README.md +207 -0
- raw_pants_train_test/metadata/pants-captions-ldm/UPLOAD_INSTRUCTIONS.md +81 -0
- raw_pants_train_test/metadata/pants-captions-ldm/audit/audit_final.md +23 -0
- raw_pants_train_test/metadata/pants-captions-ldm/audit/audit_v2.md +187 -0
- raw_pants_train_test/metadata/pants-captions-ldm/audit/phase_audit.jsonl +0 -0
- raw_pants_train_test/metadata/pants-captions-ldm/audit/phase_audit_summary.md +42 -0
- raw_pants_train_test/metadata/pants-captions-ldm/captions/captions_pilot.jsonl +0 -0
- raw_pants_train_test/metadata/pants-captions-ldm/captions/captions_v1_initial.jsonl +0 -0
- raw_pants_train_test/metadata/pants-captions-ldm/captions/captions_v7_pancreas_only.jsonl +0 -0
- raw_pants_train_test/metadata/pants-captions-ldm/code/aug_config.py +47 -0
- raw_pants_train_test/metadata/pants-captions-ldm/code/build_bucket_spec.py +75 -0
- raw_pants_train_test/metadata/pants-captions-ldm/code/build_splits.py +36 -0
- raw_pants_train_test/metadata/pants-captions-ldm/code/dry_run_train.py +240 -0
- raw_pants_train_test/metadata/pants-captions-ldm/code/pack_webdataset.py +229 -0
- raw_pants_train_test/metadata/pants-captions-ldm/code/phase_audit.py +235 -0
- raw_pants_train_test/metadata/pants-captions-ldm/code/preprocessing.py +268 -0
- raw_pants_train_test/metadata/pants-captions-ldm/code/schedule.py +145 -0
- raw_pants_train_test/metadata/pants-captions-ldm/code/test_preprocessing.py +107 -0
- raw_pants_train_test/metadata/pants-captions-ldm/code/vae_sanity_gate.py +298 -0
- raw_pants_train_test/metadata/pants-captions-ldm/docs/GPT_REVIEW.md +30 -0
- raw_pants_train_test/metadata/pants-captions-ldm/docs/LDM_TRAINING_PLAN_v4.md +416 -0
- raw_pants_train_test/metadata/pants-captions-ldm/docs/PRETRAIN_CHECKLIST_SUMMARY.md +115 -0
- raw_pants_train_test/metadata/pants-captions-ldm/docs/PROPOSAL_v1.md +191 -0
- raw_pants_train_test/metadata/pants-captions-ldm/docs/PROPOSAL_v3_1.md +158 -0
- raw_pants_train_test/metadata/pants-captions-ldm/docs/PROPOSAL_v3_2.md +187 -0
- raw_pants_train_test/metadata/pants-captions-ldm/docs/fusion_variants_design.md +99 -0
- raw_pants_train_test/metadata/pants-captions-ldm/samples/canary/annotations.json +0 -0
- raw_pants_train_test/metadata/pants-captions-ldm/samples/canary/ids.json +202 -0
- raw_pants_train_test/metadata/pants-captions-ldm/samples/pilot/annotations.json +0 -0
- raw_pants_train_test/metadata/pants-captions-ldm/samples/pilot_v2/annotations.json +0 -0
- raw_pants_train_test/metadata/pants-captions-ldm/splits/bucket_spec.json +175 -0
- raw_pants_train_test/metadata/pants-captions-ldm/splits/splits.json +0 -0
- raw_pants_train_test/metadata/raw_summary.txt +27 -0
- raw_pants_train_test/raw/RESTORE.txt +2 -0
STAGING_FILELIST_DEPTH2.tsv
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
132226 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/raw_pants_train_test/SHA256SUMS.txt
|
| 2 |
+
8971 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/model_cache_code_step8000/SHA256SUMS.txt
|
| 3 |
+
3288 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/README.md
|
| 4 |
+
2674 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/model_cache_code_step8000/README.md
|
| 5 |
+
2169 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/model_cache_code_step8000/CODE_USAGE.md
|
| 6 |
+
1806 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/model_cache_code_step8000/DATA_ASSETS.md
|
| 7 |
+
1538 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/model_cache_code_step8000/HF_UPLOAD.md
|
| 8 |
+
1113 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/raw_pants_train_test/README.md
|
| 9 |
+
359 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/UPLOAD_STATUS.md
|
| 10 |
+
317 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/model_cache_code_step8000/UPLOAD_NOTES.md
|
| 11 |
+
0 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/STAGING_FILELIST_DEPTH2.tsv
|
| 12 |
+
0 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/raw_pants_train_test/PACKAGE_COMPLETE
|
| 13 |
+
0 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/MA048181_public_upload.tmp/model_cache_code_step8000/PACKAGE_COMPLETE
|
model_cache_code_step8000/CODE_USAGE.md
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Code and Checkpoint Usage
|
| 2 |
+
|
| 3 |
+
## Code Snapshot
|
| 4 |
+
|
| 5 |
+
The code archive contains two trees:
|
| 6 |
+
|
| 7 |
+
- `TwoFrame`: experiment orchestration, evaluation, data-engine scripts, and local project code.
|
| 8 |
+
- `FastVideo`: the training stack used for this Wan2.2 fine-tune.
|
| 9 |
+
|
| 10 |
+
FastVideo matters for this run. The training log references FastVideo modules such as `fastvideo_args.py`, `training_pipeline.py`, and `training_utils.py`. The relevant dirty FastVideo changes are recorded in:
|
| 11 |
+
|
| 12 |
+
- `metadata/FastVideo.git_status.txt`
|
| 13 |
+
- `metadata/FastVideo.uncommitted.diff`
|
| 14 |
+
- `metadata/FastVideo.untracked_files.txt`
|
| 15 |
+
|
| 16 |
+
TwoFrame status is recorded similarly in `metadata/TwoFrame.*`.
|
| 17 |
+
|
| 18 |
+
## Environment
|
| 19 |
+
|
| 20 |
+
The UCSF environment used during training was:
|
| 21 |
+
|
| 22 |
+
```bash
|
| 23 |
+
source ~/.twoframe_env.sh
|
| 24 |
+
conda activate /scratch/user/yuhwang/envs/twoframe
|
| 25 |
+
export PYTHONPATH=/scratch/user/yuhwang/code/FastVideo:/scratch/user/yuhwang/code/TwoFrame:$PYTHONPATH
|
| 26 |
+
```
|
| 27 |
+
|
| 28 |
+
`~/.twoframe_env.sh` points caches and temporary directories to scratch.
|
| 29 |
+
|
| 30 |
+
## Checkpoints
|
| 31 |
+
|
| 32 |
+
Use the EMA checkpoint for inference-style evaluation unless you explicitly want non-EMA weights:
|
| 33 |
+
|
| 34 |
+
- EMA: `checkpoints/ema_checkpoint-8000/diffusion_pytorch_model.safetensors`
|
| 35 |
+
- non-EMA: `checkpoints/checkpoint-8000/transformer/diffusion_pytorch_model.safetensors`
|
| 36 |
+
|
| 37 |
+
Use the distributed checkpoint only if resuming training:
|
| 38 |
+
|
| 39 |
+
- `checkpoints/checkpoint-8000/distributed_checkpoint/`
|
| 40 |
+
|
| 41 |
+
## Base Model
|
| 42 |
+
|
| 43 |
+
The run loaded base model components from:
|
| 44 |
+
|
| 45 |
+
`/scratch/user/yuhwang/model/Wan2.2-TI2V-5B-Diffusers-merged`
|
| 46 |
+
|
| 47 |
+
That directory is symlink-composed. Its symlink map is recorded in `metadata/base_model_symlinks.txt`. The base model binaries are not duplicated in this package; this package contains the fine-tuned transformer checkpoint and reproducibility assets.
|
| 48 |
+
|
| 49 |
+
## Training Command Reconstruction
|
| 50 |
+
|
| 51 |
+
The exact shell launcher was not preserved in the final run directory. The effective training configuration is captured from the log in `metadata/train_args.json`. Use that file as the source of truth for reconstructing a resume or reproduction launch.
|
| 52 |
+
|
| 53 |
+
The original output directory was:
|
| 54 |
+
|
| 55 |
+
`/scratch/user/yuhwang/artifacts/twoframe/pants_wan22_finetune/pants_b16_9k_20260519_215532`
|
model_cache_code_step8000/DATA_ASSETS.md
ADDED
|
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Data Assets
|
| 2 |
+
|
| 3 |
+
There are three distinct data layers. Do not collapse them when uploading or restoring.
|
| 4 |
+
|
| 5 |
+
## 1. Original Raw Data
|
| 6 |
+
|
| 7 |
+
Original source data is PanTS:
|
| 8 |
+
|
| 9 |
+
`/scratch/user/yuhwang/dataset/PanTS`
|
| 10 |
+
|
| 11 |
+
Approximate size observed before packaging: 371G.
|
| 12 |
+
|
| 13 |
+
Structure:
|
| 14 |
+
|
| 15 |
+
- `data/ImageTr`: training images/volumes
|
| 16 |
+
- `data/ImageTe`: test images/volumes
|
| 17 |
+
- `data/LabelTr`: training labels
|
| 18 |
+
- `data/LabelTe`: test labels
|
| 19 |
+
- `data/metadata.xlsx`: raw metadata table
|
| 20 |
+
- `repo`: PanTS source README/license/download scripts
|
| 21 |
+
|
| 22 |
+
Raw data is packaged separately under `pants_raw_train_test_20260523_102411`.
|
| 23 |
+
|
| 24 |
+
## 2. VAE Latent Cache
|
| 25 |
+
|
| 26 |
+
Derived VAE cache used by the training dataloader:
|
| 27 |
+
|
| 28 |
+
`/scratch/user/yuhwang/dataset/pants-captions-ldm/cache/wan22_pants_v2_softwin/latents`
|
| 29 |
+
|
| 30 |
+
This stores pre-encoded Wan2.2 VAE latents as `.safetensors`. The train manifest references these via `latent_path`.
|
| 31 |
+
|
| 32 |
+
## 3. Text Cache
|
| 33 |
+
|
| 34 |
+
Derived text embedding cache:
|
| 35 |
+
|
| 36 |
+
`/scratch/user/yuhwang/dataset/pants-captions-ldm/cache/wan22_pants_v2_softwin/text_embeddings`
|
| 37 |
+
|
| 38 |
+
The main training manifest is:
|
| 39 |
+
|
| 40 |
+
`captions_embedded/source_train.jsonl`
|
| 41 |
+
|
| 42 |
+
Observed row count: 53,784.
|
| 43 |
+
|
| 44 |
+
Observed bucket counts:
|
| 45 |
+
|
| 46 |
+
- `B-whole`: 21,690
|
| 47 |
+
- `B-CA`: 11,365
|
| 48 |
+
- `B-CAP`: 10,980
|
| 49 |
+
- `P-pan`: 8,784
|
| 50 |
+
- `B-abd-pelvis`: 490
|
| 51 |
+
- `B-abd`: 475
|
| 52 |
+
|
| 53 |
+
The manifest fields include `bucket_id`, `caption`, `latent_path`, `latent_shape`, `text_embedding_path`, `text_embedding_shape`, and CT metadata fields.
|
| 54 |
+
|
| 55 |
+
## Cache Archive Included Here
|
| 56 |
+
|
| 57 |
+
`cache/wan22_pants_v2_softwin.tar.zst.part-*` is a split tar+zstd archive of the full derived cache directory. It includes:
|
| 58 |
+
|
| 59 |
+
- `latents/`: VAE latent cache
|
| 60 |
+
- `text_embeddings/`: text embedding cache
|
| 61 |
+
- `captions/`: caption jsonl files
|
| 62 |
+
- `captions_embedded/`: embedded-caption manifests used by training
|
| 63 |
+
- `manifests/`: pre-embedding manifest layer
|
| 64 |
+
- `cache_config.json`
|
model_cache_code_step8000/HF_UPLOAD.md
ADDED
|
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Hugging Face Upload Plan
|
| 2 |
+
|
| 3 |
+
Use the modern `hf` CLI, not deprecated `huggingface-cli`.
|
| 4 |
+
|
| 5 |
+
## Recommended Repos
|
| 6 |
+
|
| 7 |
+
Use two repos:
|
| 8 |
+
|
| 9 |
+
1. Model/reproducibility repo for this package:
|
| 10 |
+
- checkpoints
|
| 11 |
+
- code snapshot
|
| 12 |
+
- derived cache archive
|
| 13 |
+
- logs and metadata
|
| 14 |
+
|
| 15 |
+
2. Dataset repo for raw PanTS train/test package:
|
| 16 |
+
- raw split archive
|
| 17 |
+
- PanTS repo README/license
|
| 18 |
+
- split/caption/canonical metadata
|
| 19 |
+
|
| 20 |
+
## Upload Model Package
|
| 21 |
+
|
| 22 |
+
```bash
|
| 23 |
+
cd /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252
|
| 24 |
+
hf auth whoami
|
| 25 |
+
hf upload-large-folder Neuripsused/MA048181 . --type model --num-workers 8
|
| 26 |
+
```
|
| 27 |
+
|
| 28 |
+
For a private upload:
|
| 29 |
+
|
| 30 |
+
```bash
|
| 31 |
+
hf repos create Neuripsused/MA048181 --type model --private --exist-ok
|
| 32 |
+
hf upload-large-folder Neuripsused/MA048181 . --type model --num-workers 8
|
| 33 |
+
```
|
| 34 |
+
|
| 35 |
+
## Upload Raw Data Package
|
| 36 |
+
|
| 37 |
+
```bash
|
| 38 |
+
cd /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411
|
| 39 |
+
hf repos create Neuripsused/MA048181 --type dataset --private --exist-ok
|
| 40 |
+
hf upload-large-folder Neuripsused/MA048181 . --type dataset --num-workers 8
|
| 41 |
+
```
|
| 42 |
+
|
| 43 |
+
## Verify Before Upload
|
| 44 |
+
|
| 45 |
+
```bash
|
| 46 |
+
sha256sum -c SHA256SUMS.txt
|
| 47 |
+
find . -maxdepth 3 -type f | sort | sed -n '1,200p'
|
| 48 |
+
```
|
| 49 |
+
|
| 50 |
+
## Suggested Repo README Language
|
| 51 |
+
|
| 52 |
+
This artifact contains derived training cache and checkpoints. The raw PanTS source data is distributed separately because it is a much larger dataset-layer artifact. The last complete checkpoint is step 8000, with both EMA and non-EMA transformer weights available.
|
model_cache_code_step8000/PACKAGE_COMPLETE
ADDED
|
File without changes
|
model_cache_code_step8000/README.md
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# PanTS Wan2.2 Fine-tune: Cache, Code, and Step-8000 Checkpoint
|
| 2 |
+
|
| 3 |
+
This package is the reproducibility bundle for the UCSF run `pants_b16_9k_20260519_215532`.
|
| 4 |
+
|
| 5 |
+
The run targeted 9000 training steps, but the 2-day Slurm allocation ended after step 8016. There was no traceback, OOM, or NaN in the training log. The last complete saved checkpoint pair is step 8000.
|
| 6 |
+
|
| 7 |
+
## What Is Included
|
| 8 |
+
|
| 9 |
+
- Final non-EMA checkpoint: `checkpoints/checkpoint-8000/transformer/`
|
| 10 |
+
- Distributed resume checkpoint: `checkpoints/checkpoint-8000/distributed_checkpoint/`
|
| 11 |
+
- Final EMA checkpoint: `checkpoints/ema_checkpoint-8000/`
|
| 12 |
+
- Training cache archive: `cache/wan22_pants_v2_softwin.tar.zst.part-*`
|
| 13 |
+
- Two code snapshots: `code/twoframe_fastvideo_code_snapshot.tar.zst`
|
| 14 |
+
- Training logs and offline tracker: `logs/`
|
| 15 |
+
- Metadata and provenance: `metadata/`
|
| 16 |
+
|
| 17 |
+
The original raw PanTS train/test data is intentionally packaged separately as:
|
| 18 |
+
|
| 19 |
+
`/scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411`
|
| 20 |
+
|
| 21 |
+
That raw package is better uploaded as a Hugging Face dataset repo. This package is better uploaded as a Hugging Face model repo.
|
| 22 |
+
|
| 23 |
+
## Key Paths From Training
|
| 24 |
+
|
| 25 |
+
- Raw data root: `/scratch/user/yuhwang/dataset/PanTS`
|
| 26 |
+
- Derived cache root: `/scratch/user/yuhwang/dataset/pants-captions-ldm/cache/wan22_pants_v2_softwin`
|
| 27 |
+
- Base model path: `/scratch/user/yuhwang/model/Wan2.2-TI2V-5B-Diffusers-merged`
|
| 28 |
+
- TwoFrame code: `/scratch/user/yuhwang/code/TwoFrame`
|
| 29 |
+
- FastVideo code: `/scratch/user/yuhwang/code/FastVideo`
|
| 30 |
+
- Run output: `/scratch/user/yuhwang/artifacts/twoframe/pants_wan22_finetune/pants_b16_9k_20260519_215532`
|
| 31 |
+
|
| 32 |
+
## Training Summary
|
| 33 |
+
|
| 34 |
+
See `metadata/train_args.json` and `metadata/train_health.json` for the full captured FastVideo arguments and log health summary.
|
| 35 |
+
|
| 36 |
+
Important args:
|
| 37 |
+
|
| 38 |
+
- Model: Wan2.2 TI2V 5B Diffusers merged local directory
|
| 39 |
+
- Data cache: `wan22_pants_v2_softwin`
|
| 40 |
+
- GPUs: 8
|
| 41 |
+
- Batch size: 16
|
| 42 |
+
- Precision: bf16
|
| 43 |
+
- Optimizer: AdamW
|
| 44 |
+
- Learning rate: `1e-6`
|
| 45 |
+
- EMA: enabled, decay `0.999`, start step `1`
|
| 46 |
+
- Checkpoint interval: 4000 steps
|
| 47 |
+
- Last complete checkpoint: 8000
|
| 48 |
+
|
| 49 |
+
## Restore Cache
|
| 50 |
+
|
| 51 |
+
```bash
|
| 52 |
+
cd cache
|
| 53 |
+
cat wan22_pants_v2_softwin.tar.zst.part-* > wan22_pants_v2_softwin.tar.zst
|
| 54 |
+
tar --use-compress-program zstd -xf wan22_pants_v2_softwin.tar.zst
|
| 55 |
+
```
|
| 56 |
+
|
| 57 |
+
## Restore Code Snapshot
|
| 58 |
+
|
| 59 |
+
```bash
|
| 60 |
+
mkdir -p /scratch/user/yuhwang/code_restore
|
| 61 |
+
cd /scratch/user/yuhwang/code_restore
|
| 62 |
+
tar --use-compress-program zstd -xf /path/to/code/twoframe_fastvideo_code_snapshot.tar.zst
|
| 63 |
+
```
|
| 64 |
+
|
| 65 |
+
Then use the FastVideo and TwoFrame trees from that restored directory, or compare against the git status/diff metadata in `metadata/`.
|
| 66 |
+
|
| 67 |
+
## Check Integrity
|
| 68 |
+
|
| 69 |
+
```bash
|
| 70 |
+
sha256sum -c SHA256SUMS.txt
|
| 71 |
+
```
|
model_cache_code_step8000/SHA256SUMS.txt
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
2a71d38c621bb6603383414dcf628e129b7c1b2efef8ee2c8407e6c4df0d4645 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/CODE_USAGE.md
|
| 2 |
+
1b4f60f37f4a0baae8f650d1fcdc81dbbb652f7ea1e6ce8aabdc50ac4ab6cb24 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/DATA_ASSETS.md
|
| 3 |
+
725ab7ce9a51972f582d8b4056d17241a18d9f07c8b4c1af031aed85eb36188e /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/HF_UPLOAD.md
|
| 4 |
+
e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/PACKAGE_COMPLETE
|
| 5 |
+
84d3c9eaa746d166776db0b57528649e9d8798732ad62eb1914bcefad278bb1a /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/README.md
|
| 6 |
+
0a3c2c1e6a93a45f0444f4a20353083e2d6ca491db840f262b47c8ca44130460 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/UPLOAD_NOTES.md
|
| 7 |
+
6edc686e051c1ca375485420d63544306ac652a3c419067936753e63e90614c3 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/cache/RESTORE.txt
|
| 8 |
+
235749b3738cce42c4bffb3f5de0b8155e1a53f34eeee759977182bf20e6c3d7 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/cache/wan22_pants_v2_softwin.tar.zst.part-000
|
| 9 |
+
2886ca1b871944aa02f7e4517205b2cc649fe1d8b46096d851c8872379f9c08e /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/cache/wan22_pants_v2_softwin.tar.zst.part-001
|
| 10 |
+
838eee3cad3f4742a21da7bc6921d04d06cc5e84513fc9147ae888aa2ca627f0 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/cache/wan22_pants_v2_softwin.tar.zst.part-002
|
| 11 |
+
1d41a243d103ac7e7addd612868abf62a597ff18b0e435f58b0566b713ec48a2 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/cache/wan22_pants_v2_softwin.tar.zst.part-003
|
| 12 |
+
402eca894dcad592685c53a3e65e8bece84ec7053fdbf520761c4b3798d3cf48 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/checkpoint-8000/distributed_checkpoint/.metadata
|
| 13 |
+
cff247d9abb76826c56e4e90582e48fcd49b7dba32f422ccf5b6c1246de3bca7 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/checkpoint-8000/distributed_checkpoint/__0_0.distcp
|
| 14 |
+
b79eb539fbef3888a3659ac03bc84c5e4033a57dbc5be7cd27f983e10aaa337d /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/checkpoint-8000/distributed_checkpoint/__1_0.distcp
|
| 15 |
+
dd4e09bf820f8e27f3f7e9f790e21ad5f52f885ffc22169209005385f17a48ee /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/checkpoint-8000/distributed_checkpoint/__2_0.distcp
|
| 16 |
+
b5e1ebf97a5de44f5150be156183914d524aa21613b80badadb87c587af61252 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/checkpoint-8000/distributed_checkpoint/__3_0.distcp
|
| 17 |
+
afdf7d5a69d0f8ba7da3bdd795a7175348c44d09bd1a6e75fc2cc75e6a1b7f35 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/checkpoint-8000/distributed_checkpoint/__4_0.distcp
|
| 18 |
+
179d128735c32d559d6a6451ced63f1cf21b4d0d23de7efc4274697c82ed1f85 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/checkpoint-8000/distributed_checkpoint/__5_0.distcp
|
| 19 |
+
aaa119f4cb6ac6127dbc32f4dc9901cc6cb247785afeeb22e0a273a222297420 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/checkpoint-8000/distributed_checkpoint/__6_0.distcp
|
| 20 |
+
3c7273217b183b306f9f2b63756390a9ca144f60e1517d90cfc2c46b368585bd /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/checkpoint-8000/distributed_checkpoint/__7_0.distcp
|
| 21 |
+
9ffd02a1f215985805f12ee409dd3deb5fbc2a812ffa555e6027f71d9764ee5d /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/checkpoint-8000/transformer/config.json
|
| 22 |
+
4e1e3050d44db6234c877e88cc1587390a7ded0da5b437eaa83776b7f30add70 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/checkpoint-8000/transformer/diffusion_pytorch_model.safetensors
|
| 23 |
+
407a2c305e053419aea3f2f1fc5652dda44983d01f50a0256f737918dfac2124 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/checkpoints/ema_checkpoint-8000/diffusion_pytorch_model.safetensors
|
| 24 |
+
409149500a6f5370d3246316623c1b6a87f0ba7b70b8457acb9aed30c17a2d86 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/code/twoframe_fastvideo_code_snapshot.tar.zst
|
| 25 |
+
490cf46859bd5fd2174157b4076c8f04b1b75dfd7860b6adc16e5d91e29a33b9 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/logs/gpu_monitor.log
|
| 26 |
+
2796f6c6d197a28bffb823fc380b0721b47029040c0b872a434ce2c9ba217aae /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/logs/tracker/wandb/offline-run-20260519_215609-2dcsowl9/files/requirements.txt
|
| 27 |
+
373cd01cfeddb1659c5f1c071e04ffa9c6e435ef7f25e7e2bcf494c14f8ce14a /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/logs/tracker/wandb/offline-run-20260519_215609-2dcsowl9/logs/debug-internal.log
|
| 28 |
+
26970c3cd0312250f2a26bdb686e2462c3a48a9f3bbed08eabd3bbe3c8b769e2 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/logs/tracker/wandb/offline-run-20260519_215609-2dcsowl9/logs/debug.log
|
| 29 |
+
7453d01419e2e0c218b1419e274b8223f5a4b827824b3035de35bad3061ed974 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/logs/tracker/wandb/offline-run-20260519_215609-2dcsowl9/run-2dcsowl9.wandb
|
| 30 |
+
961a598feed253cd5d1982591025afa00ad3e35ebb6b9e12e91e85a7cfa6c1e0 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/logs/train.log
|
| 31 |
+
11817d6088dc13953a87132012c6fcef129465e256ac2a0a16101f7cea95edae /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/FastVideo.git_status.txt
|
| 32 |
+
1bd1194db330ba209098eae779e6eb03dbc8361df421ba83d8c7c1c318824e43 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/FastVideo.uncommitted.diff
|
| 33 |
+
a00ca5bec87e15b549eb416d6763a6d8a8812b88fe6220a27f1d209adb848d02 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/FastVideo.untracked_files.txt
|
| 34 |
+
7f72ebc5aee3d0331f9ea8e10283011cee73ab0ee91f71694e6fd59274a90d72 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/TwoFrame.git_status.txt
|
| 35 |
+
a714cabb398bdb2ba4833b4329de48adaf4d49f8fc7fb567e2d0877036be441c /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/TwoFrame.uncommitted.diff
|
| 36 |
+
8a5b8315d0f01fb7621e66a7639c2e63eaf202a86028b90fdfeca5c56f6874a0 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/TwoFrame.untracked_files.txt
|
| 37 |
+
e81cdde97ecfc9d1fa0bb55c63377fac36cf60b8fba1996d42daaf4be162de4a /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/asset_manifest.json
|
| 38 |
+
a687c69711100d5a40bb0a5082d852d3e9fe8a0f2f749f0aae9cc9663f8f4bde /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/base_model_symlinks.txt
|
| 39 |
+
e6df5ad453a628c2f4202986b82300a3f9cead13d7f54d10454a91c619026409 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/cache_summary.json
|
| 40 |
+
2951426e31ccc8c46b7a0627121344d951b45859b2c80e391ba0cc1fbd252af1 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/package_du.txt
|
| 41 |
+
2fdb67082026a9e2d404669c46fcb02b6fb57e446a0b24debd8df6942ab2757f /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/package_filelist.tsv
|
| 42 |
+
d97c863dd1d7091980febb4e0d0fb3a2f012096bef78e7305f6153a735c0a236 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/package_tree_depth3.txt
|
| 43 |
+
171f7b01a99eb8a642d62a07bf80aef0d86f7903f4deaa9f22f9050d3cbedea1 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/train_args.json
|
| 44 |
+
67474b319c57a3b51c1d275ce43cd2819772fba30699f5076130f811b8010ee9 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252/metadata/train_health.json
|
model_cache_code_step8000/UPLOAD_NOTES.md
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Suggested Hugging Face upload pattern after inspecting the package:
|
| 2 |
+
|
| 3 |
+
`ash
|
| 4 |
+
cd /scratch/user/yuhwang/artifacts/twoframe/hf_upload/<package-dir>
|
| 5 |
+
hf upload <repo-id> . --repo-type model
|
| 6 |
+
`
|
| 7 |
+
|
| 8 |
+
For a dataset-style upload of the training cache, upload the cache/*.part-* files and cache/RESTORE.txt as a dataset repo instead.
|
raw_pants_train_test/PACKAGE_COMPLETE
ADDED
|
File without changes
|
raw_pants_train_test/README.md
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# PanTS Raw Data Package for Wan2.2 Fine-tuning
|
| 2 |
+
|
| 3 |
+
This package contains the raw train/test data used upstream of the Wan2.2 PanTS latent/text cache.
|
| 4 |
+
|
| 5 |
+
Contents:
|
| 6 |
+
|
| 7 |
+
-
|
| 8 |
+
- Includes data/ImageTr, data/ImageTe, data/LabelTr, data/LabelTe, data/metadata.xlsx, and
|
| 9 |
+
- metadata/pants-captions-ldm/: small metadata layer used before cache construction: splits, canonical index, captions, docs, samples, audit files, and source README/license.
|
| 10 |
+
- metadata/raw_summary.txt: size/count summary at packaging time.
|
| 11 |
+
- SHA256SUMS.txt: checksums for all package files.
|
| 12 |
+
|
| 13 |
+
Relationship to cache package:
|
| 14 |
+
|
| 15 |
+
- This is the original raw data layer.
|
| 16 |
+
- The training package pants_wan22_b16_9k_step8000_* contains the derived data layer: VAE latents and text embeddings under wan22_pants_v2_softwin.
|
| 17 |
+
- The model checkpoint package should be uploaded as a model repo; this raw package is better uploaded as a dataset repo.
|
| 18 |
+
|
| 19 |
+
Restore raw archive:
|
| 20 |
+
|
| 21 |
+
`ash
|
| 22 |
+
cd raw
|
| 23 |
+
cat PanTS.tar.zst.part-* > PanTS.tar.zst
|
| 24 |
+
tar --use-compress-program zstd -xf PanTS.tar.zst
|
| 25 |
+
`
|
raw_pants_train_test/SHA256SUMS.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
raw_pants_train_test/metadata/package_du.txt
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
337G /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411
|
raw_pants_train_test/metadata/package_filelist.tsv
ADDED
|
@@ -0,0 +1,563 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-015
|
| 2 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-014
|
| 3 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-013
|
| 4 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-012
|
| 5 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-011
|
| 6 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-010
|
| 7 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-009
|
| 8 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-008
|
| 9 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-007
|
| 10 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-006
|
| 11 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-005
|
| 12 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-004
|
| 13 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-003
|
| 14 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-002
|
| 15 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-001
|
| 16 |
+
21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-000
|
| 17 |
+
17684124533 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-016
|
| 18 |
+
47855714 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/captions/captions_v3_after_regen.jsonl
|
| 19 |
+
47443603 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/captions/captions_v2_after_fusion.jsonl
|
| 20 |
+
40606805 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/captions/captions_final.jsonl
|
| 21 |
+
27351511 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/canonical/canonical_facts.jsonl
|
| 22 |
+
6061693 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/captions/captions_v1_initial.jsonl
|
| 23 |
+
3058622 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/resolution_analysis_20260519_canonical_axes/resolution_records_canonical_axes.csv
|
| 24 |
+
2786125 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/resolution_analysis_20260519_8h200/resolution_records.csv
|
| 25 |
+
2554914 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/phase_audit.jsonl
|
| 26 |
+
292750 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/canary/annotations.json
|
| 27 |
+
195689 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/captions/captions_pilot.jsonl
|
| 28 |
+
178248 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/splits/splits.json
|
| 29 |
+
162598 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/captions/captions_v7_pancreas_only.jsonl
|
| 30 |
+
106784 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/annotations.json
|
| 31 |
+
101209 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/annotations.json
|
| 32 |
+
64209 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/MANIFEST.sha256
|
| 33 |
+
53610 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_orient_sanity.jpg
|
| 34 |
+
50611 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000517_axial_raw_17.jpg
|
| 35 |
+
49448 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000517_axial_raw_0.jpg
|
| 36 |
+
48969 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000517_axial_ov_8.jpg
|
| 37 |
+
48944 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000517_axial_ov_0.jpg
|
| 38 |
+
48893 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000517_axial_ov_4.jpg
|
| 39 |
+
47416 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003449_axial_raw_19.jpg
|
| 40 |
+
47040 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/resolution_analysis_20260519_8h200/resolution_summary.json
|
| 41 |
+
46768 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002224_axial_ov_53.jpg
|
| 42 |
+
46703 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_orient_sanity.jpg
|
| 43 |
+
46678 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002224_axial_ov_46.jpg
|
| 44 |
+
46333 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002224_axial_ov_61.jpg
|
| 45 |
+
45841 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002224_axial_raw_32.jpg
|
| 46 |
+
45744 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003449_axial_raw_0.jpg
|
| 47 |
+
45123 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003449_axial_ov_25.jpg
|
| 48 |
+
44896 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003449_axial_ov_26.jpg
|
| 49 |
+
44594 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003449_axial_ov_27.jpg
|
| 50 |
+
43964 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002224_axial_raw_0.jpg
|
| 51 |
+
43897 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_orient_sanity.jpg
|
| 52 |
+
43042 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001237_axial_ov_37.jpg
|
| 53 |
+
42878 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001237_axial_raw_0.jpg
|
| 54 |
+
42672 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001237_axial_ov_55.jpg
|
| 55 |
+
41886 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001237_axial_raw_49.jpg
|
| 56 |
+
41554 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001237_axial_ov_73.jpg
|
| 57 |
+
40777 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_orient_sanity.jpg
|
| 58 |
+
39152 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002085_axial_ov_18.jpg
|
| 59 |
+
38948 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002085_axial_ov_49.jpg
|
| 60 |
+
38780 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002085_axial_ov_33.jpg
|
| 61 |
+
38444 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002085_axial_raw_48.jpg
|
| 62 |
+
38176 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/resolution_analysis_20260519_canonical_axes/resolution_summary_canonical_axes.json
|
| 63 |
+
38054 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002085_axial_raw_0.jpg
|
| 64 |
+
37910 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003485_axial_raw_40.jpg
|
| 65 |
+
36641 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003485_axial_ov_30.jpg
|
| 66 |
+
36529 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_orient_sanity.jpg
|
| 67 |
+
36158 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002224_sagittal_ov_115.jpg
|
| 68 |
+
35968 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_orient_sanity.jpg
|
| 69 |
+
35597 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003485_axial_ov_15.jpg
|
| 70 |
+
35110 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/cache/__pycache__/pants_wan22_cache.cpython-311.pyc
|
| 71 |
+
34978 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_orient_sanity.jpg
|
| 72 |
+
34815 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003485_axial_ov_0.jpg
|
| 73 |
+
34767 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003485_axial_raw_0.jpg
|
| 74 |
+
34622 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S2_axial_ov_78.jpg
|
| 75 |
+
34435 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S2_axial_ov_79.jpg
|
| 76 |
+
34416 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S2_axial_ov_77.jpg
|
| 77 |
+
34340 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S3_axial_ov_56.jpg
|
| 78 |
+
34214 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000465_axial_ov_306.jpg
|
| 79 |
+
34134 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000465_axial_ov_285.jpg
|
| 80 |
+
33881 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S2_axial_ov_42.jpg
|
| 81 |
+
33833 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S3_axial_ov_48.jpg
|
| 82 |
+
33809 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000465_axial_ov_295.jpg
|
| 83 |
+
33530 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S2_axial_ov_40.jpg
|
| 84 |
+
33518 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S2_axial_ov_38.jpg
|
| 85 |
+
33161 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S3_axial_ov_90.jpg
|
| 86 |
+
32959 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S3_axial_ov_82.jpg
|
| 87 |
+
32701 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002224_coronal_ov_118.jpg
|
| 88 |
+
32395 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000725_axial_raw_0.jpg
|
| 89 |
+
32373 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000006_axial_raw_22.jpg
|
| 90 |
+
32373 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S4_axial_raw_22.jpg
|
| 91 |
+
32373 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S1_axial_raw_22.jpg
|
| 92 |
+
32318 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S4_axial_raw_78.jpg
|
| 93 |
+
32100 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S4_axial_raw_77.jpg
|
| 94 |
+
32096 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000006_axial_ov_79.jpg
|
| 95 |
+
32072 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000006_axial_ov_77.jpg
|
| 96 |
+
31935 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S1_axial_ov_79.jpg
|
| 97 |
+
31931 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S1_axial_ov_77.jpg
|
| 98 |
+
31842 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S4_axial_raw_80.jpg
|
| 99 |
+
31287 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000725_axial_ov_0.jpg
|
| 100 |
+
31265 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000465_axial_raw_130.jpg
|
| 101 |
+
31235 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000006_axial_ov_81.jpg
|
| 102 |
+
31089 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S1_axial_ov_81.jpg
|
| 103 |
+
31064 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S4_axial_raw_41.jpg
|
| 104 |
+
31058 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000029_axial_ov_41.jpg
|
| 105 |
+
31053 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001913_axial_ov_107.jpg
|
| 106 |
+
31014 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S4_axial_raw_44.jpg
|
| 107 |
+
30994 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000029_axial_ov_45.jpg
|
| 108 |
+
30967 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001913_axial_ov_99.jpg
|
| 109 |
+
30929 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_orient_sanity.jpg
|
| 110 |
+
30877 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S1_axial_ov_41.jpg
|
| 111 |
+
30864 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S3_axial_ov_75.jpg
|
| 112 |
+
30798 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001913_axial_ov_103.jpg
|
| 113 |
+
30783 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S1_axial_ov_45.jpg
|
| 114 |
+
30739 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000029_axial_ov_38.jpg
|
| 115 |
+
30583 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000725_axial_ov_1.jpg
|
| 116 |
+
30582 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S3_axial_raw_37.jpg
|
| 117 |
+
30511 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S1_axial_ov_38.jpg
|
| 118 |
+
30464 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S4_axial_raw_38.jpg
|
| 119 |
+
30281 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000725_axial_ov_3.jpg
|
| 120 |
+
30026 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S3_axial_raw_36.jpg
|
| 121 |
+
29782 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000725_axial_raw_37.jpg
|
| 122 |
+
29762 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S3_axial_ov_67.jpg
|
| 123 |
+
29548 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001237_sagittal_ov_226.jpg
|
| 124 |
+
29424 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S3_axial_ov_80.jpg
|
| 125 |
+
29133 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S3_axial_ov_72.jpg
|
| 126 |
+
29110 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001438_axial_ov_10.jpg
|
| 127 |
+
29091 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001438_axial_ov_14.jpg
|
| 128 |
+
28961 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002085_sagittal_ov_225.jpg
|
| 129 |
+
28912 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001438_axial_ov_18.jpg
|
| 130 |
+
28766 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001438_axial_raw_15.jpg
|
| 131 |
+
28653 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001438_axial_raw_0.jpg
|
| 132 |
+
28201 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/cache/pants_wan22_cache.py
|
| 133 |
+
28137 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S2_sagittal_ov_94.jpg
|
| 134 |
+
28000 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S3_axial_ov_84.jpg
|
| 135 |
+
27947 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_orient_sanity.jpg
|
| 136 |
+
27873 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S2_axial_ov_62.jpg
|
| 137 |
+
27766 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002232_axial_ov_34.jpg
|
| 138 |
+
27647 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S3_axial_ov_76.jpg
|
| 139 |
+
27618 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000029_axial_raw_21.jpg
|
| 140 |
+
27618 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S4_axial_raw_21.jpg
|
| 141 |
+
27618 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S1_axial_raw_21.jpg
|
| 142 |
+
27500 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000029_axial_raw_0.jpg
|
| 143 |
+
27500 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S4_axial_raw_0.jpg
|
| 144 |
+
27500 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S3_axial_raw_0.jpg
|
| 145 |
+
27500 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S2_axial_raw_0.jpg
|
| 146 |
+
27500 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S1_axial_raw_0.jpg
|
| 147 |
+
27404 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S2_axial_raw_10.jpg
|
| 148 |
+
27115 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S2_axial_ov_53.jpg
|
| 149 |
+
27100 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000035_axial_ov_65.jpg
|
| 150 |
+
27100 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S2_axial_ov_61.jpg
|
| 151 |
+
27079 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S3_axial_raw_26.jpg
|
| 152 |
+
27077 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002232_axial_ov_29.jpg
|
| 153 |
+
27003 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S1_axial_ov_65.jpg
|
| 154 |
+
26932 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S2_sagittal_raw_137.jpg
|
| 155 |
+
26894 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S3_axial_ov_62.jpg
|
| 156 |
+
26892 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S3_axial_ov_54.jpg
|
| 157 |
+
26860 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S2_axial_ov_60.jpg
|
| 158 |
+
26805 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S2_axial_ov_48.jpg
|
| 159 |
+
26781 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S2_axial_ov_43.jpg
|
| 160 |
+
26780 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000002_sagittal_ov_126.jpg
|
| 161 |
+
26750 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S3_axial_ov_77.jpg
|
| 162 |
+
26746 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S2_axial_ov_50.jpg
|
| 163 |
+
26728 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S2_axial_ov_52.jpg
|
| 164 |
+
26702 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S2_sagittal_ov_188.jpg
|
| 165 |
+
26643 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S2_axial_ov_46.jpg
|
| 166 |
+
26630 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S2_axial_ov_41.jpg
|
| 167 |
+
26581 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S4_axial_raw_64.jpg
|
| 168 |
+
26530 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S2_axial_ov_51.jpg
|
| 169 |
+
26469 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S3_axial_ov_23.jpg
|
| 170 |
+
26460 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S3_axial_ov_85.jpg
|
| 171 |
+
26444 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S2_axial_ov_36.jpg
|
| 172 |
+
26433 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S3_axial_ov_16.jpg
|
| 173 |
+
26378 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S3_axial_ov_8.jpg
|
| 174 |
+
26334 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S2_axial_ov_46.jpg
|
| 175 |
+
26306 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002232_axial_ov_25.jpg
|
| 176 |
+
26291 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S2_axial_raw_9.jpg
|
| 177 |
+
26231 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_orient_sanity.jpg
|
| 178 |
+
26216 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S2_sagittal_raw_251.jpg
|
| 179 |
+
26131 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000035_axial_ov_62.jpg
|
| 180 |
+
26082 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002232_axial_raw_18.jpg
|
| 181 |
+
26069 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S2_axial_ov_2.jpg
|
| 182 |
+
26065 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S1_axial_ov_62.jpg
|
| 183 |
+
26002 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000814_axial_raw_23.jpg
|
| 184 |
+
25889 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S3_axial_ov_15.jpg
|
| 185 |
+
25872 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S2_axial_ov_40.jpg
|
| 186 |
+
25866 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S4_axial_raw_62.jpg
|
| 187 |
+
25811 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S2_axial_ov_35.jpg
|
| 188 |
+
25808 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S2_axial_ov_0.jpg
|
| 189 |
+
25800 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002097_axial_ov_27.jpg
|
| 190 |
+
25783 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003485_sagittal_ov_246.jpg
|
| 191 |
+
25742 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S2_axial_ov_4.jpg
|
| 192 |
+
25692 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002097_axial_ov_24.jpg
|
| 193 |
+
25676 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003252_axial_ov_69.jpg
|
| 194 |
+
25675 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000004_sagittal_ov_228.jpg
|
| 195 |
+
25587 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000035_axial_ov_60.jpg
|
| 196 |
+
25519 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S1_axial_ov_60.jpg
|
| 197 |
+
25485 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001030_axial_ov_95.jpg
|
| 198 |
+
25485 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000725_sagittal_ov_119.jpg
|
| 199 |
+
25472 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S2_axial_raw_7.jpg
|
| 200 |
+
25421 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003252_sagittal_ov_101.jpg
|
| 201 |
+
25380 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S4_axial_raw_60.jpg
|
| 202 |
+
25368 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S2_axial_ov_11.jpg
|
| 203 |
+
25225 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002097_axial_ov_21.jpg
|
| 204 |
+
25199 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003252_axial_ov_62.jpg
|
| 205 |
+
25110 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000001_axial_ov_55.jpg
|
| 206 |
+
25082 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001030_axial_ov_90.jpg
|
| 207 |
+
25073 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000814_axial_ov_16.jpg
|
| 208 |
+
25040 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000002_axial_ov_62.jpg
|
| 209 |
+
25019 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003127_axial_raw_0.jpg
|
| 210 |
+
25008 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000002_axial_ov_52.jpg
|
| 211 |
+
24987 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000026_axial_raw_16.jpg
|
| 212 |
+
24987 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S4_axial_raw_16.jpg
|
| 213 |
+
24987 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S1_axial_raw_16.jpg
|
| 214 |
+
24982 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S3_axial_raw_66.jpg
|
| 215 |
+
24957 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S1_axial_ov_62.jpg
|
| 216 |
+
24955 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000004_axial_ov_56.jpg
|
| 217 |
+
24937 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001030_axial_ov_85.jpg
|
| 218 |
+
24935 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S2_axial_ov_9.jpg
|
| 219 |
+
24908 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000026_axial_raw_0.jpg
|
| 220 |
+
24908 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S4_axial_raw_0.jpg
|
| 221 |
+
24908 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S3_axial_raw_0.jpg
|
| 222 |
+
24908 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S2_axial_raw_0.jpg
|
| 223 |
+
24908 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S1_axial_raw_0.jpg
|
| 224 |
+
24905 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S1_axial_ov_52.jpg
|
| 225 |
+
24871 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S1_axial_ov_55.jpg
|
| 226 |
+
24789 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S4_axial_raw_50.jpg
|
| 227 |
+
24788 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S1_axial_ov_56.jpg
|
| 228 |
+
24728 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000002_axial_raw_35.jpg
|
| 229 |
+
24728 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S4_axial_raw_35.jpg
|
| 230 |
+
24728 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S1_axial_raw_35.jpg
|
| 231 |
+
24698 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000814_axial_raw_0.jpg
|
| 232 |
+
24678 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000002_axial_ov_43.jpg
|
| 233 |
+
24655 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003631_axial_raw_0.jpg
|
| 234 |
+
24630 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002097_axial_raw_12.jpg
|
| 235 |
+
24593 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S1_axial_ov_43.jpg
|
| 236 |
+
24574 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S2_axial_ov_8.jpg
|
| 237 |
+
24569 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S4_axial_raw_43.jpg
|
| 238 |
+
24552 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S3_axial_raw_48.jpg
|
| 239 |
+
24513 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003631_axial_raw_22.jpg
|
| 240 |
+
24501 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S4_axial_raw_51.jpg
|
| 241 |
+
24462 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001237_coronal_ov_200.jpg
|
| 242 |
+
24431 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003252_axial_ov_56.jpg
|
| 243 |
+
24428 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000814_axial_ov_0.jpg
|
| 244 |
+
24401 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S2_axial_raw_16.jpg
|
| 245 |
+
24379 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S4_axial_raw_51.jpg
|
| 246 |
+
24356 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S2_coronal_ov_100.jpg
|
| 247 |
+
24352 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002097_axial_raw_0.jpg
|
| 248 |
+
24348 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000001_axial_raw_0.jpg
|
| 249 |
+
24348 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S4_axial_raw_0.jpg
|
| 250 |
+
24348 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S3_axial_raw_0.jpg
|
| 251 |
+
24348 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S2_axial_raw_0.jpg
|
| 252 |
+
24348 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S1_axial_raw_0.jpg
|
| 253 |
+
24325 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000814_axial_ov_8.jpg
|
| 254 |
+
24323 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S3_axial_raw_59.jpg
|
| 255 |
+
24321 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S4_axial_raw_58.jpg
|
| 256 |
+
24304 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000002_axial_raw_0.jpg
|
| 257 |
+
24304 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S4_axial_raw_0.jpg
|
| 258 |
+
24304 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S3_axial_raw_0.jpg
|
| 259 |
+
24304 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S2_axial_raw_0.jpg
|
| 260 |
+
24304 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S1_axial_raw_0.jpg
|
| 261 |
+
24301 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000001_axial_ov_45.jpg
|
| 262 |
+
24237 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S3_axial_ov_22.jpg
|
| 263 |
+
24227 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000465_sagittal_ov_68.jpg
|
| 264 |
+
24125 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S4_axial_raw_43.jpg
|
| 265 |
+
24119 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S3_axial_raw_74.jpg
|
| 266 |
+
24090 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S1_axial_ov_45.jpg
|
| 267 |
+
24028 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000001_axial_ov_36.jpg
|
| 268 |
+
24013 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000004_axial_ov_45.jpg
|
| 269 |
+
23932 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001878_axial_ov_58.jpg
|
| 270 |
+
23911 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003631_axial_ov_41.jpg
|
| 271 |
+
23900 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000725_coronal_ov_111.jpg
|
| 272 |
+
23827 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S1_axial_ov_45.jpg
|
| 273 |
+
23819 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S1_axial_ov_36.jpg
|
| 274 |
+
23813 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000005_axial_raw_29.jpg
|
| 275 |
+
23813 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S4_axial_raw_29.jpg
|
| 276 |
+
23813 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S1_axial_raw_29.jpg
|
| 277 |
+
23785 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003252_coronal_ov_87.jpg
|
| 278 |
+
23753 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S4_axial_raw_36.jpg
|
| 279 |
+
23738 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S2_axial_raw_18.jpg
|
| 280 |
+
23726 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S3_axial_ov_14.jpg
|
| 281 |
+
23726 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S2_axial_ov_14.jpg
|
| 282 |
+
23707 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000004_axial_ov_35.jpg
|
| 283 |
+
23553 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001878_axial_ov_61.jpg
|
| 284 |
+
23551 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S1_axial_ov_35.jpg
|
| 285 |
+
23546 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003631_axial_ov_46.jpg
|
| 286 |
+
23460 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001030_axial_raw_31.jpg
|
| 287 |
+
23457 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000001_axial_raw_39.jpg
|
| 288 |
+
23457 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S4_axial_raw_39.jpg
|
| 289 |
+
23457 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S1_axial_raw_39.jpg
|
| 290 |
+
23445 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003631_axial_ov_36.jpg
|
| 291 |
+
23376 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000005_axial_ov_8.jpg
|
| 292 |
+
23367 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S1_axial_ov_8.jpg
|
| 293 |
+
23363 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S4_axial_raw_50.jpg
|
| 294 |
+
23360 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000004_axial_raw_44.jpg
|
| 295 |
+
23360 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S4_axial_raw_44.jpg
|
| 296 |
+
23360 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S1_axial_raw_44.jpg
|
| 297 |
+
23289 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002232_axial_raw_0.jpg
|
| 298 |
+
23272 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003252_axial_raw_36.jpg
|
| 299 |
+
23245 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S4_axial_raw_43.jpg
|
| 300 |
+
23239 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000026_axial_ov_50.jpg
|
| 301 |
+
23176 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001878_axial_ov_64.jpg
|
| 302 |
+
23167 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S2_axial_raw_13.jpg
|
| 303 |
+
23152 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S4_axial_raw_53.jpg
|
| 304 |
+
23095 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S4_axial_raw_35.jpg
|
| 305 |
+
23060 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000026_axial_ov_52.jpg
|
| 306 |
+
23039 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S3_axial_raw_44.jpg
|
| 307 |
+
22979 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S4_axial_raw_12.jpg
|
| 308 |
+
22968 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S4_axial_raw_51.jpg
|
| 309 |
+
22945 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S2_axial_ov_13.jpg
|
| 310 |
+
22943 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S1_axial_ov_50.jpg
|
| 311 |
+
22932 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00005646_coronal_ov_100.jpg
|
| 312 |
+
22910 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000005_axial_ov_0.jpg
|
| 313 |
+
22898 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S1_axial_ov_0.jpg
|
| 314 |
+
22897 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S4_axial_raw_8.jpg
|
| 315 |
+
22894 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000031_axial_ov_14.jpg
|
| 316 |
+
22889 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000031_axial_raw_10.jpg
|
| 317 |
+
22889 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S4_axial_raw_10.jpg
|
| 318 |
+
22889 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S1_axial_raw_10.jpg
|
| 319 |
+
22874 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000005_axial_ov_4.jpg
|
| 320 |
+
22863 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000026_axial_ov_54.jpg
|
| 321 |
+
22857 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S1_axial_ov_4.jpg
|
| 322 |
+
22817 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S4_axial_raw_6.jpg
|
| 323 |
+
22815 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000031_axial_ov_11.jpg
|
| 324 |
+
22772 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S1_axial_ov_52.jpg
|
| 325 |
+
22771 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S4_axial_raw_3.jpg
|
| 326 |
+
22733 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000004_axial_raw_0.jpg
|
| 327 |
+
22733 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S4_axial_raw_0.jpg
|
| 328 |
+
22733 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S3_axial_raw_0.jpg
|
| 329 |
+
22733 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S2_axial_raw_0.jpg
|
| 330 |
+
22733 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S1_axial_raw_0.jpg
|
| 331 |
+
22724 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S1_axial_ov_14.jpg
|
| 332 |
+
22723 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000005_axial_raw_0.jpg
|
| 333 |
+
22723 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S4_axial_raw_0.jpg
|
| 334 |
+
22723 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S3_axial_raw_0.jpg
|
| 335 |
+
22723 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S2_axial_raw_0.jpg
|
| 336 |
+
22723 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S1_axial_raw_0.jpg
|
| 337 |
+
22706 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S3_axial_raw_16.jpg
|
| 338 |
+
22701 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S1_axial_ov_11.jpg
|
| 339 |
+
22690 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000031_axial_ov_8.jpg
|
| 340 |
+
22684 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002085_coronal_ov_258.jpg
|
| 341 |
+
22630 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S2_axial_raw_12.jpg
|
| 342 |
+
22582 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S1_axial_ov_54.jpg
|
| 343 |
+
22542 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S1_axial_ov_8.jpg
|
| 344 |
+
22347 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/docs/LDM_TRAINING_PLAN_v4.md
|
| 345 |
+
22336 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S2_axial_raw_20.jpg
|
| 346 |
+
22199 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000035_axial_raw_26.jpg
|
| 347 |
+
22199 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S4_axial_raw_26.jpg
|
| 348 |
+
22199 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S1_axial_raw_26.jpg
|
| 349 |
+
22194 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S2_axial_ov_12.jpg
|
| 350 |
+
22184 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000035_axial_raw_0.jpg
|
| 351 |
+
22184 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S4_axial_raw_0.jpg
|
| 352 |
+
22184 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S3_axial_raw_0.jpg
|
| 353 |
+
22184 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S2_axial_raw_0.jpg
|
| 354 |
+
22184 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S1_axial_raw_0.jpg
|
| 355 |
+
22131 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000001_sagittal_ov_224.jpg
|
| 356 |
+
22130 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S2_axial_raw_4.jpg
|
| 357 |
+
22107 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000003_axial_ov_15.jpg
|
| 358 |
+
22062 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003252_axial_raw_0.jpg
|
| 359 |
+
21955 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S1_axial_ov_15.jpg
|
| 360 |
+
21917 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003127_axial_ov_23.jpg
|
| 361 |
+
21852 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003449_sagittal_ov_88.jpg
|
| 362 |
+
21796 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S2_sagittal_raw_256.jpg
|
| 363 |
+
21742 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003127_axial_raw_10.jpg
|
| 364 |
+
21508 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003824_axial_raw_19.jpg
|
| 365 |
+
21433 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001030_axial_raw_0.jpg
|
| 366 |
+
21376 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000003_axial_raw_0.jpg
|
| 367 |
+
21376 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S4_axial_raw_0.jpg
|
| 368 |
+
21376 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S3_axial_raw_0.jpg
|
| 369 |
+
21376 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S2_axial_raw_0.jpg
|
| 370 |
+
21376 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S1_axial_raw_0.jpg
|
| 371 |
+
21357 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000465_axial_raw_0.jpg
|
| 372 |
+
21348 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001913_axial_raw_32.jpg
|
| 373 |
+
21269 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000031_axial_raw_0.jpg
|
| 374 |
+
21269 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S4_axial_raw_0.jpg
|
| 375 |
+
21269 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S3_axial_raw_0.jpg
|
| 376 |
+
21269 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S2_axial_raw_0.jpg
|
| 377 |
+
21269 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S1_axial_raw_0.jpg
|
| 378 |
+
21254 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003127_axial_ov_20.jpg
|
| 379 |
+
21216 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000006_axial_raw_0.jpg
|
| 380 |
+
21216 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S4_axial_raw_0.jpg
|
| 381 |
+
21216 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S3_axial_raw_0.jpg
|
| 382 |
+
21216 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S2_axial_raw_0.jpg
|
| 383 |
+
21216 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S1_axial_raw_0.jpg
|
| 384 |
+
21016 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S4_axial_raw_14.jpg
|
| 385 |
+
20964 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000003_axial_ov_13.jpg
|
| 386 |
+
20815 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S1_axial_ov_13.jpg
|
| 387 |
+
20759 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S2_axial_raw_3.jpg
|
| 388 |
+
20693 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003127_axial_ov_18.jpg
|
| 389 |
+
20689 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S2_sagittal_ov_145.jpg
|
| 390 |
+
20656 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003824_axial_ov_68.jpg
|
| 391 |
+
20616 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S4_axial_raw_13.jpg
|
| 392 |
+
20616 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S3_axial_raw_13.jpg
|
| 393 |
+
20540 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/fusion_run.py
|
| 394 |
+
20481 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000003_axial_raw_7.jpg
|
| 395 |
+
20481 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S4_axial_raw_7.jpg
|
| 396 |
+
20481 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S1_axial_raw_7.jpg
|
| 397 |
+
20479 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000003_axial_ov_12.jpg
|
| 398 |
+
20438 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S3_mip_sagittal_slab_0.jpg
|
| 399 |
+
20387 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S1_axial_ov_12.jpg
|
| 400 |
+
20371 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00005646_sagittal_ov_120.jpg
|
| 401 |
+
20294 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003824_axial_ov_65.jpg
|
| 402 |
+
20137 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S4_axial_raw_12.jpg
|
| 403 |
+
20063 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S2_coronal_raw_99.jpg
|
| 404 |
+
20023 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S3_mip_sagittal_slab_0.jpg
|
| 405 |
+
19926 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001913_axial_raw_0.jpg
|
| 406 |
+
19926 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000002_coronal_ov_115.jpg
|
| 407 |
+
19805 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001913_sagittal_ov_199.jpg
|
| 408 |
+
19766 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003485_coronal_ov_281.jpg
|
| 409 |
+
19415 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003824_axial_ov_62.jpg
|
| 410 |
+
19028 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001878_axial_raw_19.jpg
|
| 411 |
+
18744 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00005646_axial_ov_162.jpg
|
| 412 |
+
18726 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003449_coronal_ov_84.jpg
|
| 413 |
+
18604 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00005646_axial_ov_147.jpg
|
| 414 |
+
18353 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00005646_axial_ov_154.jpg
|
| 415 |
+
18081 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001878_axial_raw_0.jpg
|
| 416 |
+
18072 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/pilot_run.py
|
| 417 |
+
17915 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000002_S3_mip_coronal_slab_0.jpg
|
| 418 |
+
17585 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003824_axial_raw_0.jpg
|
| 419 |
+
17465 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S2_coronal_ov_164.jpg
|
| 420 |
+
17278 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S2_coronal_ov_151.jpg
|
| 421 |
+
16874 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000517_sagittal_ov_195.jpg
|
| 422 |
+
16419 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001030_sagittal_ov_224.jpg
|
| 423 |
+
15823 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00005646_axial_raw_42.jpg
|
| 424 |
+
15814 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S2_sagittal_ov_168.jpg
|
| 425 |
+
15730 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S2_sagittal_raw_256.jpg
|
| 426 |
+
15679 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S3_mip_sagittal_slab_0.jpg
|
| 427 |
+
15677 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S2_sagittal_raw_209.jpg
|
| 428 |
+
15624 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000004_coronal_ov_183.jpg
|
| 429 |
+
15527 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000029_sagittal_ov_221.jpg
|
| 430 |
+
15464 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000465_coronal_ov_214.jpg
|
| 431 |
+
15293 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S2_coronal_ov_179.jpg
|
| 432 |
+
15293 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S2_coronal_ov_175.jpg
|
| 433 |
+
15192 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001913_coronal_ov_201.jpg
|
| 434 |
+
15139 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000035_sagittal_ov_212.jpg
|
| 435 |
+
15058 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S2_sagittal_ov_141.jpg
|
| 436 |
+
14864 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S2_sagittal_ov_170.jpg
|
| 437 |
+
14493 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003631_sagittal_ov_225.jpg
|
| 438 |
+
14490 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S2_coronal_raw_162.jpg
|
| 439 |
+
14402 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S2_coronal_ov_212.jpg
|
| 440 |
+
14385 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S2_coronal_raw_166.jpg
|
| 441 |
+
14280 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/canonical_facts.py
|
| 442 |
+
14256 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S2_sagittal_ov_191.jpg
|
| 443 |
+
14153 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S2_coronal_ov_176.jpg
|
| 444 |
+
14113 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000006_sagittal_ov_248.jpg
|
| 445 |
+
14070 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S2_sagittal_raw_256.jpg
|
| 446 |
+
14062 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/pilot_run_v2.py
|
| 447 |
+
14045 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000001_coronal_ov_197.jpg
|
| 448 |
+
14039 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000005_sagittal_ov_244.jpg
|
| 449 |
+
13982 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000004_S3_mip_coronal_slab_0.jpg
|
| 450 |
+
13931 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S2_sagittal_raw_256.jpg
|
| 451 |
+
13638 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S2_coronal_raw_195.jpg
|
| 452 |
+
13611 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000001_S3_mip_coronal_slab_0.jpg
|
| 453 |
+
13557 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000029_coronal_ov_193.jpg
|
| 454 |
+
13534 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003824_sagittal_ov_254.jpg
|
| 455 |
+
13533 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000035_coronal_ov_223.jpg
|
| 456 |
+
13409 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/f5_audit.py
|
| 457 |
+
13363 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000517_coronal_ov_165.jpg
|
| 458 |
+
13223 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000814_sagittal_ov_193.jpg
|
| 459 |
+
13210 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002232_sagittal_ov_225.jpg
|
| 460 |
+
13157 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S3_mip_sagittal_slab_0.jpg
|
| 461 |
+
13139 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000006_coronal_ov_250.jpg
|
| 462 |
+
13044 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000035_S3_mip_coronal_slab_0.jpg
|
| 463 |
+
13022 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001878_sagittal_ov_212.jpg
|
| 464 |
+
12952 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S3_mip_coronal_slab_0.jpg
|
| 465 |
+
12846 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S2_coronal_raw_215.jpg
|
| 466 |
+
12810 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001030_coronal_ov_240.jpg
|
| 467 |
+
12777 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S2_coronal_raw_170.jpg
|
| 468 |
+
12772 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000029_S3_mip_sagittal_slab_0.jpg
|
| 469 |
+
11877 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001438_sagittal_ov_177.jpg
|
| 470 |
+
11795 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S3_mip_sagittal_slab_0.jpg
|
| 471 |
+
11723 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002232_coronal_ov_187.jpg
|
| 472 |
+
11618 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/docs/PROPOSAL_v1.md
|
| 473 |
+
11567 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000005_coronal_ov_210.jpg
|
| 474 |
+
11458 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S2_coronal_raw_215.jpg
|
| 475 |
+
11408 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S2_sagittal_ov_159.jpg
|
| 476 |
+
11379 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000006_S3_mip_coronal_slab_0.jpg
|
| 477 |
+
11287 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/vae_sanity_gate.py
|
| 478 |
+
11237 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S3_mip_coronal_slab_0.jpg
|
| 479 |
+
11190 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000026_sagittal_ov_224.jpg
|
| 480 |
+
11031 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003631_coronal_ov_169.jpg
|
| 481 |
+
10917 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S2_sagittal_raw_256.jpg
|
| 482 |
+
10721 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000005_S3_mip_sagittal_slab_0.jpg
|
| 483 |
+
10540 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00005646_axial_raw_0.jpg
|
| 484 |
+
10448 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/docs/PROPOSAL_v3_1.md
|
| 485 |
+
10320 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003824_coronal_ov_176.jpg
|
| 486 |
+
10261 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001878_coronal_ov_175.jpg
|
| 487 |
+
10258 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000814_coronal_ov_204.jpg
|
| 488 |
+
10232 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S2_coronal_ov_189.jpg
|
| 489 |
+
10089 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/docs/PROPOSAL_v3_2.md
|
| 490 |
+
9977 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/preprocessing.py
|
| 491 |
+
9935 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00001438_coronal_ov_207.jpg
|
| 492 |
+
9782 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S3_mip_sagittal_slab_0.jpg
|
| 493 |
+
9685 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/f5b_v7.py
|
| 494 |
+
9563 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/README.md
|
| 495 |
+
9461 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S2_sagittal_ov_145.jpg
|
| 496 |
+
9294 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000026_coronal_ov_222.jpg
|
| 497 |
+
9202 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002097_sagittal_ov_210.jpg
|
| 498 |
+
9163 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/dry_run_train.py
|
| 499 |
+
9105 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S2_coronal_raw_181.jpg
|
| 500 |
+
9097 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/audit_v2.md
|
| 501 |
+
8639 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/pack_webdataset.py
|
| 502 |
+
8587 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/full_vlm_run2.py
|
| 503 |
+
8578 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003127_sagittal_ov_229.jpg
|
| 504 |
+
8559 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S2_sagittal_raw_239.jpg
|
| 505 |
+
8424 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000031_sagittal_ov_213.jpg
|
| 506 |
+
8370 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/phase_audit.py
|
| 507 |
+
8276 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000026_S3_mip_coronal_slab_0.jpg
|
| 508 |
+
8159 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S2_coronal_ov_184.jpg
|
| 509 |
+
8102 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S2_sagittal_ov_190.jpg
|
| 510 |
+
7798 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/full_vlm_run.py
|
| 511 |
+
7765 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00002097_coronal_ov_188.jpg
|
| 512 |
+
7496 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000003_sagittal_ov_239.jpg
|
| 513 |
+
7450 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/docs/PRETRAIN_CHECKLIST_SUMMARY.md
|
| 514 |
+
7337 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S2_sagittal_raw_247.jpg
|
| 515 |
+
7328 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S3_mip_sagittal_slab_0.jpg
|
| 516 |
+
7214 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/f5d_regen.py
|
| 517 |
+
7184 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00003127_coronal_ov_191.jpg
|
| 518 |
+
7052 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000031_coronal_ov_216.jpg
|
| 519 |
+
7029 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S3_mip_sagittal_slab_0.jpg
|
| 520 |
+
6949 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S2_coronal_raw_181.jpg
|
| 521 |
+
6628 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/cache/pants_wan22_text_cache.py
|
| 522 |
+
6464 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000031_S3_mip_coronal_slab_0.jpg
|
| 523 |
+
6346 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/salvage_vlm.py
|
| 524 |
+
6263 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S2_coronal_ov_189.jpg
|
| 525 |
+
6000 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/fusion_retry.py
|
| 526 |
+
5789 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S2_coronal_raw_174.jpg
|
| 527 |
+
5769 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/previews/PanTS_00000003_coronal_ov_214.jpg
|
| 528 |
+
5685 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/f6_audit_final.py
|
| 529 |
+
5623 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/previews/PanTS_00000003_S3_mip_coronal_slab_0.jpg
|
| 530 |
+
5603 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/caption_generation/f6_merge.py
|
| 531 |
+
5572 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/cache/pants_wan22_decode_check.py
|
| 532 |
+
5513 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/schedule.py
|
| 533 |
+
5117 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/docs/fusion_variants_design.md
|
| 534 |
+
4235 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/resolution_analysis_20260519_8h200/resolution_report.md
|
| 535 |
+
4085 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/package_tree_depth3.txt
|
| 536 |
+
4043 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/test_preprocessing.py
|
| 537 |
+
4002 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/canary/ids.json
|
| 538 |
+
3766 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/resolution_analysis_20260519_canonical_axes/resolution_report_canonical_axes.md
|
| 539 |
+
3560 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/cache/run_pants_wan22_finetune_fullrep_8gpu.sh
|
| 540 |
+
3376 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/cache/run_pants_wan22_finetune_fullrep_8gpu_node.sh
|
| 541 |
+
3179 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/UPLOAD_INSTRUCTIONS.md
|
| 542 |
+
3094 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/resolution_analysis_20260519_canonical_axes/bucket_spec_v2_proposal.json
|
| 543 |
+
3014 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/build_bucket_spec.py
|
| 544 |
+
2661 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/splits/bucket_spec.json
|
| 545 |
+
2501 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/LICENSE
|
| 546 |
+
2283 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/cache/run_pants_wan22_finetune_smoke_8gpu.sh
|
| 547 |
+
1798 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/ids_meta.json
|
| 548 |
+
1673 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/aug_config.py
|
| 549 |
+
1630 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/docs/GPT_REVIEW.md
|
| 550 |
+
1590 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/cache/run_pants_wan22_cache_8gpu.sh
|
| 551 |
+
1491 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/phase_audit_summary.md
|
| 552 |
+
1396 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/build_splits.py
|
| 553 |
+
1379 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/resolution_analysis_20260519_canonical_axes/bucket_spec_v2_proposal.md
|
| 554 |
+
1309 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/code/cache/run_pants_wan22_text_cache_8gpu.sh
|
| 555 |
+
1113 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/README.md
|
| 556 |
+
982 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/raw_summary.txt
|
| 557 |
+
748 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/audit/audit_final.md
|
| 558 |
+
602 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot_v2/ids.json
|
| 559 |
+
180 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/samples/pilot/ids.json
|
| 560 |
+
93 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/package_du.txt
|
| 561 |
+
91 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/RESTORE.txt
|
| 562 |
+
0 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/package_filelist.tsv
|
| 563 |
+
0 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/PACKAGE_COMPLETE
|
raw_pants_train_test/metadata/package_tree_depth3.txt
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
2026-05-12 16:56 2501 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/LICENSE
|
| 2 |
+
2026-05-12 16:56 3179 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/UPLOAD_INSTRUCTIONS.md
|
| 3 |
+
2026-05-12 16:56 9563 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/README.md
|
| 4 |
+
2026-05-12 16:56 64209 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/pants-captions-ldm/MANIFEST.sha256
|
| 5 |
+
2026-05-23 10:25 91 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/RESTORE.txt
|
| 6 |
+
2026-05-23 10:25 982 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/raw_summary.txt
|
| 7 |
+
2026-05-23 10:25 1113 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/README.md
|
| 8 |
+
2026-05-23 10:26 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-000
|
| 9 |
+
2026-05-23 10:27 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-001
|
| 10 |
+
2026-05-23 10:27 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-002
|
| 11 |
+
2026-05-23 10:28 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-003
|
| 12 |
+
2026-05-23 10:29 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-004
|
| 13 |
+
2026-05-23 10:29 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-005
|
| 14 |
+
2026-05-23 10:30 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-006
|
| 15 |
+
2026-05-23 10:30 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-007
|
| 16 |
+
2026-05-23 10:31 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-008
|
| 17 |
+
2026-05-23 10:32 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-009
|
| 18 |
+
2026-05-23 10:32 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-010
|
| 19 |
+
2026-05-23 10:33 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-011
|
| 20 |
+
2026-05-23 10:34 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-012
|
| 21 |
+
2026-05-23 10:34 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-013
|
| 22 |
+
2026-05-23 10:35 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-014
|
| 23 |
+
2026-05-23 10:36 21474836480 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-015
|
| 24 |
+
2026-05-23 10:48 17684124533 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/raw/PanTS.tar.zst.part-016
|
| 25 |
+
2026-05-23 10:56 0 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/PACKAGE_COMPLETE
|
| 26 |
+
2026-05-23 10:56 4085 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/package_tree_depth3.txt
|
| 27 |
+
2026-05-23 11:00 98486 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/package_filelist.tsv
|
| 28 |
+
2026-05-23 11:06 93 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/metadata/package_du.txt
|
| 29 |
+
2026-05-23 11:06 132226 /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411/SHA256SUMS.txt
|
raw_pants_train_test/metadata/pants-captions-ldm/LICENSE
ADDED
|
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
DUAL LICENSE — this release combines code and data under different licenses.
|
| 2 |
+
|
| 3 |
+
==========================================================================
|
| 4 |
+
CODE (everything under code/)
|
| 5 |
+
==========================================================================
|
| 6 |
+
MIT License
|
| 7 |
+
|
| 8 |
+
Copyright (c) 2026 Anonymous Contributor (PanTS-Captions)
|
| 9 |
+
|
| 10 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 11 |
+
of this software and associated documentation files (the "Software"), to deal
|
| 12 |
+
in the Software without restriction, including without limitation the rights
|
| 13 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 14 |
+
copies of the Software, and to permit persons to whom the Software is
|
| 15 |
+
furnished to do so, subject to the following conditions:
|
| 16 |
+
|
| 17 |
+
The above copyright notice and this permission notice shall be included in all
|
| 18 |
+
copies or substantial portions of the Software.
|
| 19 |
+
|
| 20 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 21 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 22 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 23 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 24 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 25 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 26 |
+
SOFTWARE.
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
==========================================================================
|
| 30 |
+
DATA (captions/, canonical/, splits/, samples/, audit/, docs/)
|
| 31 |
+
==========================================================================
|
| 32 |
+
Creative Commons Attribution 4.0 International (CC-BY-4.0)
|
| 33 |
+
|
| 34 |
+
You are free to:
|
| 35 |
+
- Share — copy and redistribute the material in any medium or format
|
| 36 |
+
- Adapt — remix, transform, and build upon the material for any purpose,
|
| 37 |
+
even commercially
|
| 38 |
+
|
| 39 |
+
Under the following terms:
|
| 40 |
+
- Attribution — You must give appropriate credit, provide a link to the
|
| 41 |
+
license, and indicate if changes were made. You may do so
|
| 42 |
+
in any reasonable manner, but not in any way that suggests
|
| 43 |
+
the licensor endorses you or your use.
|
| 44 |
+
|
| 45 |
+
Full license text: https://creativecommons.org/licenses/by/4.0/legalcode
|
| 46 |
+
|
| 47 |
+
==========================================================================
|
| 48 |
+
UPSTREAM PanTS CT VOLUMES + MASKS
|
| 49 |
+
==========================================================================
|
| 50 |
+
Not redistributed by this release. Obtain from the original PanTS distributor
|
| 51 |
+
under its own license.
|
raw_pants_train_test/metadata/pants-captions-ldm/MANIFEST.sha256
ADDED
|
@@ -0,0 +1,521 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
498e8654f7822e493619b68d59e1877e0cd67aaf1a4e8a3b18ea60956b2fb80e ./LICENSE
|
| 2 |
+
8266f17e5db1c526005a82683d203f120a4b80843ddff41176964f84c4b4e37b ./README.md
|
| 3 |
+
1d268bcc63d2464d4cc6802a3f2321730b199992f216d495ae4bee948aca95a3 ./UPLOAD_INSTRUCTIONS.md
|
| 4 |
+
317c75adc491d5acd8728e7de2cb4db8499ef14788c723a6be6a6e6cf69629f6 ./audit/audit_final.md
|
| 5 |
+
baa5619b03d61498742d7ece876a60a9825c965317e418e6ca0fabf50fc8a232 ./audit/audit_v2.md
|
| 6 |
+
fdb72166d367462dd054adfb18268c70ed6f3deb7271f773fcd94dd37f471463 ./audit/phase_audit.jsonl
|
| 7 |
+
6d1af108d28ec2d93ff2bef545f371139965f18f8f6a6dcd48067e0bfa1ec67d ./audit/phase_audit_summary.md
|
| 8 |
+
8c0a095e3cfea0e492e27e30538445ff11da5d36b6ba67bee14be057791cdeff ./canonical/canonical_facts.jsonl
|
| 9 |
+
172fad40f3997eeec3307096db069f54e6797839a0b52c923ddcb2f356a4bad0 ./captions/captions_final.jsonl
|
| 10 |
+
69ab5f8714108e866c711ad3299e0be492c9626e987caf81bf77ea41ffc9910b ./captions/captions_pilot.jsonl
|
| 11 |
+
e319a8884ba7e9c492e1f2efe57ab0e2a86205c5a1e7595c7d5e253833db25a2 ./captions/captions_v1_initial.jsonl
|
| 12 |
+
1f540f20d20a92523425b34d139fe860d6fdbca456bc8c0a4d124c3a0b1a1ecf ./captions/captions_v2_after_fusion.jsonl
|
| 13 |
+
7782e3cc66c7cc4283283c6a06fbad6637baf93c8c96c86a8a14788575b275c3 ./captions/captions_v3_after_regen.jsonl
|
| 14 |
+
d74736fdb8c497fe4a7cdf02b1bc29ce62f2c1aab8e7319c4fd09675dd0ffe59 ./captions/captions_v7_pancreas_only.jsonl
|
| 15 |
+
30b83dd29caa87e4349bd9468c9f1c0873df0d71da7ad305917d2ccbd7c095d1 ./code/aug_config.py
|
| 16 |
+
5f5a21bbc3cd529da264709bc226160c09a33bd55936d3975c736d7960eb7f65 ./code/build_bucket_spec.py
|
| 17 |
+
f52b7ae61a0a94d9c617458b12fa98e8a845929b8827a512502007c6259df0fe ./code/build_splits.py
|
| 18 |
+
bf24141ffe8b659f2b9581ef62895c84613fe6442cd60693d876f635e0c9cda9 ./code/caption_generation/canonical_facts.py
|
| 19 |
+
b7fb495368800594a2e1915de91563723484e85d6520f68939a3113fdb825f8e ./code/caption_generation/f5_audit.py
|
| 20 |
+
81273ade025a7bc7594e62ef6d4f759568171aebf7952dc5315f51ccb8eb7c1d ./code/caption_generation/f5b_v7.py
|
| 21 |
+
4d15f947da8a9482dbcbb31c5f8c55e541498015134ac6546a434793093cf8c9 ./code/caption_generation/f5d_regen.py
|
| 22 |
+
c42a9c1adaca7385e8523c0a56ca86bd1702113407f649259005cff587ac8945 ./code/caption_generation/f6_audit_final.py
|
| 23 |
+
a2a5c4c7164d8d7a665fadbb6230f6d606f5d821c23f0caecb1f4e471a60de09 ./code/caption_generation/f6_merge.py
|
| 24 |
+
315d8e603939cae828302c491e57f9c0ff0f189665d31173473600f180b2f943 ./code/caption_generation/full_vlm_run.py
|
| 25 |
+
e1d70808698c578fd8630e2ea9b08006932a1c3b50baea6157e3ab2e07d1dc07 ./code/caption_generation/full_vlm_run2.py
|
| 26 |
+
a932b8f6d66c838da9e1db285a9e97fb313ddfc7cf831bb6dc9d886474607046 ./code/caption_generation/fusion_retry.py
|
| 27 |
+
af7f36ad488056b800791533e58cb6f2171c43d151531a32ceab82d5448cfb6d ./code/caption_generation/fusion_run.py
|
| 28 |
+
8e3edd4c731b0bda9365e2ac69450a098362cd8dfa51a51bb5b1b6d6c81a0191 ./code/caption_generation/pilot_run.py
|
| 29 |
+
a4dead32780ab5f8d1b9ef9aa49afa3abea5d566caff10312b6082291df82189 ./code/caption_generation/pilot_run_v2.py
|
| 30 |
+
4c969ecab65a841223cb8001da453ae2ab8e3efe15374b5e7d6724fe5867b4e3 ./code/caption_generation/salvage_vlm.py
|
| 31 |
+
ef4eb89a069ff065d6167c8d5d7528917c823a5b0df8317be4812f59b44a4056 ./code/dry_run_train.py
|
| 32 |
+
2275f30ca26fd9fe4eab70c038e8ee4492547e080998794c5dce75fa330d4ddb ./code/pack_webdataset.py
|
| 33 |
+
649cdea25d8f292de4eacbec85e36332fcfbbd8798a70d153a88f7df94811e54 ./code/phase_audit.py
|
| 34 |
+
e76016e3bf6996ff608877b3ed0b44df220c5a8937c99375c6a55ff42de72c66 ./code/preprocessing.py
|
| 35 |
+
c03959153c5ad6f2e29330cf7790207de82104b69b171011aed22257c1e2dcd0 ./code/schedule.py
|
| 36 |
+
bb1809ea4dec40f45d4b871a587729154fbd1373a7b4182cb1c102a7da009a3a ./code/test_preprocessing.py
|
| 37 |
+
deeae95643ee0cf8298fd51d3fb1c16887225be245ec2aa312a741b749bdfb7c ./code/vae_sanity_gate.py
|
| 38 |
+
3e830437bf1da5b32e295e83e2a7aa786c67dcfa2fe7e350b5d69a7b06bf75ba ./docs/GPT_REVIEW.md
|
| 39 |
+
c3fe988fb386d94c3dc0e109f4844945980f84e43f5d6370858bb2239865cc70 ./docs/LDM_TRAINING_PLAN_v4.md
|
| 40 |
+
00dfb6b78212fb24c4e113b95b785be3bab5316427976a8519ced47c331dfb75 ./docs/PRETRAIN_CHECKLIST_SUMMARY.md
|
| 41 |
+
20dc4bd3fda248f763f3702773c868d528e679be32e24d74f56a20627dbd4526 ./docs/PROPOSAL_v1.md
|
| 42 |
+
9011ffd2c1090a92839b77b6cc8e2d481bb164e31c4c578e8c5ba8057561a79d ./docs/PROPOSAL_v3_1.md
|
| 43 |
+
f258a45a0945264e0399c7ac571147032428273562ed480c8559d9da3a555526 ./docs/PROPOSAL_v3_2.md
|
| 44 |
+
bd2a5c7ea1a1147872c75429ccb9cd9c97684891d122ac3c5bf00f89f203c7ed ./docs/fusion_variants_design.md
|
| 45 |
+
22f22418a3bce670f4373d23d5d858135cdf75ed94060a6136c7cb32341ba1fe ./samples/canary/annotations.json
|
| 46 |
+
78dbc14612f95262ef70bad354b4288549ae698c493b8c7170e66c17530078a2 ./samples/canary/ids.json
|
| 47 |
+
f387a04dd6c3e95463dbb7db8d95b12857804b08798b70589d93fbfc8975eaf0 ./samples/pilot/annotations.json
|
| 48 |
+
66960251dbcd989d9822850a4f9ab50781f10e6303ec212557c76d92fb540ed6 ./samples/pilot/ids.json
|
| 49 |
+
e05b8beaf873b48f90f56afd3219653ab61f3fdcbb6f00af62f8f0d5e0274612 ./samples/pilot/previews/PanTS_00000001_S1_axial_ov_36.jpg
|
| 50 |
+
f415a0ddf2510a688eb1ac3f90d2da68820e79fd7c1527511528183e6326dfd9 ./samples/pilot/previews/PanTS_00000001_S1_axial_ov_45.jpg
|
| 51 |
+
cddc27065d54c466b16be7d271a57ad38561a422ce2a55b9c8f0743a5b4706dd ./samples/pilot/previews/PanTS_00000001_S1_axial_ov_55.jpg
|
| 52 |
+
0c33a93876518d827f81648c40c8bdd96303daf7f38e88671681d2d84418b668 ./samples/pilot/previews/PanTS_00000001_S1_axial_raw_0.jpg
|
| 53 |
+
1d27d54e758a7ed2c3740159da73a4802330afb59de6afa15bbbaf561960e556 ./samples/pilot/previews/PanTS_00000001_S1_axial_raw_39.jpg
|
| 54 |
+
13c53a5f932a320cc40b8e3a4502fa1d708ec90b8a1d4fcbd0bf6b3fa88cfd4e ./samples/pilot/previews/PanTS_00000001_S2_axial_ov_36.jpg
|
| 55 |
+
438deb6fbd6823ce1a078a03c3ea64fbb1b95c1fc79c7ebbca1815d2920ade3b ./samples/pilot/previews/PanTS_00000001_S2_axial_ov_41.jpg
|
| 56 |
+
8cbee761cdc26a8e4ab9c846562e8f3a4ab67f2ba91b7389308256e64fb7a620 ./samples/pilot/previews/PanTS_00000001_S2_axial_ov_46.jpg
|
| 57 |
+
0c33a93876518d827f81648c40c8bdd96303daf7f38e88671681d2d84418b668 ./samples/pilot/previews/PanTS_00000001_S2_axial_raw_0.jpg
|
| 58 |
+
4b7c6a30562701519bfb6b6cbc30bead679456e7664575e4cacaec7716abe066 ./samples/pilot/previews/PanTS_00000001_S2_axial_raw_18.jpg
|
| 59 |
+
016637378af44f1a3add9aa3e4d8e64c99f37d1c301e838a78c19bb280cb7155 ./samples/pilot/previews/PanTS_00000001_S2_coronal_ov_164.jpg
|
| 60 |
+
7a1adbf6f828aa4998c98854d51d953edd7c7095fa2e943a61dfe87cab02eb77 ./samples/pilot/previews/PanTS_00000001_S2_coronal_raw_166.jpg
|
| 61 |
+
ecdc45c1c7d28ad3c6115886be5b9bae36ab98fa884b09654c9f10bc795d8b37 ./samples/pilot/previews/PanTS_00000001_S2_sagittal_ov_145.jpg
|
| 62 |
+
ecaaf61c9b2a8775a907d065e5eb983f30b92c602ef75bac603b3c4e3be0cbf9 ./samples/pilot/previews/PanTS_00000001_S2_sagittal_raw_256.jpg
|
| 63 |
+
eb04851b283b0cb64f5b531b6bcd857680bdaaa9339be8bf30b8f486a61cf49a ./samples/pilot/previews/PanTS_00000001_S3_axial_ov_72.jpg
|
| 64 |
+
fd28ceade3247227516b50be475265ab5b7489b421151cccebd0ddf429a2d994 ./samples/pilot/previews/PanTS_00000001_S3_axial_ov_80.jpg
|
| 65 |
+
0c33a93876518d827f81648c40c8bdd96303daf7f38e88671681d2d84418b668 ./samples/pilot/previews/PanTS_00000001_S3_axial_raw_0.jpg
|
| 66 |
+
c4152164f1bb4931d82beee80bd2296e80060948d820acce4d56129f86941ed2 ./samples/pilot/previews/PanTS_00000001_S3_axial_raw_66.jpg
|
| 67 |
+
c55cb3bb720d2964e7ff01df158d8c9f00b44ac5320eaff252f8fed7bbf3ebec ./samples/pilot/previews/PanTS_00000001_S3_mip_coronal_slab_0.jpg
|
| 68 |
+
3f44f239d7e0cd25709055e68bfc3d118de44e862fb1414896daef76b2b9716e ./samples/pilot/previews/PanTS_00000001_S3_mip_sagittal_slab_0.jpg
|
| 69 |
+
0c33a93876518d827f81648c40c8bdd96303daf7f38e88671681d2d84418b668 ./samples/pilot/previews/PanTS_00000001_S4_axial_raw_0.jpg
|
| 70 |
+
2fb573e932ff2da1182da04881492817096ffc8391a3d7ed5529c059d585eea4 ./samples/pilot/previews/PanTS_00000001_S4_axial_raw_36.jpg
|
| 71 |
+
1d27d54e758a7ed2c3740159da73a4802330afb59de6afa15bbbaf561960e556 ./samples/pilot/previews/PanTS_00000001_S4_axial_raw_39.jpg
|
| 72 |
+
9aa5df3e902a53dd3978764df9a6d3c0b884f1b01546c20f6c4be0f974326a97 ./samples/pilot/previews/PanTS_00000001_S4_axial_raw_43.jpg
|
| 73 |
+
06e825e7fea6e2a7cee6e2b2650b9f56dbb3b4e0cc29d3c76c1f28ef656e540b ./samples/pilot/previews/PanTS_00000001_S4_axial_raw_51.jpg
|
| 74 |
+
a72a2edbcb68fb527bf47dda54accfd6cabb558b7d080bb495edd213a096006a ./samples/pilot/previews/PanTS_00000001_orient_sanity.jpg
|
| 75 |
+
6181eba1d8ea4eab37886cdfe776904adf17611bcf548322aef17acd9aa7b9fd ./samples/pilot/previews/PanTS_00000002_S1_axial_ov_43.jpg
|
| 76 |
+
93f593dca937b54a7941949a714c824a287a86e176f8c8011cc20484130bbfb2 ./samples/pilot/previews/PanTS_00000002_S1_axial_ov_52.jpg
|
| 77 |
+
9894e683410bb1bca8c7e5c619efbdca4b7a0f3331b69c4af2b372edb92c33e1 ./samples/pilot/previews/PanTS_00000002_S1_axial_ov_62.jpg
|
| 78 |
+
528667ceb887b4e6978a432d0822eacb46384d1442fdd2143bf8a1747be49745 ./samples/pilot/previews/PanTS_00000002_S1_axial_raw_0.jpg
|
| 79 |
+
48bbae938245965b5408d26853be391d7ef0ecd57958761a718d195317bd6b19 ./samples/pilot/previews/PanTS_00000002_S1_axial_raw_35.jpg
|
| 80 |
+
ceb744977c5af3d285ed302d04fddda90655bfa5eda6f086b881693ba0b1e4b6 ./samples/pilot/previews/PanTS_00000002_S2_axial_ov_43.jpg
|
| 81 |
+
1165de80ed222c2af9ab36afd81b2c51745ddde88cab2bd775e69cb47d8d35be ./samples/pilot/previews/PanTS_00000002_S2_axial_ov_48.jpg
|
| 82 |
+
97d3b899e4975a4769e0239c21574784fdcc624fb60095bac9043404d883e3fc ./samples/pilot/previews/PanTS_00000002_S2_axial_ov_53.jpg
|
| 83 |
+
528667ceb887b4e6978a432d0822eacb46384d1442fdd2143bf8a1747be49745 ./samples/pilot/previews/PanTS_00000002_S2_axial_raw_0.jpg
|
| 84 |
+
828ee355a93d52da79e2f2078583896cd6239af422071db456e6d5f4a24aa516 ./samples/pilot/previews/PanTS_00000002_S2_axial_raw_16.jpg
|
| 85 |
+
c229f4a47be384e3bf40d87745b3c346a4791177f1601bfd0a7a0244722219a0 ./samples/pilot/previews/PanTS_00000002_S2_coronal_ov_100.jpg
|
| 86 |
+
cef496932e37d7c38b7798b773ed5b0b82a9b8255596804aa71fd1f4e04cd9f7 ./samples/pilot/previews/PanTS_00000002_S2_coronal_raw_99.jpg
|
| 87 |
+
35348b47ce04a466f8aafbaa6e31e7e3dd3ae9d7ff87698f4c0e21a0bfb321b8 ./samples/pilot/previews/PanTS_00000002_S2_sagittal_ov_94.jpg
|
| 88 |
+
b13719a8d8c73d54413e4a33ddb917c35b396c45e5de0c28901089f8ec4d826a ./samples/pilot/previews/PanTS_00000002_S2_sagittal_raw_137.jpg
|
| 89 |
+
aba4582ecd050024d962bdfeed8ed7c9a8ef1f2d619f8e3614fd6b6cfd9827ac ./samples/pilot/previews/PanTS_00000002_S3_axial_ov_77.jpg
|
| 90 |
+
5690095c32b1fad9fbd683343ed535e401842884f48957b99d75913ea72a0ba8 ./samples/pilot/previews/PanTS_00000002_S3_axial_ov_85.jpg
|
| 91 |
+
528667ceb887b4e6978a432d0822eacb46384d1442fdd2143bf8a1747be49745 ./samples/pilot/previews/PanTS_00000002_S3_axial_raw_0.jpg
|
| 92 |
+
8553bd6bdd7032f04c19ec87beb8f76e052307dbb25832c211445c988d7f0a58 ./samples/pilot/previews/PanTS_00000002_S3_axial_raw_59.jpg
|
| 93 |
+
620c607aff6b3ecfd0dd587e3236aa4703e80b1c55f00e04dec8f014af2117bb ./samples/pilot/previews/PanTS_00000002_S3_mip_coronal_slab_0.jpg
|
| 94 |
+
70eaed1d69c218a0d7737e53a14bd91c0cc83c7c27a092996b4ee248cad0480c ./samples/pilot/previews/PanTS_00000002_S3_mip_sagittal_slab_0.jpg
|
| 95 |
+
528667ceb887b4e6978a432d0822eacb46384d1442fdd2143bf8a1747be49745 ./samples/pilot/previews/PanTS_00000002_S4_axial_raw_0.jpg
|
| 96 |
+
48bbae938245965b5408d26853be391d7ef0ecd57958761a718d195317bd6b19 ./samples/pilot/previews/PanTS_00000002_S4_axial_raw_35.jpg
|
| 97 |
+
e8fa148cf4b469a9c066cf622f32aaf7c391cae27cb338398c35b50e59c017b0 ./samples/pilot/previews/PanTS_00000002_S4_axial_raw_43.jpg
|
| 98 |
+
0d867f50eeb94b658f13c159cd293724a6615892d9cb94ed204ea6b756c9315f ./samples/pilot/previews/PanTS_00000002_S4_axial_raw_50.jpg
|
| 99 |
+
4a3e82059c7004dd364c223a85beebf9d32e74b8c285388b663f820a7f3dc7b9 ./samples/pilot/previews/PanTS_00000002_S4_axial_raw_58.jpg
|
| 100 |
+
13f93cbfd914e9d6feef8fefc6e7efdec733e2b9ae329fb67c3c54323b5d9583 ./samples/pilot/previews/PanTS_00000002_orient_sanity.jpg
|
| 101 |
+
696beb386bebab4ff97726cda7e111da7d062454cf82d0750921847f6244b4d4 ./samples/pilot/previews/PanTS_00000003_S1_axial_ov_12.jpg
|
| 102 |
+
1924fa50b703622834a285bd9c98bf5d5a06e0879cd3b8d8af5837f25d1682c4 ./samples/pilot/previews/PanTS_00000003_S1_axial_ov_13.jpg
|
| 103 |
+
0ceed0d204dcdf4e4fb4d09b40910853d4ab2242625b3c85ad3d9efc6a4e5b34 ./samples/pilot/previews/PanTS_00000003_S1_axial_ov_15.jpg
|
| 104 |
+
4336b704ebd7e9a4382b20d162123b096485bcdbe646a736470f9afb437b1a30 ./samples/pilot/previews/PanTS_00000003_S1_axial_raw_0.jpg
|
| 105 |
+
9f4924bcf57db2183b8139714047f823891177d9e2854cbabcb28253e1ffd310 ./samples/pilot/previews/PanTS_00000003_S1_axial_raw_7.jpg
|
| 106 |
+
68a5f9eefb9efe1c31608f8b547541419c2a6db997c5388aaa3c834884c7029a ./samples/pilot/previews/PanTS_00000003_S2_axial_ov_12.jpg
|
| 107 |
+
2e172c858489d3e13a9d607548457ffd5a8d77470b2bd5e9f718a690350823dc ./samples/pilot/previews/PanTS_00000003_S2_axial_ov_13.jpg
|
| 108 |
+
f2870d676cb7e38fa8ecc66998746bdc0535e0ff664dd227adab515e1d7a2314 ./samples/pilot/previews/PanTS_00000003_S2_axial_ov_14.jpg
|
| 109 |
+
4336b704ebd7e9a4382b20d162123b096485bcdbe646a736470f9afb437b1a30 ./samples/pilot/previews/PanTS_00000003_S2_axial_raw_0.jpg
|
| 110 |
+
9c7ceea70bd9897715f0b7a60db2b921d6a82cfc9f446780f283e911c22e7189 ./samples/pilot/previews/PanTS_00000003_S2_axial_raw_3.jpg
|
| 111 |
+
86a05fb6c32f47dcce0b4f1fd5931c4460ceb0d901ca1e5734d71814403952e0 ./samples/pilot/previews/PanTS_00000003_S2_coronal_ov_189.jpg
|
| 112 |
+
cb9c1ae18f0eb221d844e21a0827b42fd2be47fc93557df1eb50bc4b21f009e9 ./samples/pilot/previews/PanTS_00000003_S2_coronal_raw_174.jpg
|
| 113 |
+
7de0062489a6e97028492d36b40593c5e0e7c4c3241fa3982cce1153fe205163 ./samples/pilot/previews/PanTS_00000003_S2_sagittal_ov_190.jpg
|
| 114 |
+
1dab770359879f808bebd98ef7de680319d3fa5d15b57f0abc1eb2ff3a32dd4e ./samples/pilot/previews/PanTS_00000003_S2_sagittal_raw_247.jpg
|
| 115 |
+
f2870d676cb7e38fa8ecc66998746bdc0535e0ff664dd227adab515e1d7a2314 ./samples/pilot/previews/PanTS_00000003_S3_axial_ov_14.jpg
|
| 116 |
+
daf4f1b6f26b8ed4f72ea8c5e600347048b5c49165c43f595e76d8869a95fb84 ./samples/pilot/previews/PanTS_00000003_S3_axial_ov_22.jpg
|
| 117 |
+
4336b704ebd7e9a4382b20d162123b096485bcdbe646a736470f9afb437b1a30 ./samples/pilot/previews/PanTS_00000003_S3_axial_raw_0.jpg
|
| 118 |
+
0604cc7cd845f08c330e0492912cb4ca242ecfd9d6ed6cc0209ef43f9e7f664b ./samples/pilot/previews/PanTS_00000003_S3_axial_raw_13.jpg
|
| 119 |
+
8222028eb29c734518c977ae174d6377e8ff95869b40ca83f57474d263edce41 ./samples/pilot/previews/PanTS_00000003_S3_mip_coronal_slab_0.jpg
|
| 120 |
+
0ad934cd0f7640edf8a18f2f3d8b47bc838eaf231c68efcc3cdc943ebe3a279a ./samples/pilot/previews/PanTS_00000003_S3_mip_sagittal_slab_0.jpg
|
| 121 |
+
4336b704ebd7e9a4382b20d162123b096485bcdbe646a736470f9afb437b1a30 ./samples/pilot/previews/PanTS_00000003_S4_axial_raw_0.jpg
|
| 122 |
+
02cbf7363540e95c16db6b0a9480d87bb23f31d3f760fdc3eff25357c78f7393 ./samples/pilot/previews/PanTS_00000003_S4_axial_raw_12.jpg
|
| 123 |
+
0604cc7cd845f08c330e0492912cb4ca242ecfd9d6ed6cc0209ef43f9e7f664b ./samples/pilot/previews/PanTS_00000003_S4_axial_raw_13.jpg
|
| 124 |
+
2e81ac9c87deb5e3ed39f607f0e54586e98b69c9fbc2032a4d3af96e05bb046c ./samples/pilot/previews/PanTS_00000003_S4_axial_raw_14.jpg
|
| 125 |
+
9f4924bcf57db2183b8139714047f823891177d9e2854cbabcb28253e1ffd310 ./samples/pilot/previews/PanTS_00000003_S4_axial_raw_7.jpg
|
| 126 |
+
1d34a5b190f413505104aad848bf9b37fa1e4ce15d09f24492eda99ed8144b65 ./samples/pilot/previews/PanTS_00000003_orient_sanity.jpg
|
| 127 |
+
e8147937325e44707fb7f2178517f20736550b7a35554175213158c3469e3394 ./samples/pilot/previews/PanTS_00000004_S1_axial_ov_35.jpg
|
| 128 |
+
e577acfc37d30308924370f5c184c1687b197aa586ecaf930798058eef8c7dcc ./samples/pilot/previews/PanTS_00000004_S1_axial_ov_45.jpg
|
| 129 |
+
dd47954d7057f3b596426dcfe774ba7e3b6dda8a39c03aae33bc7c63f2e5e8f1 ./samples/pilot/previews/PanTS_00000004_S1_axial_ov_56.jpg
|
| 130 |
+
22531e4b5690b9e67ab4e16b13e2b4d2eabae398a3da4d07711fcca810f3fb7d ./samples/pilot/previews/PanTS_00000004_S1_axial_raw_0.jpg
|
| 131 |
+
f7a30bb163a7cd6b5360c808a30e851aec08685dc2d365d2fd30c3eec21e8204 ./samples/pilot/previews/PanTS_00000004_S1_axial_raw_44.jpg
|
| 132 |
+
959ab047eac95fbe7b3113ee5c684360bc388581c24abcff1968f72a84e651b5 ./samples/pilot/previews/PanTS_00000004_S2_axial_ov_35.jpg
|
| 133 |
+
5f0bcdd6820a9e0f614351e0f4ad22022d069d6c76658e4bc2a1464fd2b2680a ./samples/pilot/previews/PanTS_00000004_S2_axial_ov_40.jpg
|
| 134 |
+
2dd48d4288929607e340168914067db558b974806c62dce8c946782c59723bce ./samples/pilot/previews/PanTS_00000004_S2_axial_ov_46.jpg
|
| 135 |
+
22531e4b5690b9e67ab4e16b13e2b4d2eabae398a3da4d07711fcca810f3fb7d ./samples/pilot/previews/PanTS_00000004_S2_axial_raw_0.jpg
|
| 136 |
+
858cfe0f54b00fa40decd8ebcb9eff913ee27ed35195ea60dcb28cbc4736c11b ./samples/pilot/previews/PanTS_00000004_S2_axial_raw_20.jpg
|
| 137 |
+
a694b6945dba08f5f83c4d93bedeb9638060375c71ce1fdfc8d0a9b7745a80b2 ./samples/pilot/previews/PanTS_00000004_S2_coronal_ov_151.jpg
|
| 138 |
+
369e7796093761710c819f22aceb784303b1971a1ad5028402a8e22050c0ce7b ./samples/pilot/previews/PanTS_00000004_S2_coronal_raw_162.jpg
|
| 139 |
+
c2484b7a171f66577b94ea535f97cbc61e525b25d0d0a2f1fee60c95081fc7c3 ./samples/pilot/previews/PanTS_00000004_S2_sagittal_ov_188.jpg
|
| 140 |
+
5fe92511238d79c41de71b696e2c3f179bf70df1041398edf8371372644cb22a ./samples/pilot/previews/PanTS_00000004_S2_sagittal_raw_251.jpg
|
| 141 |
+
8c0ac3fb849e6538b34972108f86c48aee8feeb4757b2e23b5fa47952fecdef6 ./samples/pilot/previews/PanTS_00000004_S3_axial_ov_76.jpg
|
| 142 |
+
c82c7db7bea7530cf8f380b857b7efd7c61a122659256d24fd893b0368972f3b ./samples/pilot/previews/PanTS_00000004_S3_axial_ov_84.jpg
|
| 143 |
+
22531e4b5690b9e67ab4e16b13e2b4d2eabae398a3da4d07711fcca810f3fb7d ./samples/pilot/previews/PanTS_00000004_S3_axial_raw_0.jpg
|
| 144 |
+
64efeb251fc3649d0295e93c1d2a6fd4e03858496036584d8ade89ef97af58db ./samples/pilot/previews/PanTS_00000004_S3_axial_raw_74.jpg
|
| 145 |
+
e819ca9fd8515dadbc30cf60303d829d53d67c8ebcebd6e8090cc3e8d189c108 ./samples/pilot/previews/PanTS_00000004_S3_mip_coronal_slab_0.jpg
|
| 146 |
+
d1b0ef676340f746d97badc6bf6c16039f6261a19a7ab26e85aa156210191ece ./samples/pilot/previews/PanTS_00000004_S3_mip_sagittal_slab_0.jpg
|
| 147 |
+
22531e4b5690b9e67ab4e16b13e2b4d2eabae398a3da4d07711fcca810f3fb7d ./samples/pilot/previews/PanTS_00000004_S4_axial_raw_0.jpg
|
| 148 |
+
ca4963a990be4c760c23db5fb48b28e6244b963f17704a08e18d70fa3e29acf4 ./samples/pilot/previews/PanTS_00000004_S4_axial_raw_35.jpg
|
| 149 |
+
ce8ee783820132930be1240b6b0700e8cd22a5b0d14449bb8a8af69371aebcdd ./samples/pilot/previews/PanTS_00000004_S4_axial_raw_43.jpg
|
| 150 |
+
f7a30bb163a7cd6b5360c808a30e851aec08685dc2d365d2fd30c3eec21e8204 ./samples/pilot/previews/PanTS_00000004_S4_axial_raw_44.jpg
|
| 151 |
+
4176f9cedbba7f25f30557bcd5f979cb17f0b1cd6b212436fe63064cebdbb7eb ./samples/pilot/previews/PanTS_00000004_S4_axial_raw_51.jpg
|
| 152 |
+
11846de957ad4ef58b648d58ee0931501c20f4ea779b0e28fa88484a86a48320 ./samples/pilot/previews/PanTS_00000004_orient_sanity.jpg
|
| 153 |
+
391d853e87daa99242626af3f32ab4dd72f8f3b270be271a935ffd1b8a3f6c23 ./samples/pilot/previews/PanTS_00000005_S1_axial_ov_0.jpg
|
| 154 |
+
d99512caf44134d2b3140d22fc97a0e938c446d9d7a2cbbf0b9844c04e02c155 ./samples/pilot/previews/PanTS_00000005_S1_axial_ov_4.jpg
|
| 155 |
+
6973e9f30bb536ceec51cfb06ba3adf4373a16bdeb5a287c8173aaba2c887454 ./samples/pilot/previews/PanTS_00000005_S1_axial_ov_8.jpg
|
| 156 |
+
e15e0b9fa65d0f95d23a2f8ead313bf0aeef71865fce4b07397418045da03a9e ./samples/pilot/previews/PanTS_00000005_S1_axial_raw_0.jpg
|
| 157 |
+
b355577a8f39893379a37a4c8429a7e89c7a57c17bb52d7afef9753ccb188d79 ./samples/pilot/previews/PanTS_00000005_S1_axial_raw_29.jpg
|
| 158 |
+
ea224bb808b5af3e6bf2e23755cc8033c3f14f7661dbbf02fe3066227148ab91 ./samples/pilot/previews/PanTS_00000005_S2_axial_ov_0.jpg
|
| 159 |
+
060e01317413622ab235cece5cf27f2a8324b63eb9f719f6368a10e6531a71b4 ./samples/pilot/previews/PanTS_00000005_S2_axial_ov_2.jpg
|
| 160 |
+
a9102dc28c6c6d6112936529b6e7a8f541d36d0c047650bb26563d5de2439c7b ./samples/pilot/previews/PanTS_00000005_S2_axial_ov_4.jpg
|
| 161 |
+
e15e0b9fa65d0f95d23a2f8ead313bf0aeef71865fce4b07397418045da03a9e ./samples/pilot/previews/PanTS_00000005_S2_axial_raw_0.jpg
|
| 162 |
+
431097265d1a8a7860e5dd0f47cbcf6bcd6c285e5e15b342252c8855ba9af9fd ./samples/pilot/previews/PanTS_00000005_S2_axial_raw_13.jpg
|
| 163 |
+
535079c6bbf5c68c7217971cc7f15c1bf459dc498139e2ce69d89c27056b0981 ./samples/pilot/previews/PanTS_00000005_S2_coronal_ov_176.jpg
|
| 164 |
+
0ebac81e2cc8bd712259d93edc85e90f082a4aeadef178900344ea15984c5624 ./samples/pilot/previews/PanTS_00000005_S2_coronal_raw_215.jpg
|
| 165 |
+
e593e2ae605032e319096a709ae2027d247b66feb68a7cf0cf9f7afc303ba370 ./samples/pilot/previews/PanTS_00000005_S2_sagittal_ov_170.jpg
|
| 166 |
+
71aa251eb04d4710d0b6c3288414b6d3f36cacad62c7606dca7a8203b63936f5 ./samples/pilot/previews/PanTS_00000005_S2_sagittal_raw_256.jpg
|
| 167 |
+
d37ef2cff3b4325c737b2cdf2e082ad1f888b4dd7f290c231c9a4d98713e8e36 ./samples/pilot/previews/PanTS_00000005_S3_axial_ov_16.jpg
|
| 168 |
+
7a98ac16ad8356f1ba6da24857d6b5a9698476e898dc5a763ef62f9444defe21 ./samples/pilot/previews/PanTS_00000005_S3_axial_ov_8.jpg
|
| 169 |
+
e15e0b9fa65d0f95d23a2f8ead313bf0aeef71865fce4b07397418045da03a9e ./samples/pilot/previews/PanTS_00000005_S3_axial_raw_0.jpg
|
| 170 |
+
8668e129d10954ec14718a7cee10a84a6333eb07dc26561afc056fc325189264 ./samples/pilot/previews/PanTS_00000005_S3_axial_raw_48.jpg
|
| 171 |
+
1e6f14c076f453ad20ef3be1db25dc80699da39a66e365746a5d21906934cd2c ./samples/pilot/previews/PanTS_00000005_S3_mip_coronal_slab_0.jpg
|
| 172 |
+
5914efb248d4179fbd9fe762495a23331e8744c4e4683513f2865b3c8ce59b5e ./samples/pilot/previews/PanTS_00000005_S3_mip_sagittal_slab_0.jpg
|
| 173 |
+
e15e0b9fa65d0f95d23a2f8ead313bf0aeef71865fce4b07397418045da03a9e ./samples/pilot/previews/PanTS_00000005_S4_axial_raw_0.jpg
|
| 174 |
+
b355577a8f39893379a37a4c8429a7e89c7a57c17bb52d7afef9753ccb188d79 ./samples/pilot/previews/PanTS_00000005_S4_axial_raw_29.jpg
|
| 175 |
+
bc8e4c6c7cc811dfd3bce311164bd28c32d9e00ae740e832a67250ed7be9a953 ./samples/pilot/previews/PanTS_00000005_S4_axial_raw_3.jpg
|
| 176 |
+
fd47f9a15e8bace2b59364b77d3faf3ab892f2b89d49507f3ddb9e7be67f9c4c ./samples/pilot/previews/PanTS_00000005_S4_axial_raw_6.jpg
|
| 177 |
+
b617261711649e84346051d72a831a8390bd06d540f8abccdb5b9aa4b3d7ab93 ./samples/pilot/previews/PanTS_00000005_orient_sanity.jpg
|
| 178 |
+
56b1536859b94d904c5d5050a444a66ae215c1d708a019207171e42090848304 ./samples/pilot/previews/PanTS_00000006_S1_axial_ov_77.jpg
|
| 179 |
+
3154666cf9ea25485c61beb9f49c6758442da560baaabb401906e8cef88b604c ./samples/pilot/previews/PanTS_00000006_S1_axial_ov_79.jpg
|
| 180 |
+
a55cb72a34bdb05f83ea05605efceec3a8b4270e9fb67932b87c419606c6cb29 ./samples/pilot/previews/PanTS_00000006_S1_axial_ov_81.jpg
|
| 181 |
+
4d7d377d8bf3ae0c9d082c2431bac595812325e3a67e065e2c6c590ef8701732 ./samples/pilot/previews/PanTS_00000006_S1_axial_raw_0.jpg
|
| 182 |
+
c0c737adcd18616bbc587aa1d8eed3f15fdca3a258e6856ff42086fd8648f473 ./samples/pilot/previews/PanTS_00000006_S1_axial_raw_22.jpg
|
| 183 |
+
de6f8b920ab74603805f3dc122dcf5629b08024b6f17c53843d5fd95af2d3311 ./samples/pilot/previews/PanTS_00000006_S2_axial_ov_77.jpg
|
| 184 |
+
5a04b6d071511fa5729ea35c276ec0d15d9c3f3012d3d629a849abec50961380 ./samples/pilot/previews/PanTS_00000006_S2_axial_ov_78.jpg
|
| 185 |
+
e16bd3b01af63dcd7759204a8b76efc1a722cd636e90f20691fcf9a2615266d1 ./samples/pilot/previews/PanTS_00000006_S2_axial_ov_79.jpg
|
| 186 |
+
4d7d377d8bf3ae0c9d082c2431bac595812325e3a67e065e2c6c590ef8701732 ./samples/pilot/previews/PanTS_00000006_S2_axial_raw_0.jpg
|
| 187 |
+
97aa6db31238c2529c306c2b910f223f367151f04bc8216ab2f6ac9b091d5ace ./samples/pilot/previews/PanTS_00000006_S2_axial_raw_10.jpg
|
| 188 |
+
1ae1c8e5a9d628c58a3c9149636ce0ecf624eb090e7562edd042ea45faca8429 ./samples/pilot/previews/PanTS_00000006_S2_coronal_ov_212.jpg
|
| 189 |
+
58b54bc164943179042831bd37082e19bd4f879001db4ec0de77c85860345d1e ./samples/pilot/previews/PanTS_00000006_S2_coronal_raw_215.jpg
|
| 190 |
+
05951f76e6dfb76866ce7909f660e6963a3d6e8814e2493c46d70c8ca3247cbb ./samples/pilot/previews/PanTS_00000006_S2_sagittal_ov_191.jpg
|
| 191 |
+
c084d0e6b1b75b6b6acb8bd572f0aca148ca52fd150ca0c7f4a8a6145d055f25 ./samples/pilot/previews/PanTS_00000006_S2_sagittal_raw_256.jpg
|
| 192 |
+
e57a2ebd28822d477c2a0a821bfa7fb9481b7b6d16a515ab0151f6be846b66b4 ./samples/pilot/previews/PanTS_00000006_S3_axial_ov_82.jpg
|
| 193 |
+
62ed0148dda2c7ef87663b616541a3d821e01e75bd234285c99182c9393a8f68 ./samples/pilot/previews/PanTS_00000006_S3_axial_ov_90.jpg
|
| 194 |
+
4d7d377d8bf3ae0c9d082c2431bac595812325e3a67e065e2c6c590ef8701732 ./samples/pilot/previews/PanTS_00000006_S3_axial_raw_0.jpg
|
| 195 |
+
75554b41399c91ea8ecae92a140c60f83f7cd3583b16e3359cc6bb8e57e8a072 ./samples/pilot/previews/PanTS_00000006_S3_axial_raw_37.jpg
|
| 196 |
+
ae5998b360b355f7caa8ebaf7cb52632ba64686e92236e6da3aa1ef64df438c1 ./samples/pilot/previews/PanTS_00000006_S3_mip_coronal_slab_0.jpg
|
| 197 |
+
d33149b770aa6075788157819420fcf758b60e35be0e7e82b969bfb611ae5ac8 ./samples/pilot/previews/PanTS_00000006_S3_mip_sagittal_slab_0.jpg
|
| 198 |
+
4d7d377d8bf3ae0c9d082c2431bac595812325e3a67e065e2c6c590ef8701732 ./samples/pilot/previews/PanTS_00000006_S4_axial_raw_0.jpg
|
| 199 |
+
c0c737adcd18616bbc587aa1d8eed3f15fdca3a258e6856ff42086fd8648f473 ./samples/pilot/previews/PanTS_00000006_S4_axial_raw_22.jpg
|
| 200 |
+
66fb8fd0d782539022936edca9cae76634a14c7185224c2612677679c939683f ./samples/pilot/previews/PanTS_00000006_S4_axial_raw_77.jpg
|
| 201 |
+
41086a1af7dd9b9ff2047b8785ade782c6c79e565b6a5925309c706b68fee65d ./samples/pilot/previews/PanTS_00000006_S4_axial_raw_78.jpg
|
| 202 |
+
ef94c2d74771e0716e18e61b6edd876f5b73c5abc801fe4fe8ed2b24cc5cbf15 ./samples/pilot/previews/PanTS_00000006_S4_axial_raw_80.jpg
|
| 203 |
+
a2be6c23a1ee3ef73bd7e8c9d0228bbb7a9a45ea218797418036b3c83cb329f9 ./samples/pilot/previews/PanTS_00000006_orient_sanity.jpg
|
| 204 |
+
22fb47b02704d52696f545a1997616292ad4bfb575882e5d0beffd21e1cc5aa7 ./samples/pilot/previews/PanTS_00000026_S1_axial_ov_50.jpg
|
| 205 |
+
12f5417a58ba7fa2a4d380e1536a303137fb8d2bf3d04c995e43f1336f36889c ./samples/pilot/previews/PanTS_00000026_S1_axial_ov_52.jpg
|
| 206 |
+
58b4e3f86c1460b808a9ecd9e07a61382254b6c39b6e68750a3e23cf64b85aa4 ./samples/pilot/previews/PanTS_00000026_S1_axial_ov_54.jpg
|
| 207 |
+
a30e50d9ee80019607a28f6161c4cee47b9a3eb37f31dbbf527cbf543b27f3c2 ./samples/pilot/previews/PanTS_00000026_S1_axial_raw_0.jpg
|
| 208 |
+
704d5ff9571d7b8617a557c9cf8641ddb81283671fc289e3b840e263965a72f9 ./samples/pilot/previews/PanTS_00000026_S1_axial_raw_16.jpg
|
| 209 |
+
a398b413fa468b9b61841d789d1caa2846db2a5f4f88dde2b4c71699669f6504 ./samples/pilot/previews/PanTS_00000026_S2_axial_ov_50.jpg
|
| 210 |
+
56299c117761de7ff8664f7a729c657e55d18564abfd894378692e81e741c479 ./samples/pilot/previews/PanTS_00000026_S2_axial_ov_51.jpg
|
| 211 |
+
3dcafe20ecf9ad9870d6139d9f35e836e4da19301fb904ea8e29ecde3a96b8e6 ./samples/pilot/previews/PanTS_00000026_S2_axial_ov_52.jpg
|
| 212 |
+
a30e50d9ee80019607a28f6161c4cee47b9a3eb37f31dbbf527cbf543b27f3c2 ./samples/pilot/previews/PanTS_00000026_S2_axial_raw_0.jpg
|
| 213 |
+
ddea71bc93f8fd5e956557b7539fb8a6fa918cd01f8d57d06c94086628321f6f ./samples/pilot/previews/PanTS_00000026_S2_axial_raw_7.jpg
|
| 214 |
+
bc8b8008cf1513f6562ae9f057291d9f55701aee8b6924915986d912422953b8 ./samples/pilot/previews/PanTS_00000026_S2_coronal_ov_189.jpg
|
| 215 |
+
b678963b9871de527f8f7c76cde1d964415a3e96d7f238d419b05b05ed73b890 ./samples/pilot/previews/PanTS_00000026_S2_coronal_raw_181.jpg
|
| 216 |
+
ec446b3061fd1452f1f18f34592fa6c5d9cc8975092d969aa21b0278269d4a0e ./samples/pilot/previews/PanTS_00000026_S2_sagittal_ov_159.jpg
|
| 217 |
+
0da88649e7aed96a621b898b4ffded42efc33c4cf5809db08aace53d46504414 ./samples/pilot/previews/PanTS_00000026_S2_sagittal_raw_256.jpg
|
| 218 |
+
539720d21c03444fbcfc2dc58f8ab333a965a12c4dcf532d1e8e11ea011899f8 ./samples/pilot/previews/PanTS_00000026_S3_axial_ov_54.jpg
|
| 219 |
+
a41141836ddf24f6510ed63ce9f3f56d16861ca5f1c193cc51e36ea7e977c681 ./samples/pilot/previews/PanTS_00000026_S3_axial_ov_62.jpg
|
| 220 |
+
a30e50d9ee80019607a28f6161c4cee47b9a3eb37f31dbbf527cbf543b27f3c2 ./samples/pilot/previews/PanTS_00000026_S3_axial_raw_0.jpg
|
| 221 |
+
a8c94b8d409bf52ba66fa9803de9d5a4dbff5edcfeb79fa85672d3c1e1fae400 ./samples/pilot/previews/PanTS_00000026_S3_axial_raw_26.jpg
|
| 222 |
+
29429cb3a59ae7688e48ab807520a537f049ef41564b7f449c42a4394672cdbc ./samples/pilot/previews/PanTS_00000026_S3_mip_coronal_slab_0.jpg
|
| 223 |
+
779a1403454f7c82a56480c6a75647be00b821ce6529453d383094a5128eaa79 ./samples/pilot/previews/PanTS_00000026_S3_mip_sagittal_slab_0.jpg
|
| 224 |
+
a30e50d9ee80019607a28f6161c4cee47b9a3eb37f31dbbf527cbf543b27f3c2 ./samples/pilot/previews/PanTS_00000026_S4_axial_raw_0.jpg
|
| 225 |
+
704d5ff9571d7b8617a557c9cf8641ddb81283671fc289e3b840e263965a72f9 ./samples/pilot/previews/PanTS_00000026_S4_axial_raw_16.jpg
|
| 226 |
+
fbc14260d47a8eea83914e121e58d86dec76d7055690967b9ccfe34e0adca28c ./samples/pilot/previews/PanTS_00000026_S4_axial_raw_50.jpg
|
| 227 |
+
94e4d56480c29583c65d5881eeb473596c61b1ce2c3f98a6c8bb3cd529d725ae ./samples/pilot/previews/PanTS_00000026_S4_axial_raw_51.jpg
|
| 228 |
+
2c0237084278851e4df2a46525693e904b352557b8dc165e7093c66e2f57fd41 ./samples/pilot/previews/PanTS_00000026_S4_axial_raw_53.jpg
|
| 229 |
+
4255ce384a88d14bacfc10ff303e6d0fabcbee601cea4517b69c94b980977dab ./samples/pilot/previews/PanTS_00000026_orient_sanity.jpg
|
| 230 |
+
97be885035af046919d45357775ccb6cd37ab6d5974d7ca2cf8b21a534b188d0 ./samples/pilot/previews/PanTS_00000029_S1_axial_ov_38.jpg
|
| 231 |
+
4608bb685de41c393abc8bd06965d6cfd3622282b738794a33f62db13db8c590 ./samples/pilot/previews/PanTS_00000029_S1_axial_ov_41.jpg
|
| 232 |
+
95ba2eac885f631852fafdcb8f6479dc9ca37a225fca60dea6c3b45dc730abd1 ./samples/pilot/previews/PanTS_00000029_S1_axial_ov_45.jpg
|
| 233 |
+
cfbbb5893e0e9281efee795f4b509db3bd34dceb651942587dd5fdf9195fc27e ./samples/pilot/previews/PanTS_00000029_S1_axial_raw_0.jpg
|
| 234 |
+
d8aa605694fc30f116d7722732e9c9dbf0ad3736781c977e0389e8a60b878c6d ./samples/pilot/previews/PanTS_00000029_S1_axial_raw_21.jpg
|
| 235 |
+
077d7cb3b58065e6438e8c59bd123198f85806110d8b8873a94298a5106c8984 ./samples/pilot/previews/PanTS_00000029_S2_axial_ov_38.jpg
|
| 236 |
+
c7a2e099f4023010ded496243d9d977a9899e4342c9beb400d67a4ce86460130 ./samples/pilot/previews/PanTS_00000029_S2_axial_ov_40.jpg
|
| 237 |
+
9845307e55ac6dfc25be64eae107bbcdc2cb0e3d9f6d055f5dcb53d5211d8235 ./samples/pilot/previews/PanTS_00000029_S2_axial_ov_42.jpg
|
| 238 |
+
cfbbb5893e0e9281efee795f4b509db3bd34dceb651942587dd5fdf9195fc27e ./samples/pilot/previews/PanTS_00000029_S2_axial_raw_0.jpg
|
| 239 |
+
c3db032714be8ea9bf31ba4a911c0a7cbbc57b1cd17f7a81d8eff97c2ab497cc ./samples/pilot/previews/PanTS_00000029_S2_axial_raw_9.jpg
|
| 240 |
+
fc7a5cb325e526ddfc48727310b1499a326b2726345a1e56829848406358dff6 ./samples/pilot/previews/PanTS_00000029_S2_coronal_ov_175.jpg
|
| 241 |
+
14e7ad151190543846288ac741e4a17b2b265ba6423b7585e5820fe2a83e6295 ./samples/pilot/previews/PanTS_00000029_S2_coronal_raw_170.jpg
|
| 242 |
+
dc74d35c9250f56ba9ccf63a7a78335eabfab8b8c3c15d0581e1383484240ad9 ./samples/pilot/previews/PanTS_00000029_S2_sagittal_ov_168.jpg
|
| 243 |
+
4a04cb1accaa23b920b4c94d316a83ef20fb8f5d14637fd5c0d23d904075d9ea ./samples/pilot/previews/PanTS_00000029_S2_sagittal_raw_209.jpg
|
| 244 |
+
7635de8604df3e41252839680072f1f967e02d43d268c745c98564f922f04e2a ./samples/pilot/previews/PanTS_00000029_S3_axial_ov_48.jpg
|
| 245 |
+
e7371ae4229ff0a936ac32e234558b883e8a438e1a33673a38af2b90d0bdc954 ./samples/pilot/previews/PanTS_00000029_S3_axial_ov_56.jpg
|
| 246 |
+
cfbbb5893e0e9281efee795f4b509db3bd34dceb651942587dd5fdf9195fc27e ./samples/pilot/previews/PanTS_00000029_S3_axial_raw_0.jpg
|
| 247 |
+
ec4dba0bfff50da6371b073e69efbc745824fc323effa1f250b506fc87b4c625 ./samples/pilot/previews/PanTS_00000029_S3_axial_raw_36.jpg
|
| 248 |
+
f78bff4996916570ef88056f7886073dba4e52ca239cda0a3f7c9115549f5ba0 ./samples/pilot/previews/PanTS_00000029_S3_mip_coronal_slab_0.jpg
|
| 249 |
+
7256ef209bb8462997f024bdc5f008e1a5bb7de1c82ad2bbbcab9d026823987f ./samples/pilot/previews/PanTS_00000029_S3_mip_sagittal_slab_0.jpg
|
| 250 |
+
cfbbb5893e0e9281efee795f4b509db3bd34dceb651942587dd5fdf9195fc27e ./samples/pilot/previews/PanTS_00000029_S4_axial_raw_0.jpg
|
| 251 |
+
d8aa605694fc30f116d7722732e9c9dbf0ad3736781c977e0389e8a60b878c6d ./samples/pilot/previews/PanTS_00000029_S4_axial_raw_21.jpg
|
| 252 |
+
20bd83cef5c4e853ad68eb1eef9fa93cbe36ad52ac84cf70ea8a7f1f6ba6f8d5 ./samples/pilot/previews/PanTS_00000029_S4_axial_raw_38.jpg
|
| 253 |
+
bfd55c13b05f5e7a9fae8a3f097daec090303e83a5efc376043989e857a04344 ./samples/pilot/previews/PanTS_00000029_S4_axial_raw_41.jpg
|
| 254 |
+
30e478e2c4e030b7d7bdf990316cd40888a93cc40cb7a594e168591ff3424936 ./samples/pilot/previews/PanTS_00000029_S4_axial_raw_44.jpg
|
| 255 |
+
dd33b7d73f74924a52bff70c2fcaf63fdbc77b5486cc8ee85e878c2fe21c6216 ./samples/pilot/previews/PanTS_00000029_orient_sanity.jpg
|
| 256 |
+
ed4b2c6ad9e19b88536211b984b112e0cc6751f62d0bc2a68c280d0fe6730dc1 ./samples/pilot/previews/PanTS_00000031_S1_axial_ov_11.jpg
|
| 257 |
+
f7fdf6720fbf502397b91379780994e30861c03c39ac475b5d5d76bd87d40490 ./samples/pilot/previews/PanTS_00000031_S1_axial_ov_14.jpg
|
| 258 |
+
40c50b0a73dc631f380de68982b43ff89d402032f3a7a26ad2583853cb81343e ./samples/pilot/previews/PanTS_00000031_S1_axial_ov_8.jpg
|
| 259 |
+
5222bf7c54dcf7af5efc2351e348811f03cd75ccd595d7baa594788c82297d5a ./samples/pilot/previews/PanTS_00000031_S1_axial_raw_0.jpg
|
| 260 |
+
3913870d664f6405194ec5886d5eda88e86b6e449bce98327a3de080f437f7c8 ./samples/pilot/previews/PanTS_00000031_S1_axial_raw_10.jpg
|
| 261 |
+
5ef98fc66fad65ed9cf9b0910c9273f7f6c27e76c6e50e98abf860c567d020a7 ./samples/pilot/previews/PanTS_00000031_S2_axial_ov_11.jpg
|
| 262 |
+
4762d93062243db39a94a6c3ad28b29cada9fec14dcccd4c3ca5383527a1e020 ./samples/pilot/previews/PanTS_00000031_S2_axial_ov_8.jpg
|
| 263 |
+
ac5f394261259b4ec8126ce42c6608862636ea2d30f3f208b485e5b0a7dde6c0 ./samples/pilot/previews/PanTS_00000031_S2_axial_ov_9.jpg
|
| 264 |
+
5222bf7c54dcf7af5efc2351e348811f03cd75ccd595d7baa594788c82297d5a ./samples/pilot/previews/PanTS_00000031_S2_axial_raw_0.jpg
|
| 265 |
+
b0d0e9b5db1d3721181922a0d67cef4ce891ccfd37433d017a909f7ec16b3bbe ./samples/pilot/previews/PanTS_00000031_S2_axial_raw_4.jpg
|
| 266 |
+
d58a1b87af228b2384f7001ed42e35fec2fccfdf1a6c9c4c3b95ccef0305ac28 ./samples/pilot/previews/PanTS_00000031_S2_coronal_ov_184.jpg
|
| 267 |
+
7b843d4c3ba7d5d04889ffa8d2b9cb10b60a85ed7b7424e2c2b8f35438fe159d ./samples/pilot/previews/PanTS_00000031_S2_coronal_raw_181.jpg
|
| 268 |
+
8027b053ffeaeedd21db5bee8df77c177a66706bf99914bf983ad44b02876994 ./samples/pilot/previews/PanTS_00000031_S2_sagittal_ov_145.jpg
|
| 269 |
+
891f3e8b141eeb21fe709552ce6831c0b07a6bb299bea8014129fa1511e46bd6 ./samples/pilot/previews/PanTS_00000031_S2_sagittal_raw_239.jpg
|
| 270 |
+
e8a00f38b6a473b9784ef72bed03e55f90c3c46ba21f8136673b3e1b7795ce26 ./samples/pilot/previews/PanTS_00000031_S3_axial_ov_15.jpg
|
| 271 |
+
d3e91ace230297cbbb9a2a8f7ba07a3c25108f8c1cf24e35ea0684a51f43f4dd ./samples/pilot/previews/PanTS_00000031_S3_axial_ov_23.jpg
|
| 272 |
+
5222bf7c54dcf7af5efc2351e348811f03cd75ccd595d7baa594788c82297d5a ./samples/pilot/previews/PanTS_00000031_S3_axial_raw_0.jpg
|
| 273 |
+
6eb37839c5f7d973ab08a13a753ec32e61583710e0a3de9f8dea4a5d4c91e855 ./samples/pilot/previews/PanTS_00000031_S3_axial_raw_16.jpg
|
| 274 |
+
72e7316eba6a60fa1072db15c350d2ea60eb9eaac1ba763cd96db79bf49aaa59 ./samples/pilot/previews/PanTS_00000031_S3_mip_coronal_slab_0.jpg
|
| 275 |
+
683a40534ba8b908269af63c47f5c85965ac101d056bf4cad96cddee53e0affd ./samples/pilot/previews/PanTS_00000031_S3_mip_sagittal_slab_0.jpg
|
| 276 |
+
5222bf7c54dcf7af5efc2351e348811f03cd75ccd595d7baa594788c82297d5a ./samples/pilot/previews/PanTS_00000031_S4_axial_raw_0.jpg
|
| 277 |
+
3913870d664f6405194ec5886d5eda88e86b6e449bce98327a3de080f437f7c8 ./samples/pilot/previews/PanTS_00000031_S4_axial_raw_10.jpg
|
| 278 |
+
6a8b8d9f354833c666c0346386830d1f0703c0b83d90e4c53b2a6d4f94d62bec ./samples/pilot/previews/PanTS_00000031_S4_axial_raw_12.jpg
|
| 279 |
+
5d6619d61261693632411238bd83a25f80de1988cc4a9dca2e6323eb41fbc643 ./samples/pilot/previews/PanTS_00000031_S4_axial_raw_8.jpg
|
| 280 |
+
b4d64058cff5ba45ff0ecb7289c3679b6916ff7bfbd7bb625b689acf517e2c85 ./samples/pilot/previews/PanTS_00000031_orient_sanity.jpg
|
| 281 |
+
f8d62cab51d13db62c1f077f625e9a0ca18e1b90658a8b42e53bae25fbf34b06 ./samples/pilot/previews/PanTS_00000035_S1_axial_ov_60.jpg
|
| 282 |
+
a88fee312e16a62de72ca9d3a03a2b634db28ac86df3b7da1123a4f71563aaf2 ./samples/pilot/previews/PanTS_00000035_S1_axial_ov_62.jpg
|
| 283 |
+
695af8587c0262d7dcabaadfcd03af0fcaffe6b71ca980716043c765163d4593 ./samples/pilot/previews/PanTS_00000035_S1_axial_ov_65.jpg
|
| 284 |
+
362d4c49977f71ad7614281b2d69d0c47fdf8cc49bf4cfa899d80b88714e5950 ./samples/pilot/previews/PanTS_00000035_S1_axial_raw_0.jpg
|
| 285 |
+
314e9219a37bab317eeeb28544ec0a145bc69a73e42c149c4faa403d38baea35 ./samples/pilot/previews/PanTS_00000035_S1_axial_raw_26.jpg
|
| 286 |
+
9dd1f435d5e27dbb208d26d2e605475003cdd7435250b293b7f349bbc36d4e38 ./samples/pilot/previews/PanTS_00000035_S2_axial_ov_60.jpg
|
| 287 |
+
cd4d77b48a7f73debf57a5eccc14d9aff68aa83fedd6bb647b28a248c9e1690e ./samples/pilot/previews/PanTS_00000035_S2_axial_ov_61.jpg
|
| 288 |
+
f0cc25109ab6ca306c8dbddbaadf3c2ed1950076392688a12b2ec2786f5f4b6a ./samples/pilot/previews/PanTS_00000035_S2_axial_ov_62.jpg
|
| 289 |
+
362d4c49977f71ad7614281b2d69d0c47fdf8cc49bf4cfa899d80b88714e5950 ./samples/pilot/previews/PanTS_00000035_S2_axial_raw_0.jpg
|
| 290 |
+
258db6859d45f57edbb9396995e4e82962a01bfc6b1303dab9ead3d62d15c137 ./samples/pilot/previews/PanTS_00000035_S2_axial_raw_12.jpg
|
| 291 |
+
1eddfd87d83ae04657d74d241b93889e107cc01f8a071ebf40cbfd818eaa5a43 ./samples/pilot/previews/PanTS_00000035_S2_coronal_ov_179.jpg
|
| 292 |
+
d3ce014de2d5dc9d5e43baba04f2162a3a4f496b7ddd950eb330101f3886e356 ./samples/pilot/previews/PanTS_00000035_S2_coronal_raw_195.jpg
|
| 293 |
+
5f6ea54539ca74c332801e90b81bcff98738abe1b46589f5713bc5556e2fe794 ./samples/pilot/previews/PanTS_00000035_S2_sagittal_ov_141.jpg
|
| 294 |
+
b21b3323bf398aec5b285830f3c14db1d4f5f48118590326e386a2ef44c17b78 ./samples/pilot/previews/PanTS_00000035_S2_sagittal_raw_256.jpg
|
| 295 |
+
5b755eb1790df16a0139c93c0d5be16744dd7bc09a009d19d8955ddf13ea3930 ./samples/pilot/previews/PanTS_00000035_S3_axial_ov_67.jpg
|
| 296 |
+
0b8ac5196083cd70ac5f3e698086faaf28873d4d5b7e4727f741e3189a5cd98b ./samples/pilot/previews/PanTS_00000035_S3_axial_ov_75.jpg
|
| 297 |
+
362d4c49977f71ad7614281b2d69d0c47fdf8cc49bf4cfa899d80b88714e5950 ./samples/pilot/previews/PanTS_00000035_S3_axial_raw_0.jpg
|
| 298 |
+
f0c0d912fe5823df191fca6d766af3bd4dd1e7965899adc665ba7be8e283b496 ./samples/pilot/previews/PanTS_00000035_S3_axial_raw_44.jpg
|
| 299 |
+
4e3b371737e289bbf3dac3072da3682aa67be584361a11935540aaf2d144f9fd ./samples/pilot/previews/PanTS_00000035_S3_mip_coronal_slab_0.jpg
|
| 300 |
+
bfa32a4b38e25a3ff1955493cfbc61d6a1003d17378fa675650f64306600329d ./samples/pilot/previews/PanTS_00000035_S3_mip_sagittal_slab_0.jpg
|
| 301 |
+
362d4c49977f71ad7614281b2d69d0c47fdf8cc49bf4cfa899d80b88714e5950 ./samples/pilot/previews/PanTS_00000035_S4_axial_raw_0.jpg
|
| 302 |
+
314e9219a37bab317eeeb28544ec0a145bc69a73e42c149c4faa403d38baea35 ./samples/pilot/previews/PanTS_00000035_S4_axial_raw_26.jpg
|
| 303 |
+
0f0e6ad2e2c5cff0777266ece11e89670b31c74c099ec81560f1831ebfd227bd ./samples/pilot/previews/PanTS_00000035_S4_axial_raw_60.jpg
|
| 304 |
+
61e4fd8b1568d4e4bebb0a8e76535427ebd140d30f65d0ddd9bfa9f302209714 ./samples/pilot/previews/PanTS_00000035_S4_axial_raw_62.jpg
|
| 305 |
+
ead0f1bb4045655e5683742ef359149f602770cadc288290c5e8c0912b2c13fb ./samples/pilot/previews/PanTS_00000035_S4_axial_raw_64.jpg
|
| 306 |
+
c694c88adffeba19b4a342ce9ca4095e0007402034aa34be34213e3500a8cc92 ./samples/pilot/previews/PanTS_00000035_orient_sanity.jpg
|
| 307 |
+
e044d090f47080d6342aa94c02cb2ebaa530e3598e8cec3b397d57438aa7d960 ./samples/pilot_v2/annotations.json
|
| 308 |
+
6cc184dd6b763bb590e49ba63bdb370834d0168e5a765f0fab55aa6ca544c0d6 ./samples/pilot_v2/ids.json
|
| 309 |
+
2dd2108db832c336c079b391e94b9fca45a965a8832407d1500acd221eb7bd7f ./samples/pilot_v2/ids_meta.json
|
| 310 |
+
13984c8ae3f3336ed94ec485532636d184b2c80b0498c3e4933f9187303aec3e ./samples/pilot_v2/previews/PanTS_00000001_axial_ov_36.jpg
|
| 311 |
+
c9cd7f5faabcc2480acbc685dba6ac01e998994bc4389d1b388cdbe8852d5738 ./samples/pilot_v2/previews/PanTS_00000001_axial_ov_45.jpg
|
| 312 |
+
0c49567beeadd899887d3529a6bdcb909e3a700103a11ee42a625368b3a33445 ./samples/pilot_v2/previews/PanTS_00000001_axial_ov_55.jpg
|
| 313 |
+
0c33a93876518d827f81648c40c8bdd96303daf7f38e88671681d2d84418b668 ./samples/pilot_v2/previews/PanTS_00000001_axial_raw_0.jpg
|
| 314 |
+
1d27d54e758a7ed2c3740159da73a4802330afb59de6afa15bbbaf561960e556 ./samples/pilot_v2/previews/PanTS_00000001_axial_raw_39.jpg
|
| 315 |
+
e7d06ce995f6e7f8ed574fe1576db53376fa62f15bfedb2e1649774e93a3e26c ./samples/pilot_v2/previews/PanTS_00000001_coronal_ov_197.jpg
|
| 316 |
+
f2280f1f1a3e00bff3d377653aab3479bd2fe955693e5b5fdc04b03f3106f643 ./samples/pilot_v2/previews/PanTS_00000001_sagittal_ov_224.jpg
|
| 317 |
+
d9b27cbd87c4852f2f3929617a72dcc08b5de5525627c19c2e1c48198afccca2 ./samples/pilot_v2/previews/PanTS_00000002_axial_ov_43.jpg
|
| 318 |
+
9ccfc8121d53873389ad44c993548bb8e7cb0a00d1733470bd9e89c177b23521 ./samples/pilot_v2/previews/PanTS_00000002_axial_ov_52.jpg
|
| 319 |
+
2fb5445ac80c570ffbaf687910e850cb9ed4ccfa49a11d54d0a3e0621f143c11 ./samples/pilot_v2/previews/PanTS_00000002_axial_ov_62.jpg
|
| 320 |
+
528667ceb887b4e6978a432d0822eacb46384d1442fdd2143bf8a1747be49745 ./samples/pilot_v2/previews/PanTS_00000002_axial_raw_0.jpg
|
| 321 |
+
48bbae938245965b5408d26853be391d7ef0ecd57958761a718d195317bd6b19 ./samples/pilot_v2/previews/PanTS_00000002_axial_raw_35.jpg
|
| 322 |
+
442454c97d0660b495c54182fbe945c0a3664a6613417dcfc8061f88c69ad724 ./samples/pilot_v2/previews/PanTS_00000002_coronal_ov_115.jpg
|
| 323 |
+
0fd8e38ea1af35d7419cecbecc1396b8abfec3d4708a2290b8059571c7c00817 ./samples/pilot_v2/previews/PanTS_00000002_sagittal_ov_126.jpg
|
| 324 |
+
d85bbfcb4bb97e921b95badb37b50235c46401c78cad183d2124c0aa7ced814d ./samples/pilot_v2/previews/PanTS_00000003_axial_ov_12.jpg
|
| 325 |
+
eedf02664299d5e36797566011f1c21e6049717dc95adcf8203f150835e71296 ./samples/pilot_v2/previews/PanTS_00000003_axial_ov_13.jpg
|
| 326 |
+
2e5722e85f074efff3ad27f4637542089c345e26158ccfa193367fdad7f2b310 ./samples/pilot_v2/previews/PanTS_00000003_axial_ov_15.jpg
|
| 327 |
+
4336b704ebd7e9a4382b20d162123b096485bcdbe646a736470f9afb437b1a30 ./samples/pilot_v2/previews/PanTS_00000003_axial_raw_0.jpg
|
| 328 |
+
9f4924bcf57db2183b8139714047f823891177d9e2854cbabcb28253e1ffd310 ./samples/pilot_v2/previews/PanTS_00000003_axial_raw_7.jpg
|
| 329 |
+
104a3cc7acdc6b660fcaa12127a8cbd4ae1ac39021425d41182bb4733047f26e ./samples/pilot_v2/previews/PanTS_00000003_coronal_ov_214.jpg
|
| 330 |
+
8b03881239c12fff9897cd4ef64898a23270d43d03710d06d03b57d83ad16fe8 ./samples/pilot_v2/previews/PanTS_00000003_sagittal_ov_239.jpg
|
| 331 |
+
c846a2e214e8e6d2c5decaa64a2fcf71359421a19a539116bda1ea91e065f37a ./samples/pilot_v2/previews/PanTS_00000004_axial_ov_35.jpg
|
| 332 |
+
bf9869bf9e026f7ed3ff99207dd8b4cc34b1797616b3aec38b1835ca281a48a1 ./samples/pilot_v2/previews/PanTS_00000004_axial_ov_45.jpg
|
| 333 |
+
5439e7f0b13f812a5e90f8d97fbcac9c776e9f452184de4c682a35862b4ead7a ./samples/pilot_v2/previews/PanTS_00000004_axial_ov_56.jpg
|
| 334 |
+
22531e4b5690b9e67ab4e16b13e2b4d2eabae398a3da4d07711fcca810f3fb7d ./samples/pilot_v2/previews/PanTS_00000004_axial_raw_0.jpg
|
| 335 |
+
f7a30bb163a7cd6b5360c808a30e851aec08685dc2d365d2fd30c3eec21e8204 ./samples/pilot_v2/previews/PanTS_00000004_axial_raw_44.jpg
|
| 336 |
+
fb4e3179b5301ec89d728e6a438a8510de45f53b38fd3b86a47b43f67a8ca6a7 ./samples/pilot_v2/previews/PanTS_00000004_coronal_ov_183.jpg
|
| 337 |
+
c6da03ca0e85a0fd7093668416843e5e57a294fd2f9927b0961484e5a84efd28 ./samples/pilot_v2/previews/PanTS_00000004_sagittal_ov_228.jpg
|
| 338 |
+
d8fee5e2ac2aad9629e14d38287241b2427032c7b6af3008022b0fdc2871d8bc ./samples/pilot_v2/previews/PanTS_00000005_axial_ov_0.jpg
|
| 339 |
+
34f8ea54f2aa707f7812bf9086f9de1c2153ea082bad988d6ea93784e8fe8e04 ./samples/pilot_v2/previews/PanTS_00000005_axial_ov_4.jpg
|
| 340 |
+
2a7a902dda89714d8413c21f5c77f99719c2a9653ccdcfedd1d1103d0955cc67 ./samples/pilot_v2/previews/PanTS_00000005_axial_ov_8.jpg
|
| 341 |
+
e15e0b9fa65d0f95d23a2f8ead313bf0aeef71865fce4b07397418045da03a9e ./samples/pilot_v2/previews/PanTS_00000005_axial_raw_0.jpg
|
| 342 |
+
b355577a8f39893379a37a4c8429a7e89c7a57c17bb52d7afef9753ccb188d79 ./samples/pilot_v2/previews/PanTS_00000005_axial_raw_29.jpg
|
| 343 |
+
c387a41034f0a3fcc877c6d6389ec0a509d678c64893f2f7de3cb43505bd10c2 ./samples/pilot_v2/previews/PanTS_00000005_coronal_ov_210.jpg
|
| 344 |
+
a2963879b55e13f73f8072a139f2bc3b6dc5fd3b75c1af6074176e1beb94e6b7 ./samples/pilot_v2/previews/PanTS_00000005_sagittal_ov_244.jpg
|
| 345 |
+
5e4ab21808f512a68c679bd77fe5c300dcf1041f5fd8c7c5d81a658bce1835d0 ./samples/pilot_v2/previews/PanTS_00000006_axial_ov_77.jpg
|
| 346 |
+
a5b865a49c93f66f9bbdac3b62395446de8ea53214e94170b27e72a7d500145d ./samples/pilot_v2/previews/PanTS_00000006_axial_ov_79.jpg
|
| 347 |
+
1dc2d2d88744484dabc5aaf07111d59b9d861dc801a9ccb832ee2a5909b0f636 ./samples/pilot_v2/previews/PanTS_00000006_axial_ov_81.jpg
|
| 348 |
+
4d7d377d8bf3ae0c9d082c2431bac595812325e3a67e065e2c6c590ef8701732 ./samples/pilot_v2/previews/PanTS_00000006_axial_raw_0.jpg
|
| 349 |
+
c0c737adcd18616bbc587aa1d8eed3f15fdca3a258e6856ff42086fd8648f473 ./samples/pilot_v2/previews/PanTS_00000006_axial_raw_22.jpg
|
| 350 |
+
573c8c47839c6fbf9ccd60a21d68a6a8e9961b48ea6abdc2347ea645b5586bbb ./samples/pilot_v2/previews/PanTS_00000006_coronal_ov_250.jpg
|
| 351 |
+
40118d5fad6deb7bb507c253cf6bde1ed2bd75fbb838ea20546d8afde56759b3 ./samples/pilot_v2/previews/PanTS_00000006_sagittal_ov_248.jpg
|
| 352 |
+
33d9f990f60c55fc6509538c1024d35d362a92e4ec6b2505dd79ea65afd3c774 ./samples/pilot_v2/previews/PanTS_00000026_axial_ov_50.jpg
|
| 353 |
+
707c25c4b433e983ab68d4b659bdc35517de47dfa20458636f0143555024fe73 ./samples/pilot_v2/previews/PanTS_00000026_axial_ov_52.jpg
|
| 354 |
+
59eec2262a54bea95e33c390c2b1333800144fc95cf327a207a5ad3b0bd5d89a ./samples/pilot_v2/previews/PanTS_00000026_axial_ov_54.jpg
|
| 355 |
+
a30e50d9ee80019607a28f6161c4cee47b9a3eb37f31dbbf527cbf543b27f3c2 ./samples/pilot_v2/previews/PanTS_00000026_axial_raw_0.jpg
|
| 356 |
+
704d5ff9571d7b8617a557c9cf8641ddb81283671fc289e3b840e263965a72f9 ./samples/pilot_v2/previews/PanTS_00000026_axial_raw_16.jpg
|
| 357 |
+
2efd101e4720a4288b81e6205af866fd970691b7e3dee74caf73e3edd8e076d4 ./samples/pilot_v2/previews/PanTS_00000026_coronal_ov_222.jpg
|
| 358 |
+
a9e189b896029e77879c189e5631e1aeedcb0a589273090ff4a505da426896c8 ./samples/pilot_v2/previews/PanTS_00000026_sagittal_ov_224.jpg
|
| 359 |
+
4f814277c9eea2e4da440fc5ea0d40bdd5ef1d6b406c1c4130ff8e303913f3e1 ./samples/pilot_v2/previews/PanTS_00000029_axial_ov_38.jpg
|
| 360 |
+
4be7848be41c8a84a3c2ce0c7c126c00ff773017de97ee4cbf688e10f90f1ec5 ./samples/pilot_v2/previews/PanTS_00000029_axial_ov_41.jpg
|
| 361 |
+
1a3f87ea3b35460c38f7406d84ce5e5a640d0ad74f50d3c74a1f90ce9e22442b ./samples/pilot_v2/previews/PanTS_00000029_axial_ov_45.jpg
|
| 362 |
+
cfbbb5893e0e9281efee795f4b509db3bd34dceb651942587dd5fdf9195fc27e ./samples/pilot_v2/previews/PanTS_00000029_axial_raw_0.jpg
|
| 363 |
+
d8aa605694fc30f116d7722732e9c9dbf0ad3736781c977e0389e8a60b878c6d ./samples/pilot_v2/previews/PanTS_00000029_axial_raw_21.jpg
|
| 364 |
+
059789de036c753b2e545ad5d2cdfe689daf13f3f3d18d0b148507265de9a422 ./samples/pilot_v2/previews/PanTS_00000029_coronal_ov_193.jpg
|
| 365 |
+
160d0877206f1947c5493ede934c5ff388a96c4a9000d3120325f9de551c2a16 ./samples/pilot_v2/previews/PanTS_00000029_sagittal_ov_221.jpg
|
| 366 |
+
daaf7c6b3dbe132671ec27eabf611f893efb785e049dd3a3655a9ea743e174cf ./samples/pilot_v2/previews/PanTS_00000031_axial_ov_11.jpg
|
| 367 |
+
54d081ed6d8ab52103c98c2b4d5667c10303805d2b6dfc8e581ccb9fe60fd08c ./samples/pilot_v2/previews/PanTS_00000031_axial_ov_14.jpg
|
| 368 |
+
cd964bf221c9aff2f2e491b6413392b8d772256d1ee6f97e68646df50b7ff55e ./samples/pilot_v2/previews/PanTS_00000031_axial_ov_8.jpg
|
| 369 |
+
5222bf7c54dcf7af5efc2351e348811f03cd75ccd595d7baa594788c82297d5a ./samples/pilot_v2/previews/PanTS_00000031_axial_raw_0.jpg
|
| 370 |
+
3913870d664f6405194ec5886d5eda88e86b6e449bce98327a3de080f437f7c8 ./samples/pilot_v2/previews/PanTS_00000031_axial_raw_10.jpg
|
| 371 |
+
9ae45469ea230716a664629732945cab36e3df8326177f25fad086a87c3a0647 ./samples/pilot_v2/previews/PanTS_00000031_coronal_ov_216.jpg
|
| 372 |
+
a4d246d5b2f80d3118aacf75ddae8481845563a0457db9910500c52290f0b183 ./samples/pilot_v2/previews/PanTS_00000031_sagittal_ov_213.jpg
|
| 373 |
+
dbc63e64506b380bcf5d85007380cebf78e859f92ff413f8d52965deff0abe13 ./samples/pilot_v2/previews/PanTS_00000035_axial_ov_60.jpg
|
| 374 |
+
e4f815924b60758497bc48aae9f7219589821f4f04f991633f2f8c08399725fc ./samples/pilot_v2/previews/PanTS_00000035_axial_ov_62.jpg
|
| 375 |
+
5df99620edbe8e0137d7aca9b48768c8c8ae7efc56e42b058b53f97e5bb29ae7 ./samples/pilot_v2/previews/PanTS_00000035_axial_ov_65.jpg
|
| 376 |
+
362d4c49977f71ad7614281b2d69d0c47fdf8cc49bf4cfa899d80b88714e5950 ./samples/pilot_v2/previews/PanTS_00000035_axial_raw_0.jpg
|
| 377 |
+
314e9219a37bab317eeeb28544ec0a145bc69a73e42c149c4faa403d38baea35 ./samples/pilot_v2/previews/PanTS_00000035_axial_raw_26.jpg
|
| 378 |
+
ad75f6e6c2825f05f686d8c4e81250d2473b2b4e3797f1e9bfae4bb906167721 ./samples/pilot_v2/previews/PanTS_00000035_coronal_ov_223.jpg
|
| 379 |
+
7c16b038c46f8d547c4fdbb314716c6fe46eb26da6b63019f18daf88acf3a546 ./samples/pilot_v2/previews/PanTS_00000035_sagittal_ov_212.jpg
|
| 380 |
+
c1e721797c419f184a192e54a7f32489d4a7b142e456241903162a840d23b8bc ./samples/pilot_v2/previews/PanTS_00000465_axial_ov_285.jpg
|
| 381 |
+
d31d86de25436443f64e56f7e26661571019ae459c8d2c221a95f7cf0fe53dae ./samples/pilot_v2/previews/PanTS_00000465_axial_ov_295.jpg
|
| 382 |
+
f54c87d44c246549341ce5bd908f8dabf8809a38de018a543adbe1a08c6f1629 ./samples/pilot_v2/previews/PanTS_00000465_axial_ov_306.jpg
|
| 383 |
+
5593735a26f6f8370380abe5a15fce06fb6b0eec4d9913bfaaed27f6ec0741da ./samples/pilot_v2/previews/PanTS_00000465_axial_raw_0.jpg
|
| 384 |
+
ae93d8dc915521fcca92db95a8ea2fa95c0835034f6f49a1c1ce9a2cf6cbaca7 ./samples/pilot_v2/previews/PanTS_00000465_axial_raw_130.jpg
|
| 385 |
+
24e7d29b586b92b157c7f34c424b6262704837fe1d23902e4f0762962a9cb74b ./samples/pilot_v2/previews/PanTS_00000465_coronal_ov_214.jpg
|
| 386 |
+
fab3838c6c70aa808aeb7ba089cb9a3bf445753e373df72175bb15bd1d707797 ./samples/pilot_v2/previews/PanTS_00000465_sagittal_ov_68.jpg
|
| 387 |
+
e434e73fd756b65226595e200f909ca2c94c06cdeb4793ca0431456346f4666b ./samples/pilot_v2/previews/PanTS_00000517_axial_ov_0.jpg
|
| 388 |
+
23fd6a2953b05f8f854ba9a707ca512185b68ba92f5cee2762732d17bc2ebe7f ./samples/pilot_v2/previews/PanTS_00000517_axial_ov_4.jpg
|
| 389 |
+
3e2b7312b5364c979601e10865d3168f6c5bb21dd6b71650761355608d9e05b1 ./samples/pilot_v2/previews/PanTS_00000517_axial_ov_8.jpg
|
| 390 |
+
f541a35b1afc4d140d2e0532880af50cd39776c721f8d7fd51b2df02110254ae ./samples/pilot_v2/previews/PanTS_00000517_axial_raw_0.jpg
|
| 391 |
+
3e5142de6d7087f1323412059f137fcb7f69f311c01c05940865fb08d0e8a857 ./samples/pilot_v2/previews/PanTS_00000517_axial_raw_17.jpg
|
| 392 |
+
9290930a1a886a3ffbfbd08ef75baa048ac700c3974eae1499b7aca9cc132880 ./samples/pilot_v2/previews/PanTS_00000517_coronal_ov_165.jpg
|
| 393 |
+
325c5a59cfb40922eb96cb5655c051704083f682ba6aab25355dfffefcc3ffdd ./samples/pilot_v2/previews/PanTS_00000517_sagittal_ov_195.jpg
|
| 394 |
+
fe3a6591fbc54367c1bcb2233ff6343d460a41b9fcc2ec5bbefc3b472521d927 ./samples/pilot_v2/previews/PanTS_00000725_axial_ov_0.jpg
|
| 395 |
+
5559a4ed145d3749e54d2b38fc21de6252bb148e40a603c141528b8e7ad4675e ./samples/pilot_v2/previews/PanTS_00000725_axial_ov_1.jpg
|
| 396 |
+
b91fc88f53605a4e01ad1ef5d3b17d6b2b149f3cc53a2188605968c94fd0b9eb ./samples/pilot_v2/previews/PanTS_00000725_axial_ov_3.jpg
|
| 397 |
+
d717a0a80d9e4ec4692a428039ca61d42c007d6c31af94360caea96f235478c0 ./samples/pilot_v2/previews/PanTS_00000725_axial_raw_0.jpg
|
| 398 |
+
7e9a3c63a6d466fb8c9795b2096135cbea096f8a53f36f6c0f70736d5bd5ed85 ./samples/pilot_v2/previews/PanTS_00000725_axial_raw_37.jpg
|
| 399 |
+
127c847693e5201d966119ce4df00b446211d51d6328d049526cdcccff5cb239 ./samples/pilot_v2/previews/PanTS_00000725_coronal_ov_111.jpg
|
| 400 |
+
5517666aff813d7ef22bc3c6382a8100f70dc1100c2a2a78ae167f5d6fd21dc4 ./samples/pilot_v2/previews/PanTS_00000725_sagittal_ov_119.jpg
|
| 401 |
+
7195387b1e54013176c1487a3fba580cef878f7e673591a9579373e4878f3402 ./samples/pilot_v2/previews/PanTS_00000814_axial_ov_0.jpg
|
| 402 |
+
83d162470fcad143da249676411293754cd9a32776f4e492212f67048ae2e767 ./samples/pilot_v2/previews/PanTS_00000814_axial_ov_16.jpg
|
| 403 |
+
936a96f030b504c1938dacec57cbdc7697dca8d0eb5638f6747c79872299fadf ./samples/pilot_v2/previews/PanTS_00000814_axial_ov_8.jpg
|
| 404 |
+
20937d43ffda293c8c88da0a171d2d2a60ffb5b9991572572308e648e0c68099 ./samples/pilot_v2/previews/PanTS_00000814_axial_raw_0.jpg
|
| 405 |
+
8895f2fbc0bd8831a3cbaa96ffa427805d38844a0a7150369ed25aebccdc2609 ./samples/pilot_v2/previews/PanTS_00000814_axial_raw_23.jpg
|
| 406 |
+
42bfa8f94476df4ca17b89847c9a5d9a304d0867b7fc9b0ce373db1b74721e53 ./samples/pilot_v2/previews/PanTS_00000814_coronal_ov_204.jpg
|
| 407 |
+
4f6689d4cb631c8bba028dfdaabec29c27870bf3430747ba449aa91e43a988a8 ./samples/pilot_v2/previews/PanTS_00000814_sagittal_ov_193.jpg
|
| 408 |
+
fb35e4ec87e0dbdd0baafd5edf18aebd5f1521035844184bb1b4687c0ef0ef62 ./samples/pilot_v2/previews/PanTS_00001030_axial_ov_85.jpg
|
| 409 |
+
77dbe3fd56e8d3350da4cfe5221eb47e3ad9552f6b8514176a713d0c93819fa2 ./samples/pilot_v2/previews/PanTS_00001030_axial_ov_90.jpg
|
| 410 |
+
2a19eb1f77af29b44b6a3dc2614513baf964b6b8d72eaefc8e6046aa8867bc58 ./samples/pilot_v2/previews/PanTS_00001030_axial_ov_95.jpg
|
| 411 |
+
7f052ccf803046117a1971736f00481ab86895d8f23cd54b6842e2e8d0456cc5 ./samples/pilot_v2/previews/PanTS_00001030_axial_raw_0.jpg
|
| 412 |
+
722b6e3f180f3e030807a23cdab8409da5174b1afa7717e0fc2058dabb9b9059 ./samples/pilot_v2/previews/PanTS_00001030_axial_raw_31.jpg
|
| 413 |
+
0ca311cc8bdea122368fe38d8cfba94cf58b3bb244504662da8dad39243fc1fc ./samples/pilot_v2/previews/PanTS_00001030_coronal_ov_240.jpg
|
| 414 |
+
bc917d35ca2f33b78f17ee106e5a41630047f82dd2c887923dcb8e70b7d4337f ./samples/pilot_v2/previews/PanTS_00001030_sagittal_ov_224.jpg
|
| 415 |
+
13c3002ea350ab42325c84758acfa22dee42b523eac6398bc566f0ec894df03e ./samples/pilot_v2/previews/PanTS_00001237_axial_ov_37.jpg
|
| 416 |
+
06c1c7988d9740dbef79605e123b5811d89a8403482ddcb4f36061fab0836e7b ./samples/pilot_v2/previews/PanTS_00001237_axial_ov_55.jpg
|
| 417 |
+
162f576c9f400559f98b52bf74c1452ec7345c1c4c8d5229866f4f80f7b2a98c ./samples/pilot_v2/previews/PanTS_00001237_axial_ov_73.jpg
|
| 418 |
+
9290b5a1044df7b3e268943b5d6ec090b17c8958de212d4fbb187e371c921ced ./samples/pilot_v2/previews/PanTS_00001237_axial_raw_0.jpg
|
| 419 |
+
53e08b6b4bee93e4234e2ef89a50973bfc798db91717076ba6edf8881a9ae929 ./samples/pilot_v2/previews/PanTS_00001237_axial_raw_49.jpg
|
| 420 |
+
f68ce0c22bbc9769e640c3d2330b7d9a6e7dcfb5f23977262ab39f44403641ae ./samples/pilot_v2/previews/PanTS_00001237_coronal_ov_200.jpg
|
| 421 |
+
1d493eb46f80c1e914150055ce335658aed97376fa4dff88474587c257b7ef7f ./samples/pilot_v2/previews/PanTS_00001237_sagittal_ov_226.jpg
|
| 422 |
+
37aca480803178934e92289000b4068176c0672c3925e45b09f37e963cb6c71c ./samples/pilot_v2/previews/PanTS_00001438_axial_ov_10.jpg
|
| 423 |
+
700d366c74b28fccd30064428ddb578d726c8087b7654db1049a814c4f41e2b5 ./samples/pilot_v2/previews/PanTS_00001438_axial_ov_14.jpg
|
| 424 |
+
92fa8f074fca4fea71157561f6b442d1fb7afd1477d69e470e873c8ee9b2cb8a ./samples/pilot_v2/previews/PanTS_00001438_axial_ov_18.jpg
|
| 425 |
+
66e35c1a6842bf0852df2241feb2a138dff096d160652dfdb336649b229a5b80 ./samples/pilot_v2/previews/PanTS_00001438_axial_raw_0.jpg
|
| 426 |
+
9354434dcfdc5005346bf28cafd9edcd6e46e3333902db76a3b69d1d57db9fbf ./samples/pilot_v2/previews/PanTS_00001438_axial_raw_15.jpg
|
| 427 |
+
86fb2601c725f36d486913bdc027f2d51736e7252bb8e059d9da3721febad3c1 ./samples/pilot_v2/previews/PanTS_00001438_coronal_ov_207.jpg
|
| 428 |
+
f46b95cae2857a19f3e8e864b2b2fd5ebb41674b6ddd2f487340a5ea4ef820ac ./samples/pilot_v2/previews/PanTS_00001438_sagittal_ov_177.jpg
|
| 429 |
+
c54b4bc8f58e0b180eaae5e23c8d0ae1c1df6a615031e57391b1ecc738c54489 ./samples/pilot_v2/previews/PanTS_00001878_axial_ov_58.jpg
|
| 430 |
+
073ffb21a3127f5c09e59e70160dc2fdb24a1356fade3183e395a4903eede14b ./samples/pilot_v2/previews/PanTS_00001878_axial_ov_61.jpg
|
| 431 |
+
3fd8d0e4c38f1ece78a76f75b8484c614a40d897d659b571f9ec126f3f663478 ./samples/pilot_v2/previews/PanTS_00001878_axial_ov_64.jpg
|
| 432 |
+
87e1acc1429c913db2e422869ddec5f81f91d3febd78ee0f57bd7c3e6a468789 ./samples/pilot_v2/previews/PanTS_00001878_axial_raw_0.jpg
|
| 433 |
+
5af6aac8e504c1dafeb0f5dfad543fd62f7866df5aee1f0c94fef733d9a311ce ./samples/pilot_v2/previews/PanTS_00001878_axial_raw_19.jpg
|
| 434 |
+
22636914c96ae3f240aab08ab6667fa34f6f4f2edd593a3b6c45f699299d9170 ./samples/pilot_v2/previews/PanTS_00001878_coronal_ov_175.jpg
|
| 435 |
+
211a4ef07b5e6fae5f6a9048db384644eae0385461644c50cbdbe927862c78d3 ./samples/pilot_v2/previews/PanTS_00001878_sagittal_ov_212.jpg
|
| 436 |
+
fc23558371bb5426b4466f09c8e4dd91b80d829506c4cf0236bcd9c7a20e0ff9 ./samples/pilot_v2/previews/PanTS_00001913_axial_ov_103.jpg
|
| 437 |
+
4811222bd64a176a3a7eb90cf495402a4a6143245d89f9182edd2078055951f2 ./samples/pilot_v2/previews/PanTS_00001913_axial_ov_107.jpg
|
| 438 |
+
f00ab5a6ff1e8cdd70be546a37a61245dca634482535621bee366d0809cda5b5 ./samples/pilot_v2/previews/PanTS_00001913_axial_ov_99.jpg
|
| 439 |
+
520b29e77748fb09b75a6321646a11362d15c15139060cb5430af5c3cd9d773c ./samples/pilot_v2/previews/PanTS_00001913_axial_raw_0.jpg
|
| 440 |
+
4a8641a705c09b9a6e044cb0b562b6b7bbe659e34c013ad75caba42e17750f01 ./samples/pilot_v2/previews/PanTS_00001913_axial_raw_32.jpg
|
| 441 |
+
9e3573cc534cab81beaaab44f77bcb07a55912a6df94318cb41f7615e9740938 ./samples/pilot_v2/previews/PanTS_00001913_coronal_ov_201.jpg
|
| 442 |
+
16c22920ae7beb26da642f1987994b7dbf943e9dcd7102c5032c6db4993940a2 ./samples/pilot_v2/previews/PanTS_00001913_sagittal_ov_199.jpg
|
| 443 |
+
6c3aec1081d20711b937b1d3c9d8e1622ce93aef0ea268b3a174d883314a55f6 ./samples/pilot_v2/previews/PanTS_00002085_axial_ov_18.jpg
|
| 444 |
+
7d379d0f413d4aaa787a9a2a316031f2ff4b840ac53c9983b44c21cc32eb1259 ./samples/pilot_v2/previews/PanTS_00002085_axial_ov_33.jpg
|
| 445 |
+
06c663dd93389d7be9075e73edfe0aaee982d694c1d9de1b1ecfdfa51e92a561 ./samples/pilot_v2/previews/PanTS_00002085_axial_ov_49.jpg
|
| 446 |
+
33bed0790f86c43f259cb344bc4272433c2988f004042001dc7f24fb19524c57 ./samples/pilot_v2/previews/PanTS_00002085_axial_raw_0.jpg
|
| 447 |
+
3a24db62ce1b3d0a4e91dc7f9b894738757fc758fd69d8bf455618093fee3166 ./samples/pilot_v2/previews/PanTS_00002085_axial_raw_48.jpg
|
| 448 |
+
b5db252f06b5f4aab5ed2e076264d908ac902f2ee3d9dbb0a53783b605949205 ./samples/pilot_v2/previews/PanTS_00002085_coronal_ov_258.jpg
|
| 449 |
+
6b28ccc2b3501363219871979b77844398e0ef239323eabf0d37a8e7944c2eaa ./samples/pilot_v2/previews/PanTS_00002085_sagittal_ov_225.jpg
|
| 450 |
+
64e4e2e1513f28b97efd9223aa77c0dd71d42057014d54a857c8d42acc66fe6a ./samples/pilot_v2/previews/PanTS_00002097_axial_ov_21.jpg
|
| 451 |
+
b0c8ffa8cddc349ef45aebeb11e8461a1f7b62064b4f6a33dc99889f345223c9 ./samples/pilot_v2/previews/PanTS_00002097_axial_ov_24.jpg
|
| 452 |
+
2c47a3ebacdcbb317d8084f3f084face3b8a342383a6bc5129c1bc4b9d4ce701 ./samples/pilot_v2/previews/PanTS_00002097_axial_ov_27.jpg
|
| 453 |
+
d39d86f49426048cf25a7f5f06926c04236cc23dfc900b9b9be66bb700aac261 ./samples/pilot_v2/previews/PanTS_00002097_axial_raw_0.jpg
|
| 454 |
+
301b72238679d753b54b87d4521e84e7093d991797cf06ef8d9d887d9a0507a4 ./samples/pilot_v2/previews/PanTS_00002097_axial_raw_12.jpg
|
| 455 |
+
7cee38bfb5d3b5e1eb49b9c5df406ac8152a0023007f20b65862e8d3634b61de ./samples/pilot_v2/previews/PanTS_00002097_coronal_ov_188.jpg
|
| 456 |
+
185fd47882381dd65455536507572d0ff0cb524aa2f43959d6e3af380699940e ./samples/pilot_v2/previews/PanTS_00002097_sagittal_ov_210.jpg
|
| 457 |
+
a5dd42740429875842f6d9bd009968a6fe5a18faa7140528ee464c7503c9f60c ./samples/pilot_v2/previews/PanTS_00002224_axial_ov_46.jpg
|
| 458 |
+
1822daa7e592091b294d63f8792203df1dc22b05fd128e0a942c1e60beb7d359 ./samples/pilot_v2/previews/PanTS_00002224_axial_ov_53.jpg
|
| 459 |
+
9536f20e7e96311aa76ac8ba3a32307e7a927ff141d65044e3feedb6ba2f6e07 ./samples/pilot_v2/previews/PanTS_00002224_axial_ov_61.jpg
|
| 460 |
+
6ce09ca4cd0113a6c74da4b37479ecebb5479c9ab763e811ff0ef28272b8ead8 ./samples/pilot_v2/previews/PanTS_00002224_axial_raw_0.jpg
|
| 461 |
+
e2851e3e61469e84644326b66968c5fb87f8d9f06f796c8f52eb0553777fc9e1 ./samples/pilot_v2/previews/PanTS_00002224_axial_raw_32.jpg
|
| 462 |
+
761fcb51e0a2d2cb2eb5d1c40fabeb885ca55106c68166ed9cb058c25b3435f1 ./samples/pilot_v2/previews/PanTS_00002224_coronal_ov_118.jpg
|
| 463 |
+
2a29b9a3726591f0b601698451c5c594e890575e423785f5333986ff8f0bd02f ./samples/pilot_v2/previews/PanTS_00002224_sagittal_ov_115.jpg
|
| 464 |
+
242642f6ebec5602e4c1911fbc9c4848719fd219f3de989635a8242b267711e1 ./samples/pilot_v2/previews/PanTS_00002232_axial_ov_25.jpg
|
| 465 |
+
d9b679acdee47c205d9ac9de23686ffc57cd58248b065ca51c0660a773adcdc7 ./samples/pilot_v2/previews/PanTS_00002232_axial_ov_29.jpg
|
| 466 |
+
4c8433b5da1a7625131b18c020b9c386c91564e5811c937022a161aec9f2ac83 ./samples/pilot_v2/previews/PanTS_00002232_axial_ov_34.jpg
|
| 467 |
+
5d97f715502b2fb10bdb75740bf4e61971e371138c81f0f2be1f40e93e3923e3 ./samples/pilot_v2/previews/PanTS_00002232_axial_raw_0.jpg
|
| 468 |
+
564c1875880558cde60d7122d99c5c12297c335d9c3735d693611ced2d285973 ./samples/pilot_v2/previews/PanTS_00002232_axial_raw_18.jpg
|
| 469 |
+
3abf57550150498342a85c44355a099c2fb1d7ce0abd391eaca1110ba35258fe ./samples/pilot_v2/previews/PanTS_00002232_coronal_ov_187.jpg
|
| 470 |
+
c386394d20a2c94332232ae2fe3334b9a7a154c91a4ff4491b0672ea00af198d ./samples/pilot_v2/previews/PanTS_00002232_sagittal_ov_225.jpg
|
| 471 |
+
eb19057456a30c225ab1f80029742b557240874dea539daf2bfd99d6f33422dc ./samples/pilot_v2/previews/PanTS_00003127_axial_ov_18.jpg
|
| 472 |
+
d6b5a4c234957c6b3b690386e2be2b2b59617fb313cf0d2ad47353165b3e39fd ./samples/pilot_v2/previews/PanTS_00003127_axial_ov_20.jpg
|
| 473 |
+
588a5e3a71a391097f74f7f320c2c2ecf773b58b2794b230a584cf7739ece2be ./samples/pilot_v2/previews/PanTS_00003127_axial_ov_23.jpg
|
| 474 |
+
2246f5e34590deb69d46c82d8ac9be391df876958b61555e4d23304dc47b81a6 ./samples/pilot_v2/previews/PanTS_00003127_axial_raw_0.jpg
|
| 475 |
+
30e7db0ea575db99bd98903cf48e6d312d0288db5fe6083fbeaae49ee27c2ba0 ./samples/pilot_v2/previews/PanTS_00003127_axial_raw_10.jpg
|
| 476 |
+
dbe4c50039b2706089d3e1d21cf30c3e74d0a5eb677c246f6455172f37cde134 ./samples/pilot_v2/previews/PanTS_00003127_coronal_ov_191.jpg
|
| 477 |
+
d5c1390619bd181477e1ae864b816f43112b977c81158a00e534198d61e91d07 ./samples/pilot_v2/previews/PanTS_00003127_sagittal_ov_229.jpg
|
| 478 |
+
fab54f627cca8a42c43386db041e1e9223834df9281302fc78ca148ddd6c91db ./samples/pilot_v2/previews/PanTS_00003252_axial_ov_56.jpg
|
| 479 |
+
37b03145079d2e86a4f9f4b4530af3d002693b443ac322ffb8a6110c07b7599b ./samples/pilot_v2/previews/PanTS_00003252_axial_ov_62.jpg
|
| 480 |
+
a205d7ffb2317585bddc4698d6f1d51247bab4b76ae0b44dbd5393495c623937 ./samples/pilot_v2/previews/PanTS_00003252_axial_ov_69.jpg
|
| 481 |
+
29915f633cb6d0e9486bdd7b362c3702349c4b1766a9521003844793984a2afa ./samples/pilot_v2/previews/PanTS_00003252_axial_raw_0.jpg
|
| 482 |
+
9196ce1719d43f144b002a8dc579f83b4db8a9b20c5c057978cfb40381fffcc7 ./samples/pilot_v2/previews/PanTS_00003252_axial_raw_36.jpg
|
| 483 |
+
ba3a7e0881d197dff056c6357f62c64cafd47168608ea4cbe700962cd6a1f0b7 ./samples/pilot_v2/previews/PanTS_00003252_coronal_ov_87.jpg
|
| 484 |
+
d1ed57b9c7cad55a283480e84fb40907766a7c8644c37d97059002427c749a00 ./samples/pilot_v2/previews/PanTS_00003252_sagittal_ov_101.jpg
|
| 485 |
+
43a66909b4774b28c77cc4381db0d2365822298eb13b798362ca1e2e8e54c6b5 ./samples/pilot_v2/previews/PanTS_00003449_axial_ov_25.jpg
|
| 486 |
+
1aeb19fc894481870f739982768b5bb7fde354a44cadc9d3b2cf0df4a90583c6 ./samples/pilot_v2/previews/PanTS_00003449_axial_ov_26.jpg
|
| 487 |
+
ae06e69e44404a09d1082094eace6d7a07ccb4593e9df450728a02821da3ec1b ./samples/pilot_v2/previews/PanTS_00003449_axial_ov_27.jpg
|
| 488 |
+
124b985fb26ac2daf49d2ff7a1eff58b36a4b2eb656a1aca984741b9d33bd88a ./samples/pilot_v2/previews/PanTS_00003449_axial_raw_0.jpg
|
| 489 |
+
299a3c891814f79196d2bfe927b1caecdb815d96986461af958da401db46c79e ./samples/pilot_v2/previews/PanTS_00003449_axial_raw_19.jpg
|
| 490 |
+
7d02d253b67a5a6e0e09ffa9cc64122d625c1199d9b79274886f26c945c0e328 ./samples/pilot_v2/previews/PanTS_00003449_coronal_ov_84.jpg
|
| 491 |
+
8b6f0bb9c9eba01d94a8a6e4334d8f95de116b8ed12901beafdc4fca94af8821 ./samples/pilot_v2/previews/PanTS_00003449_sagittal_ov_88.jpg
|
| 492 |
+
61c4ae3145ea0b99ba757a7059bc64b16eccf7e16286291c3dd383d26be4be8b ./samples/pilot_v2/previews/PanTS_00003485_axial_ov_0.jpg
|
| 493 |
+
e78da7fcf1fa7a320bce4130990d5498f7fb73afb686a0a3304dbf9d84a63dc9 ./samples/pilot_v2/previews/PanTS_00003485_axial_ov_15.jpg
|
| 494 |
+
3e0887b176c4f5fdd100665bf47c6e5761e9b3545154757b8132fb7163168af4 ./samples/pilot_v2/previews/PanTS_00003485_axial_ov_30.jpg
|
| 495 |
+
701a8831994f00c7a44c34b80a3fab0836d5f5f67e82fdfb42dfe98c655d9ae8 ./samples/pilot_v2/previews/PanTS_00003485_axial_raw_0.jpg
|
| 496 |
+
45f7c9455883e526d8691cd366b13aff4a91d6c6018e597bbc3e0d4cc64655bf ./samples/pilot_v2/previews/PanTS_00003485_axial_raw_40.jpg
|
| 497 |
+
ae08a2325e4316d4996fb7cf3acdc3eaf503fb1c9820f937f4edda4a215278c2 ./samples/pilot_v2/previews/PanTS_00003485_coronal_ov_281.jpg
|
| 498 |
+
043a1e3337a452283a5fc73e123c646f3890a86fa794a4b4282906e44ee84681 ./samples/pilot_v2/previews/PanTS_00003485_sagittal_ov_246.jpg
|
| 499 |
+
53045663be5cb1247c2321cf647904a2fb7cff25ed61fd57611caba999f9c45a ./samples/pilot_v2/previews/PanTS_00003631_axial_ov_36.jpg
|
| 500 |
+
847f02c881a4d003785c7c39d0ba438ed6d2398c4cd06d1febda6e02b6213203 ./samples/pilot_v2/previews/PanTS_00003631_axial_ov_41.jpg
|
| 501 |
+
b569edd06e7a6781fcbaf405c3d4c1dcfb7622845e3800dd8a34b287e59df29f ./samples/pilot_v2/previews/PanTS_00003631_axial_ov_46.jpg
|
| 502 |
+
0ac9b83111d146f1c992692224050a056ceb16bb71154f3b630772f34b87c72f ./samples/pilot_v2/previews/PanTS_00003631_axial_raw_0.jpg
|
| 503 |
+
5dff00b366fc28e19f67d1ea90d1ccde7c8cdaf8454f8e6768366dbf4403690d ./samples/pilot_v2/previews/PanTS_00003631_axial_raw_22.jpg
|
| 504 |
+
e04b934e7086c70f02c038f7a83bcd9d014f0b1938e963c10ad3e46ac8344293 ./samples/pilot_v2/previews/PanTS_00003631_coronal_ov_169.jpg
|
| 505 |
+
5bead6648a2b4b5ba7334dd83164af72ee9167315a1bacb2eef2402c2d6f6445 ./samples/pilot_v2/previews/PanTS_00003631_sagittal_ov_225.jpg
|
| 506 |
+
354c4f1c215461ce75f9744a037c285c49627142a8b23bb9d26de9927523e2bf ./samples/pilot_v2/previews/PanTS_00003824_axial_ov_62.jpg
|
| 507 |
+
e3eaad6f1c901bfc40c6b53039c53d1d52b7dacf38385a2c9312ac4d358030b6 ./samples/pilot_v2/previews/PanTS_00003824_axial_ov_65.jpg
|
| 508 |
+
811fcb770a4ba02e5c8f4ec7c7cc61f7005da9916168424baf1e695a04cd8af4 ./samples/pilot_v2/previews/PanTS_00003824_axial_ov_68.jpg
|
| 509 |
+
109c92162681200b17b023168b340814e332c8e12ac2fbde148382531eab50a3 ./samples/pilot_v2/previews/PanTS_00003824_axial_raw_0.jpg
|
| 510 |
+
b9a908298b359267df9cf2e44ca16f09808f4177b80a11617e0163b6451e9a22 ./samples/pilot_v2/previews/PanTS_00003824_axial_raw_19.jpg
|
| 511 |
+
bfb083fd8fef6fd6554ba3656c50c69b6fa5aec763fd1b8c2018d68068d75e54 ./samples/pilot_v2/previews/PanTS_00003824_coronal_ov_176.jpg
|
| 512 |
+
f47580247bbc98bed5a27efdb01498502bac23bfe85386424dcd1fe8d05bca19 ./samples/pilot_v2/previews/PanTS_00003824_sagittal_ov_254.jpg
|
| 513 |
+
06d00b25c8660d67580aaf1dbc6dfe0ac5319807826be14170b167e6104de774 ./samples/pilot_v2/previews/PanTS_00005646_axial_ov_147.jpg
|
| 514 |
+
847ed025dbaa032cc5e98897e94c09bcb6b4f402326cb0bd2741bd6e065a9a21 ./samples/pilot_v2/previews/PanTS_00005646_axial_ov_154.jpg
|
| 515 |
+
0c8b3dec24003fec00a3029912a8067892bfc07de2c1e92bdeff819c45f3662d ./samples/pilot_v2/previews/PanTS_00005646_axial_ov_162.jpg
|
| 516 |
+
160e75ec50161d9185072708a38b9da5921d7b4677f6bc8ff86cc368c65a8bc2 ./samples/pilot_v2/previews/PanTS_00005646_axial_raw_0.jpg
|
| 517 |
+
a0a16a98ba489b537e7225d7bbf44cc18a8bc48059ad2476d4a6ec2a93662f30 ./samples/pilot_v2/previews/PanTS_00005646_axial_raw_42.jpg
|
| 518 |
+
af879de3192fb299da6ca6a60511d1120effb3f5f5ae71ce09767d88783b5f90 ./samples/pilot_v2/previews/PanTS_00005646_coronal_ov_100.jpg
|
| 519 |
+
c5c990af155afbef91aabf595b8ef44f9a72359a709de7bf7b61708cee3c3789 ./samples/pilot_v2/previews/PanTS_00005646_sagittal_ov_120.jpg
|
| 520 |
+
5f78f6f7b50b1753fe3b8cccc5444ffa55ca9da03a2fb0816bfd88d303ed1567 ./splits/bucket_spec.json
|
| 521 |
+
9068552fdba7f7a09236266f971f27c28bbbf2c663f799f5ca06b1f3c12bdb02 ./splits/splits.json
|
raw_pants_train_test/metadata/pants-captions-ldm/README.md
ADDED
|
@@ -0,0 +1,207 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
language: en
|
| 4 |
+
pretty_name: PanTS-Captions — 7-variant CT captions + LDM training pipeline
|
| 5 |
+
task_categories:
|
| 6 |
+
- text-to-image
|
| 7 |
+
- image-to-text
|
| 8 |
+
- other
|
| 9 |
+
tags:
|
| 10 |
+
- medical
|
| 11 |
+
- ct
|
| 12 |
+
- 3d
|
| 13 |
+
- pancreas
|
| 14 |
+
- latent-diffusion
|
| 15 |
+
- captions
|
| 16 |
+
- radiology
|
| 17 |
+
size_categories:
|
| 18 |
+
- 1K<n<10K
|
| 19 |
+
---
|
| 20 |
+
|
| 21 |
+
# PanTS-Captions
|
| 22 |
+
|
| 23 |
+
A companion dataset for the **PanTS** pancreatic CT corpus: per-case structured
|
| 24 |
+
conditioning + **7 caption variants** (narrative, terse impression, organ bullet,
|
| 25 |
+
tag string, layered findings, Q/A, pancreas-only) over all 9,901 PanTS cases,
|
| 26 |
+
together with the full training-pipeline scaffolds for a 3D text-to-image
|
| 27 |
+
latent-diffusion model (LDM) on pancreatic CT.
|
| 28 |
+
|
| 29 |
+
**This release contains NO CT volumes, NO per-voxel masks, and NO patient-level
|
| 30 |
+
identifiers.** The CT + mask data itself is the original PanTS release and must
|
| 31 |
+
be obtained from its upstream distributor under its own license. What is
|
| 32 |
+
distributed here is:
|
| 33 |
+
|
| 34 |
+
1. A **captions corpus** (`captions/captions_final.jsonl`) keyed by PanTS case ID.
|
| 35 |
+
2. A **structured canonical-facts file** (`canonical/canonical_facts.jsonl`) —
|
| 36 |
+
organ volumes, HU statistics, FOV, phase, lesion list, mask-derived pancreas
|
| 37 |
+
bbox, demographics — parsed deterministically from the PanTS radiology reports
|
| 38 |
+
and mask files.
|
| 39 |
+
3. **Official dataset splits** (`splits/splits.json`) — 8,181 train / 819 val /
|
| 40 |
+
901 test, with the PanTS-te holdout preserved intact.
|
| 41 |
+
4. **Bucket spec** (`splits/bucket_spec.json`) — 6 FOV-bucketed training shapes
|
| 42 |
+
with DiT-3D token counts under an 8,192-token cap.
|
| 43 |
+
5. **Pilot + canary rating annotations** (`samples/`) — 40 pilot cases + 200
|
| 44 |
+
canary cases across multiple caption formats, the labeled evaluation samples
|
| 45 |
+
that drove the v7 caption policy.
|
| 46 |
+
6. **Phase QC** (`audit/phase_audit.jsonl`) — a 4-way cross-check between
|
| 47 |
+
LLM-parsed report phase, organ-HU heuristic, caption mention, and report text;
|
| 48 |
+
flags 4,967 cases where the HU heuristic disagrees with the canonical phase
|
| 49 |
+
label (the heuristic is a noisy secondary signal; canonical is the source of
|
| 50 |
+
truth).
|
| 51 |
+
7. **All code** (`code/`) — caption generation, merging, auditing, preprocessing,
|
| 52 |
+
noise schedule, augmentation policy, WebDataset packaging, VAE sanity gate,
|
| 53 |
+
dry-run training.
|
| 54 |
+
8. **Planning docs** (`docs/`) — proposal + 4 training-plan revisions + the
|
| 55 |
+
pre-training checklist summary.
|
| 56 |
+
|
| 57 |
+
## Quick stats
|
| 58 |
+
|
| 59 |
+
- N = 9,901 cases (9,000 PanTS-tr + 901 PanTS-te) × 5 always-present variants = 49,505 captions
|
| 60 |
+
- V6_qa (lesion+ only): 1,122 additional Q/A caption pairs
|
| 61 |
+
- V7 pancreas_only (mask-grounded, lesion+): 256 pancreas-only captions
|
| 62 |
+
- Final banned-word audit pass rate per variant: V1 96.3% / V2 97.8% / V3 93.6% / V4 94.0% / V5 96.5% / V6_qa 100% / V7 97.3%
|
| 63 |
+
- Phase split (from canonical): Non-contrast 4,485 / Venous 2,897 / Arterial 2,450 / Delay 68 / MISSING 1
|
| 64 |
+
|
| 65 |
+
## Directory layout
|
| 66 |
+
|
| 67 |
+
```
|
| 68 |
+
.
|
| 69 |
+
├── README.md # this file
|
| 70 |
+
├── captions/
|
| 71 |
+
│ ├── captions_final.jsonl # PRIMARY corpus — 9,901 rows, {id, cond, captions}
|
| 72 |
+
│ ├── captions_pilot.jsonl # 40 pilot cases, multiple caption variants
|
| 73 |
+
│ ├── captions_v1_initial.jsonl # VLM first pass (pre-fusion)
|
| 74 |
+
│ ├── captions_v2_after_fusion.jsonl # after canonical-facts fusion
|
| 75 |
+
│ ├── captions_v3_after_regen.jsonl # after truncation regen
|
| 76 |
+
│ └── captions_v7_pancreas_only.jsonl # V7 raw output (pre-V7_BAD_RE filter)
|
| 77 |
+
│
|
| 78 |
+
├── canonical/
|
| 79 |
+
│ └── canonical_facts.jsonl # 9,901 rows: report + mask + canonical dicts
|
| 80 |
+
│
|
| 81 |
+
├── splits/
|
| 82 |
+
│ ├── splits.json # train 8181 / val 819 / test 901
|
| 83 |
+
│ └── bucket_spec.json # 6 FOV buckets, DiT-3D token math
|
| 84 |
+
│
|
| 85 |
+
├── samples/
|
| 86 |
+
│ ├── pilot/ { annotations.json (10 cases × 4 volumetric-render options), ids.json, previews/ (258 jpg renders — axial / coronal / sagittal / MIP, raw vs organ-overlay) }
|
| 87 |
+
│ ├── pilot_v2/ { annotations.json (30 cases × 2 prompt styles), ids.json, ids_meta.json, previews/ (210 jpg renders) }
|
| 88 |
+
│ └── canary/ { annotations.json (200 cases: JSON-only vs prose), ids.json — no previews, this was a text-only study }
|
| 89 |
+
│
|
| 90 |
+
├── audit/
|
| 91 |
+
│ ├── audit_final.md # per-variant banned-word QC report
|
| 92 |
+
│ ├── audit_v2.md # intermediate audit report
|
| 93 |
+
│ ├── phase_audit.jsonl # 9,901 per-case 4-way phase check
|
| 94 |
+
│ └── phase_audit_summary.md # aggregate phase-source confusion
|
| 95 |
+
│
|
| 96 |
+
├── code/
|
| 97 |
+
│ ├── preprocessing.py # RAS+ / HU clip / resample / pan-crop (order-fix)
|
| 98 |
+
│ ├── schedule.py # VP linear β + ZTSNR + v-prediction + Min-SNR-γ=5
|
| 99 |
+
│ ├── aug_config.py # latent noise + crop jitter (LR/HU/spacing off)
|
| 100 |
+
│ ├── build_splits.py
|
| 101 |
+
│ ├── build_bucket_spec.py
|
| 102 |
+
│ ├── phase_audit.py
|
| 103 |
+
│ ├── vae_sanity_gate.py # scaffold — needs Wan 2.2 VAE weights
|
| 104 |
+
│ ├── dry_run_train.py # 5k-step dry run, stand-in DiT-3D
|
| 105 |
+
│ ├── pack_webdataset.py # HF WebDataset tar packing
|
| 106 |
+
│ ├── test_preprocessing.py # 8 unit tests, all pass
|
| 107 |
+
│ └── caption_generation/
|
| 108 |
+
│ ├── canonical_facts.py
|
| 109 |
+
│ ├── f5_audit.py # full-corpus QC
|
| 110 |
+
│ ├── f5d_regen.py # truncation regen
|
| 111 |
+
│ ├── f5b_v7.py # V7 pancreas-only fusion
|
| 112 |
+
│ ├── f6_merge.py # final merge across variants
|
| 113 |
+
│ ├── f6_audit_final.py # post-merge QC
|
| 114 |
+
│ ├── fusion_run.py / fusion_retry.py # canonical × VLM fusion
|
| 115 |
+
│ ├── full_vlm_run.py / full_vlm_run2.py # first + second VLM passes
|
| 116 |
+
│ ├── pilot_run.py / pilot_run_v2.py # pilot annotation studies
|
| 117 |
+
│ └── salvage_vlm.py
|
| 118 |
+
│
|
| 119 |
+
└── docs/
|
| 120 |
+
├── PROPOSAL_v1.md, PROPOSAL_v3_1.md, PROPOSAL_v3_2.md
|
| 121 |
+
├── LDM_TRAINING_PLAN_v4.md
|
| 122 |
+
├── PRETRAIN_CHECKLIST_SUMMARY.md # v4 §14 10-item status
|
| 123 |
+
├── fusion_variants_design.md # the 7-variant design doc
|
| 124 |
+
└── GPT_REVIEW.md # external review notes
|
| 125 |
+
```
|
| 126 |
+
|
| 127 |
+
## Schema: `captions/captions_final.jsonl`
|
| 128 |
+
|
| 129 |
+
One JSON object per line:
|
| 130 |
+
|
| 131 |
+
```json
|
| 132 |
+
{
|
| 133 |
+
"id": "PanTS_00000001",
|
| 134 |
+
"cond": {
|
| 135 |
+
"lesion_present": false,
|
| 136 |
+
"phase": "Non-contrast",
|
| 137 |
+
"fov": "chest_abdomen_pelvis",
|
| 138 |
+
"sex": "M",
|
| 139 |
+
"age": 63.0,
|
| 140 |
+
"pancreas_volume_cc": 82.3,
|
| 141 |
+
"gallbladder_present": true,
|
| 142 |
+
"z_physical_mm": 270.0,
|
| 143 |
+
"study_type": "ct thorax-abdomen-pelvis",
|
| 144 |
+
"manufacturer": "SIEMENS",
|
| 145 |
+
"lesion_list": [],
|
| 146 |
+
"lesion_components_mask": []
|
| 147 |
+
},
|
| 148 |
+
"captions": {
|
| 149 |
+
"V1_long_narrative": "...",
|
| 150 |
+
"V2_terse_impression": "...",
|
| 151 |
+
"V3_organ_bullet": "...",
|
| 152 |
+
"V4_tag_string": "...",
|
| 153 |
+
"V5_layered_findings": "...",
|
| 154 |
+
"V6_qa_pair": "...", // lesion+ only
|
| 155 |
+
"V7_pancreas_only": "..." // lesion+ w/ mask components only
|
| 156 |
+
}
|
| 157 |
+
}
|
| 158 |
+
```
|
| 159 |
+
|
| 160 |
+
## Schema: `canonical/canonical_facts.jsonl`
|
| 161 |
+
|
| 162 |
+
One JSON object per line, with sub-dicts `report` (organ statuses + lesion list
|
| 163 |
+
from the free-text report), `mask` (organ volumes + bbox + pancreas geometry
|
| 164 |
+
from the segmentation masks), and `canonical` (merged, deduplicated view used
|
| 165 |
+
by the captioners). See `docs/LDM_TRAINING_PLAN_v4.md` for the full spec.
|
| 166 |
+
|
| 167 |
+
## Phase QC caveat
|
| 168 |
+
|
| 169 |
+
`audit/phase_audit.jsonl` flags canonical-phase vs HU-heuristic disagreement.
|
| 170 |
+
Canonical phase is the LLM-parsed value from the report and is treated as the
|
| 171 |
+
source of truth for training. The HU heuristic is a noisy secondary signal: it
|
| 172 |
+
agrees on 49.8% of cases and disagrees on 50.2%. **Do not filter out the
|
| 173 |
+
disagreements — use canonical.** The flag is provided for downstream QC.
|
| 174 |
+
|
| 175 |
+
## What's NOT here (and why)
|
| 176 |
+
|
| 177 |
+
- **CT volumes and masks**: the upstream PanTS release is the source. This
|
| 178 |
+
add-on is metadata + captions only.
|
| 179 |
+
- **VAE latents**: the Wan 2.2 VAE weights are not yet bundled with the
|
| 180 |
+
training pipeline. `code/vae_sanity_gate.py` and `code/pack_webdataset.py`
|
| 181 |
+
run end-to-end once latents are cached.
|
| 182 |
+
- **DiT-3D backbone**: `code/dry_run_train.py` uses a 3-conv stand-in so the
|
| 183 |
+
training loop is verifiable in isolation.
|
| 184 |
+
- **Patient identifiers**: the upstream PanTS dataset is already de-identified;
|
| 185 |
+
no additional identifiers are added here.
|
| 186 |
+
|
| 187 |
+
## Reproduction
|
| 188 |
+
|
| 189 |
+
All caption generation is deterministic given the canonical-facts file and the
|
| 190 |
+
VLM prompts in `code/caption_generation/`. A subset of intermediate artifacts
|
| 191 |
+
(`captions_v1_initial.jsonl` → `..._v2_after_fusion.jsonl` → `..._v3_after_regen.jsonl`
|
| 192 |
+
→ `captions_final.jsonl`) is included to let downstream researchers audit or
|
| 193 |
+
re-run any stage.
|
| 194 |
+
|
| 195 |
+
## License
|
| 196 |
+
|
| 197 |
+
Code in `code/` is released under the MIT license.
|
| 198 |
+
Captions and structured metadata in `captions/`, `canonical/`, `splits/`,
|
| 199 |
+
`samples/`, and `audit/` are released under CC-BY-4.0.
|
| 200 |
+
|
| 201 |
+
Upstream PanTS CT volumes and masks are covered by their own license — this
|
| 202 |
+
add-on dataset does not redistribute or relicense them.
|
| 203 |
+
|
| 204 |
+
## Citation
|
| 205 |
+
|
| 206 |
+
Please cite the upstream PanTS release and this companion dataset.
|
| 207 |
+
Bibtex for this companion dataset will be added upon public release.
|
raw_pants_train_test/metadata/pants-captions-ldm/UPLOAD_INSTRUCTIONS.md
ADDED
|
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Upload instructions (anonymous, public)
|
| 2 |
+
|
| 3 |
+
## Prerequisites
|
| 4 |
+
|
| 5 |
+
1. Create a HuggingFace account with no personal-identifying info
|
| 6 |
+
(username + throwaway email). Do NOT use your real name or
|
| 7 |
+
institutional email.
|
| 8 |
+
2. Generate a token: https://huggingface.co/settings/tokens
|
| 9 |
+
→ "Create new token" → name: "pants-captions-upload" → permission:
|
| 10 |
+
"Write access to contents/settings of selected repos" → create.
|
| 11 |
+
3. The token starts with `hf_` — keep it secret, we will need it once.
|
| 12 |
+
|
| 13 |
+
## Option A — `huggingface_hub` Python client (simplest)
|
| 14 |
+
|
| 15 |
+
```bash
|
| 16 |
+
pip install --upgrade huggingface_hub
|
| 17 |
+
|
| 18 |
+
# one-time login (pastes token into ~/.cache/huggingface/token):
|
| 19 |
+
huggingface-cli login # paste the hf_... token when prompted
|
| 20 |
+
|
| 21 |
+
# 1) create the dataset repo (public)
|
| 22 |
+
huggingface-cli repo create pants-captions --type dataset
|
| 23 |
+
# or programmatically:
|
| 24 |
+
# from huggingface_hub import create_repo
|
| 25 |
+
# create_repo("anon/pants-captions", repo_type="dataset", private=False)
|
| 26 |
+
|
| 27 |
+
# 2) upload the whole staging dir
|
| 28 |
+
huggingface-cli upload <anon-username>/pants-captions \
|
| 29 |
+
/home/ubuntu/pants_hf_release \
|
| 30 |
+
. \
|
| 31 |
+
--repo-type dataset
|
| 32 |
+
```
|
| 33 |
+
|
| 34 |
+
The `.` as the second positional is the in-repo path (root).
|
| 35 |
+
|
| 36 |
+
## Option B — single tarball
|
| 37 |
+
|
| 38 |
+
If you prefer to upload one file, we also build `pants_hf_release.tar.gz`
|
| 39 |
+
(see `MANIFEST.sha256` for integrity). Upload it as an asset, then untar
|
| 40 |
+
in-place in a CI action, or release it as a release artifact.
|
| 41 |
+
|
| 42 |
+
```bash
|
| 43 |
+
cd /home/ubuntu
|
| 44 |
+
tar -czf pants_hf_release.tar.gz pants_hf_release/
|
| 45 |
+
sha256sum pants_hf_release.tar.gz > pants_hf_release.tar.gz.sha256
|
| 46 |
+
# total size: ~166 MB uncompressed, expected ~60-80 MB compressed — well
|
| 47 |
+
# under the 50 GB per-tarball cap.
|
| 48 |
+
|
| 49 |
+
huggingface-cli upload <anon-username>/pants-captions-tarball \
|
| 50 |
+
pants_hf_release.tar.gz \
|
| 51 |
+
--repo-type dataset
|
| 52 |
+
```
|
| 53 |
+
|
| 54 |
+
## Anonymity checklist (BEFORE upload)
|
| 55 |
+
|
| 56 |
+
- [ ] You have verified your HF username contains no PII and is not linked
|
| 57 |
+
to any other account.
|
| 58 |
+
- [ ] You are uploading over a throwaway network / TOR if institutional
|
| 59 |
+
IP-level correlation matters for your threat model.
|
| 60 |
+
- [ ] You have reviewed `code/` — no absolute paths of the form
|
| 61 |
+
`/home/<your-username>/...` remain (verified: we only use
|
| 62 |
+
`os.environ.get("PANTS_DATA_ROOT", ...)` and
|
| 63 |
+
`Path(__file__).resolve().parents[1]`).
|
| 64 |
+
- [ ] You have reviewed `docs/` — the planning markdowns are scrubbed of
|
| 65 |
+
absolute paths, but they do contain prose you may want to double-check.
|
| 66 |
+
- [ ] You have reviewed `samples/*/annotations.json` — they contain case IDs
|
| 67 |
+
(`PanTS_00000001`-style) which are the upstream dataset's own IDs and
|
| 68 |
+
contain no patient info.
|
| 69 |
+
|
| 70 |
+
## Files to review (largest / most likely to contain forgotten info)
|
| 71 |
+
|
| 72 |
+
| file | size | type |
|
| 73 |
+
|------|------|------|
|
| 74 |
+
| captions/captions_final.jsonl | 40 MB | 7 captions × 9,901 rows |
|
| 75 |
+
| captions/captions_full_v2.jsonl | 47 MB | intermediate |
|
| 76 |
+
| captions/captions_full_v3.jsonl | 48 MB | intermediate |
|
| 77 |
+
| canonical/canonical_facts.jsonl | 27 MB | structured facts |
|
| 78 |
+
| audit/phase_audit.jsonl | 2.5 MB | 9,901 phase-QC rows |
|
| 79 |
+
| everything else | small | code/docs |
|
| 80 |
+
|
| 81 |
+
Total: ~166 MB, one tarball will be well under the 50 GB cap.
|
raw_pants_train_test/metadata/pants-captions-ldm/audit/audit_final.md
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# F6 Final Corpus QA (2026-04-20)
|
| 2 |
+
|
| 3 |
+
- Source: `${REPO_ROOT}/captions/captions_final.jsonl`
|
| 4 |
+
- Cases: 9901
|
| 5 |
+
- Flagged rows (case × variant): 2163
|
| 6 |
+
|
| 7 |
+
| Variant | Total | OK | Pass % | P50 words | P95 words |
|
| 8 |
+
|---|---|---|---|---|---|
|
| 9 |
+
| V1_long_narrative | 9901 | 9536 | 96.3% | 213 | 239 |
|
| 10 |
+
| V2_terse_impression | 9901 | 9684 | 97.8% | 20 | 26 |
|
| 11 |
+
| V3_organ_bullet | 9901 | 9270 | 93.6% | 149 | 178 |
|
| 12 |
+
| V4_tag_string | 9901 | 9302 | 94.0% | 32 | 44 |
|
| 13 |
+
| V5_layered_findings | 9901 | 9557 | 96.5% | 127 | 149 |
|
| 14 |
+
| V6_qa_pair | 958 | 958 | 100.0% | 35 | 51 |
|
| 15 |
+
| V7_pancreas_only | 256 | 249 | 97.3% | 69 | 85 |
|
| 16 |
+
|
| 17 |
+
## Top violations
|
| 18 |
+
|
| 19 |
+
- `too_short`: 1562
|
| 20 |
+
- `too_long`: 461
|
| 21 |
+
- `hallucinated_lesion_mention`: 95
|
| 22 |
+
- `missing_lesion_mention`: 44
|
| 23 |
+
- `size_not_echoed`: 29
|
raw_pants_train_test/metadata/pants-captions-ldm/audit/audit_v2.md
ADDED
|
@@ -0,0 +1,187 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# F5 Caption Audit Report (2026-04-20)
|
| 2 |
+
|
| 3 |
+
**TL;DR**
|
| 4 |
+
- 9,901 cases, 6 variants each; all cases have all variants filled (zero full-failure rows).
|
| 5 |
+
- Pass rates after incidental-cyst and V6_negative deprecation fixes:
|
| 6 |
+
V1 91.9% / V2 97.1% / V3 90.2% / V4 94.0% / V5 92.9% / V6_qa 99.7%.
|
| 7 |
+
- Dominant residual issue: **token truncation** on V1/V3/V5 (~7-10% per variant),
|
| 8 |
+
totalling **2,704 (case × variant) rows** that need regen with a higher max_tokens.
|
| 9 |
+
- Real hallucinations (pancreatic lesion mention on lesion- case): **64 / 8,943 = 0.72%**.
|
| 10 |
+
- Real missing-lesion (lesion+ case but text silent): **46 / 958 = 4.8%** — mostly V2/V4.
|
| 11 |
+
- Banned-word hits: 1 ("adenocarcinoma" in PanTS_00000789 V1) — queued for regen.
|
| 12 |
+
- V6_negative_descriptive is deprecated by user decision 2026-04-20; will not ship.
|
| 13 |
+
|
| 14 |
+
## Raw counts
|
| 15 |
+
|
| 16 |
+
- Cases audited: 9901
|
| 17 |
+
- Canonical intersection misses: 0
|
| 18 |
+
- Total flagged (case, variant) rows: 3733
|
| 19 |
+
- Regen list (truncation + banned): 2,705 rows across 2,407 cases
|
| 20 |
+
|
| 21 |
+
---
|
| 22 |
+
|
| 23 |
+
## Final corpus QA (post F5d regen + V7 merge + V6_neg drop, 2026-04-20)
|
| 24 |
+
|
| 25 |
+
Final corpus `captions_final.jsonl` contains 9,901 cases. V6_negative_descriptive dropped
|
| 26 |
+
per user directive; V7_pancreas_only added for 256 / 293 V7-eligible cases (37 dropped
|
| 27 |
+
by the banned-upgrade-word filter — they contained "invasion" / "encasement" despite
|
| 28 |
+
the whitelist prompt). Pass rates are now V1 96.3% / V2 97.8% / V3 93.6% / V4 94.0% /
|
| 29 |
+
V5 96.5% / V6_qa 100% / V7 97.3%. Residual flags are dominated by 1,562 "too_short"
|
| 30 |
+
(some V3 bullet captions come in at 50-60 words on slim-report cases) and 461
|
| 31 |
+
"too_long" (V3 and V5 occasionally overshoot by 10-20 words). Real content issues
|
| 32 |
+
are small: 95 "hallucinated_lesion_mention" (0.97% of lesion- rows, mostly V1 prose
|
| 33 |
+
mentioning an incidental renal cyst the INCIDENTAL_CYST regex didn't catch) and 44
|
| 34 |
+
"missing_lesion_mention" on lesion+ rows (largely V2/V4 where the tag/impression
|
| 35 |
+
format lost the term). Banned-word hits: 0. This corpus is the training-ready one;
|
| 36 |
+
see `audit_final.md` for the full table.
|
| 37 |
+
|
| 38 |
+
## Per-variant pass rate
|
| 39 |
+
|
| 40 |
+
| Variant | Total | OK | Pass % | P50 words | P95 words |
|
| 41 |
+
|---|---|---|---|---|---|
|
| 42 |
+
| V1_long_narrative | 9901 | 9100 | 91.9% | 211 | 243 |
|
| 43 |
+
| V2_terse_impression | 9901 | 9614 | 97.1% | 20 | 28 |
|
| 44 |
+
| V3_organ_bullet | 9901 | 8934 | 90.2% | 148 | 181 |
|
| 45 |
+
| V4_tag_string | 9901 | 9302 | 94.0% | 32 | 44 |
|
| 46 |
+
| V5_layered_findings | 9901 | 9202 | 92.9% | 127 | 154 |
|
| 47 |
+
| V6_negative_descriptive | 8943 | 8566 | 95.8% | 102 | 112 |
|
| 48 |
+
| V6_qa_pair | 958 | 955 | 99.7% | 35 | 52 |
|
| 49 |
+
|
| 50 |
+
## Violation breakdown
|
| 51 |
+
|
| 52 |
+
### V1_long_narrative
|
| 53 |
+
|
| 54 |
+
- `too_short`: 708
|
| 55 |
+
- `too_long`: 59
|
| 56 |
+
- `hallucinated_lesion_mention`: 28
|
| 57 |
+
- `size_not_echoed`: 13
|
| 58 |
+
- `missing_subregion`: 3
|
| 59 |
+
- `missing_lesion_mention`: 2
|
| 60 |
+
- `banned`: 1
|
| 61 |
+
|
| 62 |
+
### V2_terse_impression
|
| 63 |
+
|
| 64 |
+
- `too_short`: 183
|
| 65 |
+
- `too_long`: 69
|
| 66 |
+
- `missing_lesion_mention`: 31
|
| 67 |
+
- `size_not_echoed`: 4
|
| 68 |
+
- `hallucinated_lesion_mention`: 1
|
| 69 |
+
|
| 70 |
+
### V3_organ_bullet
|
| 71 |
+
|
| 72 |
+
- `too_long`: 526
|
| 73 |
+
- `too_short`: 435
|
| 74 |
+
- `missing_subregion`: 8
|
| 75 |
+
- `size_not_echoed`: 6
|
| 76 |
+
- `hallucinated_lesion_mention`: 4
|
| 77 |
+
- `missing_lesion_mention`: 3
|
| 78 |
+
|
| 79 |
+
### V4_tag_string
|
| 80 |
+
|
| 81 |
+
- `too_short`: 576
|
| 82 |
+
- `hallucinated_lesion_mention`: 22
|
| 83 |
+
- `size_not_echoed`: 7
|
| 84 |
+
- `missing_lesion_mention`: 5
|
| 85 |
+
|
| 86 |
+
### V5_layered_findings
|
| 87 |
+
|
| 88 |
+
- `too_short`: 517
|
| 89 |
+
- `too_long`: 170
|
| 90 |
+
- `hallucinated_lesion_mention`: 9
|
| 91 |
+
- `size_not_echoed`: 6
|
| 92 |
+
- `missing_subregion`: 3
|
| 93 |
+
- `missing_lesion_mention`: 3
|
| 94 |
+
|
| 95 |
+
### V6_negative_descriptive
|
| 96 |
+
|
| 97 |
+
- `too_short`: 218
|
| 98 |
+
- `too_long`: 159
|
| 99 |
+
|
| 100 |
+
### V6_qa_pair
|
| 101 |
+
|
| 102 |
+
- `size_not_echoed`: 3
|
| 103 |
+
- `missing_lesion_mention`: 2
|
| 104 |
+
|
| 105 |
+
## Sample flagged rows (first 20)
|
| 106 |
+
|
| 107 |
+
- `PanTS_00000032` / `V3_organ_bullet` / lesion+=False / viol=['too_short:57<60']
|
| 108 |
+
- text: - Pancreas: The pancreas demonstrates normal size and morphology. The parenchymal texture appears homogeneous with clean fat planes.
|
| 109 |
+
- Liver: The liver shows normal size. No focal ...
|
| 110 |
+
- `PanTS_00000007` / `V5_layered_findings` / lesion+=False / viol=['too_short:60<80']
|
| 111 |
+
- text: TECHNIQUE: A venous phase whole-body CT was performed, demonstrating prominent beam-hardening streaks from high-density contrast within the urinary bladder and vascular structures ...
|
| 112 |
+
- `PanTS_00000034` / `V6_negative_descriptive` / lesion+=False / viol=['too_long:153>150']
|
| 113 |
+
- text: This arterial phase CT of the chest, abdomen, and pelvis demonstrates a 71-year-old female patient with an average to broad habitus, featuring moderate subcutaneous adipose tissue ...
|
| 114 |
+
- `PanTS_00000057` / `V5_layered_findings` / lesion+=False / viol=['too_long:270>220']
|
| 115 |
+
- text: TECHNIQUE: Non-contrast whole-body CT imaging was performed, demonstrating beam-hardening artifacts adjacent to high-density material within the aorta and renal collecting systems,...
|
| 116 |
+
- `PanTS_00000011` / `V4_tag_string` / lesion+=False / viol=['too_short:3<6tags']
|
| 117 |
+
- text: abdominal ct, non-contrast phase, whole body fov": "The user wants me to write an SDXL-style comma-separated tag prompt.
|
| 118 |
+
* **Scan:
|
| 119 |
+
* **Scan:
|
| 120 |
+
* **Patient:
|
| 121 |
+
* ...
|
| 122 |
+
- `PanTS_00000011` / `V5_layered_findings` / lesion+=False / viol=['too_long:439>220']
|
| 123 |
+
- text: TECHNIQUE: A non-contrast computed tomography scan of the whole body was performed, covering a z-axis of 187.2 mm with no visible artifacts are present. No medical scan of the abdo...
|
| 124 |
+
- `PanTS_00000041` / `V5_layered_findings` / lesion+=False / viol=['too_long:221>220']
|
| 125 |
+
- text: TECHNIQUE: The user wants a radiology report in 3-section format.
|
| 126 |
+
|
| 127 |
+
1. **1. **Assessing the provided.
|
| 128 |
+
|
| 129 |
+
1. **Assessing the user wants me.
|
| 130 |
+
|
| 131 |
+
</think>
|
| 132 |
+
|
| 133 |
+
TECHNIQUE: The user wants a radi...
|
| 134 |
+
- `PanTS_00000062` / `V4_tag_string` / lesion+=False / viol=['too_short:1<6tags']
|
| 135 |
+
- text: abdominion
|
| 136 |
+
The user wants me to the user wants me to the user wants me to the user wants me to the user wants me to write to write a caption for an SDXL-style comma-separated tag p...
|
| 137 |
+
- `PanTS_00000053` / `V3_organ_bullet` / lesion+=False / viol=['too_short:24<60']
|
| 138 |
+
- text: - Pancreas: Non-contrast chest and pelvis demonstrates an enlarged pancreas; the liver and spleen appear normal in size and attenuation are without focal lesions....
|
| 139 |
+
- `PanTS_00000016` / `V3_organ_bullet` / lesion+=False / viol=['too_long:181>180']
|
| 140 |
+
- text: - **Analyzing the provided a per-organ structured bulleted caption.
|
| 141 |
+
|
| 142 |
+
FACTS:
|
| 143 |
+
{
|
| 144 |
+
"scan:
|
| 145 |
+
|
| 146 |
+
- **Pancreased on:
|
| 147 |
+
- **Pancreas.
|
| 148 |
+
|
| 149 |
+
** - **Masked images and RULES:
|
| 150 |
+
- **1. **Analyze the inpu...
|
| 151 |
+
- `PanTS_00000046` / `V5_layered_findings` / lesion+=False / viol=['too_long:258>220']
|
| 152 |
+
- text: TECHNIQUE: A non-contrast computed tomography scan of the chest and abdomen was performed, covering a z-axis range of 17.
|
| 153 |
+
|
| 154 |
+
Okay, let's break down the provided JSON data.
|
| 155 |
+
2. **Anal...
|
| 156 |
+
- `PanTS_00000055` / `V5_layered_findings` / lesion+=False / viol=['too_long:275>220']
|
| 157 |
+
- text: TECHNIQUE: Venous phase whole-body CT was not provided with a specific format.
|
| 158 |
+
|
| 159 |
+
The user wants a radiology report in 3-section format.
|
| 160 |
+
|
| 161 |
+
The user wants a radiology report in 3-secti...
|
| 162 |
+
- `PanTS_00000064` / `V1_long_narrative` / lesion+=False / viol=['too_short:117<120']
|
| 163 |
+
- text: This non-contrast computed tomography scan covers the chest, abdomen, and pelvis with a z-axis coverage of 390.0 mm in a 78-year-old male patient. The patient exhibits an average b...
|
| 164 |
+
- `PanTS_00000064` / `V3_organ_bullet` / lesion+=False / viol=['too_long:194>180']
|
| 165 |
+
- text: - Pancreas: The gland demonstrates normal size with a volume of 69.1 cc and a mean attenuation of 44.2 HU. The parenchyma appears homogeneous with clean fat planes surrounding the ...
|
| 166 |
+
- `PanTS_00000078` / `V5_layered_findings` / lesion+=False / viol=['too_short:48<80']
|
| 167 |
+
- text: TECHNIQUE: A non-contrast whole-body CT scan covering 172.5 cm in the z-axis reveals significant beam-hardening artifacts radiating from high-density: A non-contrast whole-body CT ...
|
| 168 |
+
- `PanTS_00000056` / `V4_tag_string` / lesion+=False / viol=['too_short:3<6tags']
|
| 169 |
+
- text: abdominal ct, non-contrast phase non-contrast whole_body>
|
| 170 |
+
|
| 171 |
+
1. **Analyze the user wants a SDXL-style comma-separated tag prompt.
|
| 172 |
+
* **Facts:**
|
| 173 |
+
* **Facts:**
|
| 174 |
+
* *...
|
| 175 |
+
- `PanTS_00000072` / `V1_long_narrative` / lesion+=False / viol=['too_short:40<120']
|
| 176 |
+
- text: This arterial phase whole-body CT scan encompasses anatomic ct, whole-body CT scan phase: "The user wants a caption:** The user wants a caption for training a text-to-image medical...
|
| 177 |
+
- `PanTS_00000052` / `V3_organ_bullet` / lesion+=False / viol=['too_short:44<60']
|
| 178 |
+
- text: - Pancreas: The gland demonstrates homogeneous texture with a clean fat plane surrounding the organ; no focal lesion is observed.
|
| 179 |
+
- Liver: Parenchyma: The liver shows normal attenu...
|
| 180 |
+
- `PanTS_00000081` / `V1_long_narrative` / lesion+=False / viol=['too_short:57<120']
|
| 181 |
+
- text: This venous phase whole-body CT scan encompasses the abdomen with moderate subcutaneous fat with average patient habitus is visualized. The pancreas demonstrates a homogeneous text...
|
| 182 |
+
- `PanTS_00000067` / `V1_long_narrative` / lesion+=False / viol=['too_short:116<120']
|
| 183 |
+
- text: This arterial phase whole-body CT scan covers a very clearly, I need to write a caption for training a text-to-image model.
|
| 184 |
+
- The input image.
|
| 185 |
+
- The input image is.
|
| 186 |
+
- The u...
|
| 187 |
+
|
raw_pants_train_test/metadata/pants-captions-ldm/audit/phase_audit.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
raw_pants_train_test/metadata/pants-captions-ldm/audit/phase_audit_summary.md
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Phase audit summary (N=9901)
|
| 2 |
+
|
| 3 |
+
## Phase distribution by source
|
| 4 |
+
| phase | canonical | hu_heur | caption | report |
|
| 5 |
+
|---------------|-----------|---------|---------|--------|
|
| 6 |
+
| Arterial | 2450 | 994 | 2449 | 0 |
|
| 7 |
+
| Delay | 68 | 368 | 68 | 0 |
|
| 8 |
+
| Non-contrast | 4485 | 3540 | 4484 | 0 |
|
| 9 |
+
| Venous | 2897 | 3275 | 2897 | 0 |
|
| 10 |
+
| unknown | 1 | 1724 | 3 | 9901 |
|
| 11 |
+
|
| 12 |
+
## Agreement across sources (k = max votes for any single label)
|
| 13 |
+
k=1: 3 (0.0%)
|
| 14 |
+
k=2: 6689 (67.6%)
|
| 15 |
+
k=3: 3209 (32.4%)
|
| 16 |
+
|
| 17 |
+
## Majority-vote agreed phase
|
| 18 |
+
Non-contrast : 4484 (45.3%)
|
| 19 |
+
Venous : 2898 (29.3%)
|
| 20 |
+
Arterial : 2449 (24.7%)
|
| 21 |
+
Delay : 68 (0.7%)
|
| 22 |
+
none : 2 (0.0%)
|
| 23 |
+
|
| 24 |
+
## canonical vs HU-heuristic confusion (top 15)
|
| 25 |
+
canonical \ hu count
|
| 26 |
+
Non-contrast Non-contrast 1843
|
| 27 |
+
Non-contrast Venous 1339
|
| 28 |
+
Venous Venous 1081
|
| 29 |
+
Venous Non-contrast 905
|
| 30 |
+
Arterial Venous 831
|
| 31 |
+
Arterial Non-contrast 767
|
| 32 |
+
Non-contrast unknown 748
|
| 33 |
+
Venous unknown 491
|
| 34 |
+
Arterial unknown 475
|
| 35 |
+
Non-contrast Arterial 399
|
| 36 |
+
Venous Arterial 304
|
| 37 |
+
Arterial Arterial 283
|
| 38 |
+
Non-contrast Delay 156
|
| 39 |
+
Venous Delay 116
|
| 40 |
+
Arterial Delay 94
|
| 41 |
+
|
| 42 |
+
canonical-vs-hu strong disagreement (both non-unknown, labels differ): 4967 (50.2%)
|
raw_pants_train_test/metadata/pants-captions-ldm/captions/captions_pilot.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
raw_pants_train_test/metadata/pants-captions-ldm/captions/captions_v1_initial.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
raw_pants_train_test/metadata/pants-captions-ldm/captions/captions_v7_pancreas_only.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
raw_pants_train_test/metadata/pants-captions-ldm/code/aug_config.py
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
P5: Augmentation config for PanTS LDM training (per v4 §5).
|
| 3 |
+
|
| 4 |
+
Summary of v4 expert blockers that led to this policy:
|
| 5 |
+
|
| 6 |
+
LR flip OFF:
|
| 7 |
+
- Pancreatic head/body/tail laterality is non-bilateral (head is right,
|
| 8 |
+
tail is left; duct geometry is chiral).
|
| 9 |
+
- Gallbladder is on the right only.
|
| 10 |
+
- Kidney sizes, renal cysts — left/right matter.
|
| 11 |
+
- Flipping destroys the supervision that the caption carries.
|
| 12 |
+
|
| 13 |
+
Spacing jitter OFF:
|
| 14 |
+
- The bucket-level voxel_spacing_mm is a conditioning input. Jittering
|
| 15 |
+
it means the model is asked to generate the same scan at a
|
| 16 |
+
re-interpreted spacing without re-encoding the latent cache. Breaks
|
| 17 |
+
latent/prompt alignment.
|
| 18 |
+
|
| 19 |
+
HU jitter OFF:
|
| 20 |
+
- Captions encode HU_mean / enhancement (iso/hypo/hyper-attenuating)
|
| 21 |
+
from canonical_facts. Jittering HU contradicts the text.
|
| 22 |
+
|
| 23 |
+
Latent noise ON (σ = 0.02 · std(z)):
|
| 24 |
+
- Small per-channel Gaussian on the VAE-encoded latent. Regularizes
|
| 25 |
+
against latent-space mode collapse.
|
| 26 |
+
|
| 27 |
+
Crop-origin jitter ON (±8 voxels each axis, uniform):
|
| 28 |
+
- Cheap translation augmentation that keeps the scale and content
|
| 29 |
+
consistent (only shifts the window).
|
| 30 |
+
"""
|
| 31 |
+
from dataclasses import dataclass
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
@dataclass(frozen=True)
|
| 35 |
+
class AugmentationConfig:
|
| 36 |
+
lr_flip: bool = False # OFF — laterality matters
|
| 37 |
+
spacing_jitter_pct: float = 0 # OFF — cond mismatch
|
| 38 |
+
hu_jitter: float = 0.0 # OFF — caption mismatch
|
| 39 |
+
latent_noise_sigma_rel: float = 0.02 # σ = 0.02 · std(z), per channel
|
| 40 |
+
crop_origin_jitter_vox: int = 8 # uniform(-8, +8) per axis
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
DEFAULT = AugmentationConfig()
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
if __name__ == "__main__":
|
| 47 |
+
print(DEFAULT)
|
raw_pants_train_test/metadata/pants-captions-ldm/code/build_bucket_spec.py
ADDED
|
@@ -0,0 +1,75 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
"""
|
| 3 |
+
P3: Build bucket_spec.json.
|
| 4 |
+
|
| 5 |
+
Bucket = (FOV class, shape (D,H,W)).
|
| 6 |
+
Training uses D % 4 == 0 (Wan 2.2 VAE pads internally for D=4k+1).
|
| 7 |
+
Tokens counted AFTER DiT-3D patch embedding.
|
| 8 |
+
|
| 9 |
+
Wan 2.2 VAE compression: D → D/4 (ceil((D+1)/4)), H → H/16, W → W/16, C=48
|
| 10 |
+
DiT-3D patch size: (pD, pH, pW) = (1, 2, 2)
|
| 11 |
+
Tokens per sample = ceil(D_lat / pD) * ceil(H_lat / pH) * ceil(W_lat / pW)
|
| 12 |
+
|
| 13 |
+
Token cap: 8192 (memory envelope for DiT-XL/2 on 8×A100 80GB).
|
| 14 |
+
"""
|
| 15 |
+
import json, os, math
|
| 16 |
+
from pathlib import Path
|
| 17 |
+
|
| 18 |
+
OUT = Path(__file__).resolve().parent / "bucket_spec.json"
|
| 19 |
+
TOKEN_CAP = 8192
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
def wan_latent_shape(D, H, W):
|
| 23 |
+
"""Wan 2.2 VAE 4x16x16 compression with 48 channels."""
|
| 24 |
+
return (math.ceil((D + 1) / 4) if D % 4 else D // 4 + 1, H // 16, W // 16)
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
def dit_tokens(D_lat, H_lat, W_lat, patch=(1, 2, 2)):
|
| 28 |
+
pD, pH, pW = patch
|
| 29 |
+
return math.ceil(D_lat / pD) * math.ceil(H_lat / pH) * math.ceil(W_lat / pW)
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
# Design: cover whole-body / CAP / CA / abd / pan_crop with varied spacing.
|
| 33 |
+
BUCKETS = [
|
| 34 |
+
# name fov D H W voxel_spacing (sz, sy, sx) in mm stage
|
| 35 |
+
("B-whole", "whole_body", 128, 256, 256, [5.0, 2.0, 2.0], "A"),
|
| 36 |
+
("B-CAP", "chest_abdomen_pelvis", 96, 320, 256, [4.0, 2.0, 2.0], "A_B"),
|
| 37 |
+
("B-CA", "chest_abdomen", 80, 320, 256, [3.5, 2.0, 2.0], "B"),
|
| 38 |
+
("B-abd", "abdomen_only", 64, 256, 256, [3.0, 2.0, 2.0], "B"),
|
| 39 |
+
("B-abd-pelvis", "abdomen_pelvis", 64, 256, 256, [3.0, 2.0, 2.0], "B"),
|
| 40 |
+
("B-pan", "pan_crop", 64, 192, 256, [2.0, 1.0, 1.0], "C"),
|
| 41 |
+
]
|
| 42 |
+
|
| 43 |
+
spec = {"buckets": [], "token_cap": TOKEN_CAP,
|
| 44 |
+
"vae": {"compression": [4, 16, 16], "channels": 48, "family": "wan22"},
|
| 45 |
+
"dit_patch": [1, 2, 2]}
|
| 46 |
+
for name, fov, D, H, W, sp, stage in BUCKETS:
|
| 47 |
+
assert D % 4 == 0, f"D for {name} must be divisible by 4 (training convention)"
|
| 48 |
+
D_lat, H_lat, W_lat = wan_latent_shape(D, H, W)
|
| 49 |
+
toks = dit_tokens(D_lat, H_lat, W_lat)
|
| 50 |
+
assert H % 16 == 0 and W % 16 == 0, f"H,W for {name} must be multiples of 16"
|
| 51 |
+
assert toks <= TOKEN_CAP, f"bucket {name} tokens {toks} > cap {TOKEN_CAP}"
|
| 52 |
+
cover_mm = [round(D * sp[0], 0), round(H * sp[1], 0), round(W * sp[2], 0)]
|
| 53 |
+
spec["buckets"].append({
|
| 54 |
+
"name": name,
|
| 55 |
+
"fov": fov,
|
| 56 |
+
"shape_dhw": [D, H, W],
|
| 57 |
+
"voxel_spacing_mm": sp,
|
| 58 |
+
"coverage_mm": cover_mm,
|
| 59 |
+
"latent_shape_dhw": [D_lat, H_lat, W_lat],
|
| 60 |
+
"tokens_after_patch": toks,
|
| 61 |
+
"stage": stage,
|
| 62 |
+
})
|
| 63 |
+
|
| 64 |
+
OUT.parent.mkdir(parents=True, exist_ok=True)
|
| 65 |
+
with open(OUT, "w") as f:
|
| 66 |
+
json.dump(spec, f, indent=2)
|
| 67 |
+
|
| 68 |
+
print(f"wrote {OUT}")
|
| 69 |
+
print()
|
| 70 |
+
print(f"{'bucket':<16}{'fov':<24}{'shape':<16}{'cover mm':<22}{'latent':<14}{'tokens':<8}stage")
|
| 71 |
+
for b in spec["buckets"]:
|
| 72 |
+
s = "x".join(str(x) for x in b["shape_dhw"])
|
| 73 |
+
lat = "x".join(str(x) for x in b["latent_shape_dhw"])
|
| 74 |
+
cov = "x".join(str(int(x)) for x in b["coverage_mm"])
|
| 75 |
+
print(f"{b['name']:<16}{b['fov']:<24}{s:<16}{cov:<22}{lat:<14}{b['tokens_after_patch']:<8}{b['stage']}")
|
raw_pants_train_test/metadata/pants-captions-ldm/code/build_splits.py
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
"""
|
| 3 |
+
P2: Build splits.json from official PanTS-tr / PanTS-te.
|
| 4 |
+
|
| 5 |
+
- PanTS-tr: 9,000 cases from ${PANTS_DATA_ROOT}/ImageTr/
|
| 6 |
+
- PanTS-te: 901 cases from ${PANTS_DATA_ROOT}/ImageTe/ (strictly held out)
|
| 7 |
+
- train/val carve from PanTS-tr: deterministic every-11th id by sorted order → 818 val, 8,182 train
|
| 8 |
+
- test = all 901 from PanTS-te
|
| 9 |
+
|
| 10 |
+
Emit:
|
| 11 |
+
splits.json {train: [...], val: [...], test: [...]}
|
| 12 |
+
"""
|
| 13 |
+
import json, os
|
| 14 |
+
from pathlib import Path
|
| 15 |
+
|
| 16 |
+
TR = Path(os.environ.get("PANTS_DATA_ROOT", "./PanTS_data")) / "ImageTr"
|
| 17 |
+
TE = Path(os.environ.get("PANTS_DATA_ROOT", "./PanTS_data")) / "ImageTe"
|
| 18 |
+
OUT = Path(__file__).resolve().parent / "splits.json"
|
| 19 |
+
|
| 20 |
+
tr_ids = sorted([p.name for p in TR.iterdir() if p.is_dir()])
|
| 21 |
+
te_ids = sorted([p.name for p in TE.iterdir() if p.is_dir()])
|
| 22 |
+
assert len(tr_ids) == 9000, f"expected 9000 in PanTS-tr, got {len(tr_ids)}"
|
| 23 |
+
assert len(te_ids) == 901, f"expected 901 in PanTS-te, got {len(te_ids)}"
|
| 24 |
+
|
| 25 |
+
# every 11th id → val; remaining → train
|
| 26 |
+
val_ids = tr_ids[::11] # 818 cases (len 9000 → ceil(9000/11)=819, [::11] gives 819)
|
| 27 |
+
train_ids = [i for i in tr_ids if i not in set(val_ids)]
|
| 28 |
+
|
| 29 |
+
split = {"train": train_ids, "val": val_ids, "test": te_ids}
|
| 30 |
+
OUT.parent.mkdir(parents=True, exist_ok=True)
|
| 31 |
+
with open(OUT, "w") as f:
|
| 32 |
+
json.dump(split, f)
|
| 33 |
+
print(f"train: {len(train_ids)}")
|
| 34 |
+
print(f"val: {len(val_ids)}")
|
| 35 |
+
print(f"test: {len(te_ids)}")
|
| 36 |
+
print(f"wrote {OUT}")
|
raw_pants_train_test/metadata/pants-captions-ldm/code/dry_run_train.py
ADDED
|
@@ -0,0 +1,240 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
P8: 5k-step dry run on 200-case subset (v4 §14 item 8).
|
| 3 |
+
|
| 4 |
+
Intent
|
| 5 |
+
------
|
| 6 |
+
Before committing to full Stage-A (9,901 × T steps) training, run a 5k-step
|
| 7 |
+
dry-run on a 200-case subset to confirm:
|
| 8 |
+
* loss curve descends (log-scale) — not flat, not NaN
|
| 9 |
+
* no collate shape mismatches across buckets
|
| 10 |
+
* memory fits on 1x A100 80GB at batch=1 grad-accum=k for every bucket
|
| 11 |
+
* loss weight distribution under Min-SNR-γ=5 is well-behaved
|
| 12 |
+
* data I/O is not the bottleneck (< 20% of step time on cache-warm runs)
|
| 13 |
+
|
| 14 |
+
STATUS
|
| 15 |
+
------
|
| 16 |
+
Scaffold only — the model code (DiT-3D backbone + text encoder wrapping)
|
| 17 |
+
lives outside this repo and is not available on this machine. The script
|
| 18 |
+
assembles the full DataLoader, noise schedule, loss, and training loop
|
| 19 |
+
*except* for the DiT forward pass, which is replaced by a small stand-in
|
| 20 |
+
(a 3-layer conv net) so the pipeline can be exercised end-to-end.
|
| 21 |
+
|
| 22 |
+
When the real model is dropped in, replace `StandInDiT3D` with the actual
|
| 23 |
+
backbone and flip `--stand-in` to False.
|
| 24 |
+
|
| 25 |
+
Success gate (tracked in Weights & Biases if WANDB_PROJECT is set):
|
| 26 |
+
- train loss at step 5000 < 0.75 * loss at step 100 (on identical batches)
|
| 27 |
+
- no NaNs in 5000 steps
|
| 28 |
+
- sample-per-second throughput logged
|
| 29 |
+
"""
|
| 30 |
+
from __future__ import annotations
|
| 31 |
+
import argparse
|
| 32 |
+
import json
|
| 33 |
+
import math
|
| 34 |
+
import os
|
| 35 |
+
import random
|
| 36 |
+
import sys
|
| 37 |
+
import time
|
| 38 |
+
from pathlib import Path
|
| 39 |
+
|
| 40 |
+
import numpy as np
|
| 41 |
+
import torch
|
| 42 |
+
import torch.nn as nn
|
| 43 |
+
|
| 44 |
+
ROOT = Path(__file__).resolve().parents[1]
|
| 45 |
+
sys.path.insert(0, str(ROOT / "ldm_training"))
|
| 46 |
+
from schedule import ( # noqa: E402
|
| 47 |
+
make_schedule, snr, forward_q_sample, v_prediction_target,
|
| 48 |
+
min_snr_gamma_vpred_weight,
|
| 49 |
+
)
|
| 50 |
+
from aug_config import DEFAULT as AUG # noqa: E402
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
# --- stand-in model --------------------------------------------------------
|
| 54 |
+
|
| 55 |
+
class StandInDiT3D(nn.Module):
|
| 56 |
+
"""Cheap 3-layer 3D conv stand-in for the real DiT-3D.
|
| 57 |
+
Takes x_t (B, 48, D', H', W'), returns v-prediction of same shape.
|
| 58 |
+
The token conditioning is ignored. This exists only to exercise the
|
| 59 |
+
training loop; loss should still descend because the net has enough
|
| 60 |
+
capacity to memorize a few latents from a tiny dataset.
|
| 61 |
+
"""
|
| 62 |
+
def __init__(self, ch: int = 48):
|
| 63 |
+
super().__init__()
|
| 64 |
+
h = 128
|
| 65 |
+
self.net = nn.Sequential(
|
| 66 |
+
nn.Conv3d(ch, h, 3, padding=1),
|
| 67 |
+
nn.GELU(),
|
| 68 |
+
nn.Conv3d(h, h, 3, padding=1),
|
| 69 |
+
nn.GELU(),
|
| 70 |
+
nn.Conv3d(h, ch, 3, padding=1),
|
| 71 |
+
)
|
| 72 |
+
|
| 73 |
+
def forward(self, x_t, t, cond=None): # noqa: ARG002
|
| 74 |
+
return self.net(x_t)
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
# --- data ------------------------------------------------------------------
|
| 78 |
+
|
| 79 |
+
class DryRunLatentDataset(torch.utils.data.Dataset):
|
| 80 |
+
"""Loads {id}.latent.npy files from a directory, or synthesizes random
|
| 81 |
+
latents if --synthetic is passed. Conditioning dict is loaded per case
|
| 82 |
+
but unused by the stand-in model.
|
| 83 |
+
|
| 84 |
+
The synthetic mode exists so the dry-run pipeline can be validated
|
| 85 |
+
before any real VAE encoding is available.
|
| 86 |
+
"""
|
| 87 |
+
def __init__(self, ids: list[str], latent_dir: Path | None, bucket: dict,
|
| 88 |
+
synthetic: bool = False):
|
| 89 |
+
self.ids = ids
|
| 90 |
+
self.latent_dir = latent_dir
|
| 91 |
+
self.bucket = bucket
|
| 92 |
+
self.synthetic = synthetic
|
| 93 |
+
# compute latent shape from bucket
|
| 94 |
+
D, H, W = bucket["shape_dhw"]
|
| 95 |
+
self.latent_shape = (48,
|
| 96 |
+
math.ceil((D + 1) / 4) if D % 4 else D // 4 + 1,
|
| 97 |
+
H // 16, W // 16)
|
| 98 |
+
|
| 99 |
+
def __len__(self):
|
| 100 |
+
return len(self.ids)
|
| 101 |
+
|
| 102 |
+
def __getitem__(self, idx):
|
| 103 |
+
cid = self.ids[idx]
|
| 104 |
+
if self.synthetic or self.latent_dir is None:
|
| 105 |
+
z = np.random.randn(*self.latent_shape).astype(np.float32)
|
| 106 |
+
else:
|
| 107 |
+
p = self.latent_dir / f"{cid}.latent.npy"
|
| 108 |
+
z = np.load(p).astype(np.float32)
|
| 109 |
+
# latent noise augmentation: σ = sigma_rel · std(z)
|
| 110 |
+
if AUG.latent_noise_sigma_rel > 0:
|
| 111 |
+
z = z + AUG.latent_noise_sigma_rel * z.std() * np.random.randn(*z.shape).astype(np.float32)
|
| 112 |
+
return torch.from_numpy(z), cid
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
# --- training --------------------------------------------------------------
|
| 116 |
+
|
| 117 |
+
def train(args):
|
| 118 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 119 |
+
print(f"device: {device}")
|
| 120 |
+
|
| 121 |
+
# load bucket + 200-case subset
|
| 122 |
+
bucket_spec = json.loads((ROOT / "ldm_training" / "bucket_spec.json").read_text())
|
| 123 |
+
bucket = {b["name"]: b for b in bucket_spec["buckets"]}[args.bucket]
|
| 124 |
+
splits = json.loads((ROOT / "ldm_training" / "splits.json").read_text())
|
| 125 |
+
rng = random.Random(42)
|
| 126 |
+
ids = rng.sample(splits["train"], k=min(args.n_cases, len(splits["train"])))
|
| 127 |
+
print(f"bucket: {args.bucket} shape {bucket['shape_dhw']}")
|
| 128 |
+
print(f"subset: {len(ids)} cases")
|
| 129 |
+
|
| 130 |
+
# schedule
|
| 131 |
+
sch = make_schedule(T=1000, ztsnr=True)
|
| 132 |
+
alpha_bar = sch["alpha_bar"].to(device=device, dtype=torch.float32)
|
| 133 |
+
T = sch["T"]
|
| 134 |
+
|
| 135 |
+
# data
|
| 136 |
+
latent_dir = Path(args.latent_cache) if args.latent_cache else None
|
| 137 |
+
ds = DryRunLatentDataset(ids, latent_dir, bucket, synthetic=args.synthetic)
|
| 138 |
+
dl = torch.utils.data.DataLoader(ds, batch_size=args.batch, shuffle=True,
|
| 139 |
+
num_workers=args.workers, pin_memory=True,
|
| 140 |
+
drop_last=True)
|
| 141 |
+
print(f"latents: {'synthetic noise' if args.synthetic or latent_dir is None else str(latent_dir)}")
|
| 142 |
+
|
| 143 |
+
# model
|
| 144 |
+
model = StandInDiT3D(ch=48).to(device)
|
| 145 |
+
opt = torch.optim.AdamW(model.parameters(), lr=args.lr, betas=(0.9, 0.95),
|
| 146 |
+
weight_decay=0.01)
|
| 147 |
+
|
| 148 |
+
use_wandb = bool(os.environ.get("WANDB_PROJECT"))
|
| 149 |
+
if use_wandb:
|
| 150 |
+
import wandb
|
| 151 |
+
wandb.init(project=os.environ["WANDB_PROJECT"],
|
| 152 |
+
config=vars(args), name=f"p8-dryrun-{args.bucket}")
|
| 153 |
+
|
| 154 |
+
model.train()
|
| 155 |
+
step = 0
|
| 156 |
+
t0 = time.time()
|
| 157 |
+
loss_hist = []
|
| 158 |
+
nan_count = 0
|
| 159 |
+
while step < args.steps:
|
| 160 |
+
for x0, _cid in dl:
|
| 161 |
+
x0 = x0.to(device, non_blocking=True)
|
| 162 |
+
B = x0.shape[0]
|
| 163 |
+
t = torch.randint(0, T, (B,), device=device)
|
| 164 |
+
ab_t = alpha_bar[t].view(B, 1, 1, 1, 1)
|
| 165 |
+
noise = torch.randn_like(x0)
|
| 166 |
+
xt = forward_q_sample(x0, noise, ab_t)
|
| 167 |
+
v_tgt = v_prediction_target(x0, noise, ab_t)
|
| 168 |
+
v_pred = model(xt, t)
|
| 169 |
+
# per-sample MSE
|
| 170 |
+
diff = (v_pred - v_tgt).float()
|
| 171 |
+
per_sample = diff.flatten(1).pow(2).mean(dim=1)
|
| 172 |
+
snr_t = snr(ab_t.squeeze().float())
|
| 173 |
+
w = min_snr_gamma_vpred_weight(snr_t.view(-1), gamma=5.0)
|
| 174 |
+
loss = (w * per_sample).mean()
|
| 175 |
+
|
| 176 |
+
opt.zero_grad(set_to_none=True)
|
| 177 |
+
loss.backward()
|
| 178 |
+
nn.utils.clip_grad_norm_(model.parameters(), 1.0)
|
| 179 |
+
opt.step()
|
| 180 |
+
|
| 181 |
+
step += 1
|
| 182 |
+
loss_hist.append(float(loss.detach().cpu().item()))
|
| 183 |
+
if not math.isfinite(loss_hist[-1]):
|
| 184 |
+
nan_count += 1
|
| 185 |
+
if step % 100 == 0:
|
| 186 |
+
recent = np.mean(loss_hist[-100:])
|
| 187 |
+
sec_per_step = (time.time() - t0) / step
|
| 188 |
+
print(f"step {step:5d} loss={recent:.4f} "
|
| 189 |
+
f"sec/step={sec_per_step:.3f} "
|
| 190 |
+
f"samples/s={B / sec_per_step:.1f}")
|
| 191 |
+
if use_wandb:
|
| 192 |
+
wandb.log({"loss": recent, "step": step,
|
| 193 |
+
"sec_per_step": sec_per_step}, step=step)
|
| 194 |
+
if step >= args.steps:
|
| 195 |
+
break
|
| 196 |
+
|
| 197 |
+
# final gate check
|
| 198 |
+
loss_start = float(np.mean(loss_hist[50:150])) if len(loss_hist) > 150 else float('nan')
|
| 199 |
+
loss_end = float(np.mean(loss_hist[-100:])) if len(loss_hist) >= 100 else float('nan')
|
| 200 |
+
ratio = loss_end / loss_start if loss_start > 0 else float('nan')
|
| 201 |
+
passed = math.isfinite(loss_start) and math.isfinite(loss_end) and ratio < 0.75 and nan_count == 0
|
| 202 |
+
|
| 203 |
+
report = {
|
| 204 |
+
"steps": step,
|
| 205 |
+
"loss_start": loss_start,
|
| 206 |
+
"loss_end": loss_end,
|
| 207 |
+
"ratio": ratio,
|
| 208 |
+
"nan_count": nan_count,
|
| 209 |
+
"pass_gate": bool(passed),
|
| 210 |
+
"wall_sec": time.time() - t0,
|
| 211 |
+
}
|
| 212 |
+
out = ROOT / "ldm_training" / "dry_run_report.json"
|
| 213 |
+
out.write_text(json.dumps(report, indent=2))
|
| 214 |
+
print(json.dumps(report, indent=2))
|
| 215 |
+
print(f"wrote {out}")
|
| 216 |
+
if use_wandb:
|
| 217 |
+
import wandb
|
| 218 |
+
wandb.log(report)
|
| 219 |
+
wandb.finish()
|
| 220 |
+
|
| 221 |
+
|
| 222 |
+
def main():
|
| 223 |
+
ap = argparse.ArgumentParser()
|
| 224 |
+
ap.add_argument("--n-cases", type=int, default=200)
|
| 225 |
+
ap.add_argument("--bucket", default="B-pan")
|
| 226 |
+
ap.add_argument("--steps", type=int, default=5000)
|
| 227 |
+
ap.add_argument("--batch", type=int, default=2)
|
| 228 |
+
ap.add_argument("--workers", type=int, default=4)
|
| 229 |
+
ap.add_argument("--lr", type=float, default=1e-4)
|
| 230 |
+
ap.add_argument("--latent-cache", default=os.environ.get("LATENT_CACHE", ""))
|
| 231 |
+
ap.add_argument("--synthetic", action="store_true",
|
| 232 |
+
help="Use synthetic random-noise latents (plumbing test).")
|
| 233 |
+
ap.add_argument("--stand-in", action="store_true", default=True,
|
| 234 |
+
help="Use the 3-conv stand-in instead of a real DiT-3D.")
|
| 235 |
+
args = ap.parse_args()
|
| 236 |
+
train(args)
|
| 237 |
+
|
| 238 |
+
|
| 239 |
+
if __name__ == "__main__":
|
| 240 |
+
main()
|
raw_pants_train_test/metadata/pants-captions-ldm/code/pack_webdataset.py
ADDED
|
@@ -0,0 +1,229 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
P10: HuggingFace WebDataset packaging scaffold (v4 §14 item 10).
|
| 3 |
+
|
| 4 |
+
Packs the PanTS LDM training corpus as tar shards in the WebDataset format,
|
| 5 |
+
ready to upload to a HuggingFace dataset repo or to stream via `datasets`.
|
| 6 |
+
|
| 7 |
+
Shard layout
|
| 8 |
+
------------
|
| 9 |
+
One sample ≡ {CT volume latent, conditioning, all caption variants}:
|
| 10 |
+
{key}.latent.npy # (C=48, D_lat, H_lat, W_lat) float16 — after Wan VAE encode
|
| 11 |
+
{key}.cond.json # conditioning dict (bucket, spacing, ccrop, phase, fov, …)
|
| 12 |
+
{key}.captions.json # all 6/7 caption variants (V1..V5, V6_qa if lesion+, V7 if mask+)
|
| 13 |
+
{key}.manifest.json # tiny pointer dict: {id, bucket, tokens, split}
|
| 14 |
+
|
| 15 |
+
Shards are grouped by bucket so that a dataloader can build homogeneous
|
| 16 |
+
batches. Within each bucket, we shuffle then chunk into ~500-sample shards.
|
| 17 |
+
|
| 18 |
+
Stage of the pipeline
|
| 19 |
+
---------------------
|
| 20 |
+
This script assumes that the VAE-encoded latents already exist at
|
| 21 |
+
$LATENT_CACHE/{id}.latent.npy
|
| 22 |
+
If they do not, the script runs in --skip-latents mode: the tar still
|
| 23 |
+
contains cond + captions + manifest, and the consumer is expected to
|
| 24 |
+
compute latents on the fly (slower, but useful for debugging).
|
| 25 |
+
|
| 26 |
+
CLI
|
| 27 |
+
---
|
| 28 |
+
python pack_webdataset.py --split train --out /path/to/webdataset/
|
| 29 |
+
python pack_webdataset.py --split val --out /path/to/webdataset/
|
| 30 |
+
python pack_webdataset.py --split test --out /path/to/webdataset/
|
| 31 |
+
|
| 32 |
+
The script writes {split}-{bucket}-{idx:05d}.tar files and a
|
| 33 |
+
{split}_index.json that maps id → (bucket, shard, offset).
|
| 34 |
+
|
| 35 |
+
STATUS
|
| 36 |
+
------
|
| 37 |
+
Scaffold only — runs in --skip-latents mode to verify the packing + shard
|
| 38 |
+
layout + HF upload metadata. End-to-end packing is blocked on the VAE
|
| 39 |
+
sanity gate (P7) passing, which is blocked on Wan 2.2 weights being
|
| 40 |
+
available.
|
| 41 |
+
"""
|
| 42 |
+
from __future__ import annotations
|
| 43 |
+
import argparse
|
| 44 |
+
import hashlib
|
| 45 |
+
import io
|
| 46 |
+
import json
|
| 47 |
+
import os
|
| 48 |
+
import tarfile
|
| 49 |
+
from pathlib import Path
|
| 50 |
+
|
| 51 |
+
ROOT = Path(__file__).resolve().parents[1]
|
| 52 |
+
SPLITS = ROOT / "ldm_training" / "splits.json"
|
| 53 |
+
BUCKETS = ROOT / "ldm_training" / "bucket_spec.json"
|
| 54 |
+
CAPS = ROOT / "captions_out" / "captions_final.jsonl"
|
| 55 |
+
CANON = ROOT / "canonical_out" / "canonical_facts.jsonl"
|
| 56 |
+
DEFAULT_OUT = ROOT / "ldm_training" / "webdataset"
|
| 57 |
+
|
| 58 |
+
SHARD_SIZE = 500 # samples per shard
|
| 59 |
+
|
| 60 |
+
# One bucket per case. For PanTS v4, the intended bucket assignment per case
|
| 61 |
+
# comes from canonical.fov:
|
| 62 |
+
FOV_TO_BUCKET = {
|
| 63 |
+
"whole_body": "B-whole",
|
| 64 |
+
"chest_abdomen_pelvis": "B-CAP",
|
| 65 |
+
"chest_abdomen": "B-CA",
|
| 66 |
+
"abdomen_only": "B-abd",
|
| 67 |
+
"abdomen_pelvis": "B-abd-pelvis",
|
| 68 |
+
# "pan_crop" is a derived bucket produced by preprocess_pan_crop (stage C);
|
| 69 |
+
# raw cases do not start in this bucket.
|
| 70 |
+
}
|
| 71 |
+
DEFAULT_BUCKET = "B-whole"
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
# --- loaders ---------------------------------------------------------------
|
| 75 |
+
|
| 76 |
+
def load_captions() -> dict[str, dict]:
|
| 77 |
+
out = {}
|
| 78 |
+
with open(CAPS) as f:
|
| 79 |
+
for line in f:
|
| 80 |
+
if not line.strip():
|
| 81 |
+
continue
|
| 82 |
+
d = json.loads(line)
|
| 83 |
+
out[d["id"]] = d
|
| 84 |
+
return out
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
def load_canonical() -> dict[str, dict]:
|
| 88 |
+
out = {}
|
| 89 |
+
with open(CANON) as f:
|
| 90 |
+
for line in f:
|
| 91 |
+
if not line.strip():
|
| 92 |
+
continue
|
| 93 |
+
d = json.loads(line)
|
| 94 |
+
out[d["id"]] = d
|
| 95 |
+
return out
|
| 96 |
+
|
| 97 |
+
|
| 98 |
+
# --- per-sample serialization ---------------------------------------------
|
| 99 |
+
|
| 100 |
+
def build_sample(cid: str, caps_row: dict, canon_row: dict, latent_path: Path | None
|
| 101 |
+
) -> dict:
|
| 102 |
+
"""Return dict of {filename : bytes} for one tar sample."""
|
| 103 |
+
files: dict[str, bytes] = {}
|
| 104 |
+
|
| 105 |
+
if latent_path is not None and latent_path.exists():
|
| 106 |
+
files[f"{cid}.latent.npy"] = latent_path.read_bytes()
|
| 107 |
+
|
| 108 |
+
files[f"{cid}.captions.json"] = json.dumps(
|
| 109 |
+
caps_row["captions"], separators=(",", ":")).encode("utf-8")
|
| 110 |
+
|
| 111 |
+
# conditioning — union of caps_row.cond and a subset of canonical fields
|
| 112 |
+
cond = dict(caps_row["cond"])
|
| 113 |
+
# add physical size + mask shape for coarse conditioning reconstruction
|
| 114 |
+
mask_meta = canon_row.get("mask", {})
|
| 115 |
+
cond["orig_shape_dhw"] = list(mask_meta.get("shape", []))
|
| 116 |
+
cond["orig_spacing_mm"] = list(mask_meta.get("spacing", []))
|
| 117 |
+
cond["bucket"] = assign_bucket(canon_row)
|
| 118 |
+
files[f"{cid}.cond.json"] = json.dumps(cond, separators=(",", ":")).encode("utf-8")
|
| 119 |
+
|
| 120 |
+
files[f"{cid}.manifest.json"] = json.dumps({
|
| 121 |
+
"id": cid,
|
| 122 |
+
"bucket": cond["bucket"],
|
| 123 |
+
"has_latent": f"{cid}.latent.npy" in files,
|
| 124 |
+
}, separators=(",", ":")).encode("utf-8")
|
| 125 |
+
|
| 126 |
+
return files
|
| 127 |
+
|
| 128 |
+
|
| 129 |
+
def assign_bucket(canon_row: dict) -> str:
|
| 130 |
+
fov = canon_row.get("canonical", {}).get("fov") or canon_row.get("mask", {}).get("fov")
|
| 131 |
+
return FOV_TO_BUCKET.get(fov, DEFAULT_BUCKET)
|
| 132 |
+
|
| 133 |
+
|
| 134 |
+
# --- packing ----------------------------------------------------------------
|
| 135 |
+
|
| 136 |
+
def deterministic_shuffle(ids: list[str], split: str) -> list[str]:
|
| 137 |
+
"""Shuffle deterministically by hashing id||split."""
|
| 138 |
+
def key(i):
|
| 139 |
+
return hashlib.sha1(f"{split}::{i}".encode()).digest()
|
| 140 |
+
return sorted(ids, key=key)
|
| 141 |
+
|
| 142 |
+
|
| 143 |
+
def write_shard(sample_batch: list[tuple[str, dict]], out_path: Path) -> int:
|
| 144 |
+
"""Write a tar shard. Returns sample count."""
|
| 145 |
+
out_path.parent.mkdir(parents=True, exist_ok=True)
|
| 146 |
+
with tarfile.open(out_path, "w") as tf:
|
| 147 |
+
for cid, files in sample_batch:
|
| 148 |
+
for fname, data in files.items():
|
| 149 |
+
info = tarfile.TarInfo(name=fname)
|
| 150 |
+
info.size = len(data)
|
| 151 |
+
tf.addfile(info, io.BytesIO(data))
|
| 152 |
+
return len(sample_batch)
|
| 153 |
+
|
| 154 |
+
|
| 155 |
+
def main():
|
| 156 |
+
ap = argparse.ArgumentParser()
|
| 157 |
+
ap.add_argument("--split", choices=["train", "val", "test"], required=True)
|
| 158 |
+
ap.add_argument("--out", default=str(DEFAULT_OUT))
|
| 159 |
+
ap.add_argument("--latent-cache", default=os.environ.get("LATENT_CACHE", ""),
|
| 160 |
+
help="Dir with {id}.latent.npy files. If empty or absent, latents are skipped.")
|
| 161 |
+
ap.add_argument("--shard-size", type=int, default=SHARD_SIZE)
|
| 162 |
+
ap.add_argument("--dry-count", type=int, default=0,
|
| 163 |
+
help="If > 0, only pack this many samples (scaffold test).")
|
| 164 |
+
args = ap.parse_args()
|
| 165 |
+
|
| 166 |
+
out_dir = Path(args.out)
|
| 167 |
+
splits = json.loads(SPLITS.read_text())
|
| 168 |
+
ids = splits[args.split]
|
| 169 |
+
|
| 170 |
+
caps = load_captions()
|
| 171 |
+
canon = load_canonical()
|
| 172 |
+
|
| 173 |
+
# drop IDs missing canonical or caption
|
| 174 |
+
missing = [i for i in ids if i not in caps or i not in canon]
|
| 175 |
+
if missing:
|
| 176 |
+
print(f"[warn] {len(missing)} ids in {args.split} are missing caps/canonical; skipping")
|
| 177 |
+
ids = [i for i in ids if i in caps and i in canon]
|
| 178 |
+
|
| 179 |
+
# group by bucket
|
| 180 |
+
by_bucket: dict[str, list[str]] = {}
|
| 181 |
+
for cid in ids:
|
| 182 |
+
b = assign_bucket(canon[cid])
|
| 183 |
+
by_bucket.setdefault(b, []).append(cid)
|
| 184 |
+
|
| 185 |
+
latent_dir = Path(args.latent_cache) if args.latent_cache else None
|
| 186 |
+
|
| 187 |
+
index = {} # id → {bucket, shard, pos}
|
| 188 |
+
total_samples = 0
|
| 189 |
+
total_shards = 0
|
| 190 |
+
for bucket, bids in by_bucket.items():
|
| 191 |
+
bids = deterministic_shuffle(bids, args.split)
|
| 192 |
+
if args.dry_count > 0:
|
| 193 |
+
bids = bids[: args.dry_count]
|
| 194 |
+
print(f"[{args.split}] bucket {bucket}: {len(bids)} cases → "
|
| 195 |
+
f"{(len(bids) + args.shard_size - 1) // args.shard_size} shards")
|
| 196 |
+
shard_idx = 0
|
| 197 |
+
batch: list[tuple[str, dict]] = []
|
| 198 |
+
for pos, cid in enumerate(bids):
|
| 199 |
+
latent_path = (latent_dir / f"{cid}.latent.npy") if latent_dir else None
|
| 200 |
+
files = build_sample(cid, caps[cid], canon[cid], latent_path)
|
| 201 |
+
batch.append((cid, files))
|
| 202 |
+
index[cid] = {"bucket": bucket, "shard": shard_idx, "pos": pos % args.shard_size}
|
| 203 |
+
if len(batch) >= args.shard_size:
|
| 204 |
+
tar_name = f"{args.split}-{bucket}-{shard_idx:05d}.tar"
|
| 205 |
+
n = write_shard(batch, out_dir / tar_name)
|
| 206 |
+
total_samples += n; total_shards += 1
|
| 207 |
+
shard_idx += 1
|
| 208 |
+
batch.clear()
|
| 209 |
+
if batch:
|
| 210 |
+
tar_name = f"{args.split}-{bucket}-{shard_idx:05d}.tar"
|
| 211 |
+
n = write_shard(batch, out_dir / tar_name)
|
| 212 |
+
total_samples += n; total_shards += 1
|
| 213 |
+
|
| 214 |
+
idx_path = out_dir / f"{args.split}_index.json"
|
| 215 |
+
idx_path.write_text(json.dumps({
|
| 216 |
+
"split": args.split,
|
| 217 |
+
"n_samples": total_samples,
|
| 218 |
+
"n_shards": total_shards,
|
| 219 |
+
"shard_size": args.shard_size,
|
| 220 |
+
"buckets": {b: len(v) for b, v in by_bucket.items()},
|
| 221 |
+
"has_latents": latent_dir is not None,
|
| 222 |
+
"index": index,
|
| 223 |
+
}, separators=(",", ":")))
|
| 224 |
+
print(f"[{args.split}] wrote {total_shards} shards ({total_samples} samples) → {out_dir}")
|
| 225 |
+
print(f"[{args.split}] index: {idx_path}")
|
| 226 |
+
|
| 227 |
+
|
| 228 |
+
if __name__ == "__main__":
|
| 229 |
+
main()
|
raw_pants_train_test/metadata/pants-captions-ldm/code/phase_audit.py
ADDED
|
@@ -0,0 +1,235 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
P9: 4-way phase audit (v4 §14 item 2).
|
| 3 |
+
|
| 4 |
+
For each of the 9,901 PanTS cases, cross-compares CT phase from multiple
|
| 5 |
+
sources and flags disagreements.
|
| 6 |
+
|
| 7 |
+
Sources used:
|
| 8 |
+
(1) canonical — canonical_facts.canonical.phase (LLM-parsed from report)
|
| 9 |
+
(2) hu — heuristic from liver + spleen mean HU (report organ_status)
|
| 10 |
+
Non-contrast: liver_hu < 60 AND spleen_hu < 55
|
| 11 |
+
Arterial : 60 <= liver_hu < 80 AND spleen_hu > 70 (heterogeneous bright spleen)
|
| 12 |
+
Venous : liver_hu >= 85
|
| 13 |
+
Delay : 65 <= liver_hu < 85 AND abs(liver-spleen) < 15
|
| 14 |
+
Otherwise : unknown
|
| 15 |
+
(3) caption — regex on V1_long_narrative / V5_layered_findings for
|
| 16 |
+
phase mentions (non-contrast / arterial / venous / delayed)
|
| 17 |
+
(4) report_text — regex on canonical_facts.report.raw_report
|
| 18 |
+
|
| 19 |
+
DICOM metadata is not distributed with PanTS — that source is not used.
|
| 20 |
+
|
| 21 |
+
Output (phase_audit.jsonl):
|
| 22 |
+
{ id, phase_canonical, phase_hu, phase_caption, phase_report,
|
| 23 |
+
agreement_k (int in 0..4), phase_agreed (majority or None),
|
| 24 |
+
disagreement_flag (bool) }
|
| 25 |
+
|
| 26 |
+
Summary report printed to stdout and also dropped as phase_audit_summary.md.
|
| 27 |
+
"""
|
| 28 |
+
from __future__ import annotations
|
| 29 |
+
import json
|
| 30 |
+
import re
|
| 31 |
+
from collections import Counter
|
| 32 |
+
from pathlib import Path
|
| 33 |
+
|
| 34 |
+
ROOT = Path(__file__).resolve().parents[1]
|
| 35 |
+
CANON_PATH = ROOT / "canonical_out" / "canonical_facts.jsonl"
|
| 36 |
+
CAPS_PATH = ROOT / "captions_out" / "captions_final.jsonl"
|
| 37 |
+
OUT_JSONL = ROOT / "ldm_training" / "phase_audit.jsonl"
|
| 38 |
+
OUT_SUMMARY = ROOT / "ldm_training" / "phase_audit_summary.md"
|
| 39 |
+
|
| 40 |
+
PHASES = ("Non-contrast", "Arterial", "Venous", "Delay")
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
# --- text patterns ----------------------------------------------------------
|
| 44 |
+
|
| 45 |
+
_RX_NONCON = re.compile(
|
| 46 |
+
r"\b(non[-\s]?contrast|unenhanced|without\s+contrast|no\s+iv\s+contrast|plain\s+ct)\b",
|
| 47 |
+
re.IGNORECASE,
|
| 48 |
+
)
|
| 49 |
+
_RX_ART = re.compile(r"\b(arterial\s+phase|late\s+arterial|hepatic\s+arterial)\b", re.IGNORECASE)
|
| 50 |
+
_RX_VEN = re.compile(
|
| 51 |
+
r"\b(venous\s+phase|portal\s+venous|portovenous|portal[-\s]?phase)\b",
|
| 52 |
+
re.IGNORECASE,
|
| 53 |
+
)
|
| 54 |
+
_RX_DEL = re.compile(r"\b(delay(ed)?\s+phase|equilibrium\s+phase)\b", re.IGNORECASE)
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
def _phase_from_text(text: str) -> str:
|
| 58 |
+
if not text:
|
| 59 |
+
return "unknown"
|
| 60 |
+
# precedence: more specific terms first. check all and if multiple, prefer the
|
| 61 |
+
# one closest to the start (captions usually open with the phase).
|
| 62 |
+
hits = []
|
| 63 |
+
for rx, label in ((_RX_NONCON, "Non-contrast"),
|
| 64 |
+
(_RX_ART, "Arterial"),
|
| 65 |
+
(_RX_VEN, "Venous"),
|
| 66 |
+
(_RX_DEL, "Delay")):
|
| 67 |
+
m = rx.search(text)
|
| 68 |
+
if m:
|
| 69 |
+
hits.append((m.start(), label))
|
| 70 |
+
if not hits:
|
| 71 |
+
return "unknown"
|
| 72 |
+
hits.sort()
|
| 73 |
+
return hits[0][1]
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
# --- HU heuristic -----------------------------------------------------------
|
| 77 |
+
|
| 78 |
+
def _phase_from_hu(liver_hu: float | None, spleen_hu: float | None) -> str:
|
| 79 |
+
if liver_hu is None or spleen_hu is None:
|
| 80 |
+
return "unknown"
|
| 81 |
+
# Typical post-injection enhancement curves:
|
| 82 |
+
# non-contrast: liver ~55-65, spleen ~45-55
|
| 83 |
+
# arterial: liver ~65-80 (modest uptake via hepatic artery), spleen heterogeneously bright 80-120
|
| 84 |
+
# venous: liver 85-110 (portal phase peaks), spleen homogeneous ~100-130
|
| 85 |
+
# delay: liver + spleen equilibrate back ~65-90, close to each other
|
| 86 |
+
if liver_hu >= 85:
|
| 87 |
+
return "Venous"
|
| 88 |
+
if liver_hu < 60 and spleen_hu < 60:
|
| 89 |
+
return "Non-contrast"
|
| 90 |
+
if 60 <= liver_hu < 80 and spleen_hu >= 70:
|
| 91 |
+
return "Arterial"
|
| 92 |
+
if 65 <= liver_hu < 85 and abs(liver_hu - spleen_hu) < 15:
|
| 93 |
+
return "Delay"
|
| 94 |
+
return "unknown"
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
# --- main -------------------------------------------------------------------
|
| 98 |
+
|
| 99 |
+
def main():
|
| 100 |
+
# load canonical facts into dict-by-id
|
| 101 |
+
canon = {}
|
| 102 |
+
with open(CANON_PATH) as f:
|
| 103 |
+
for line in f:
|
| 104 |
+
if not line.strip():
|
| 105 |
+
continue
|
| 106 |
+
d = json.loads(line)
|
| 107 |
+
canon[d["id"]] = d
|
| 108 |
+
|
| 109 |
+
# load captions
|
| 110 |
+
caps = {}
|
| 111 |
+
with open(CAPS_PATH) as f:
|
| 112 |
+
for line in f:
|
| 113 |
+
if not line.strip():
|
| 114 |
+
continue
|
| 115 |
+
d = json.loads(line)
|
| 116 |
+
caps[d["id"]] = d
|
| 117 |
+
|
| 118 |
+
ids = sorted(set(canon) & set(caps))
|
| 119 |
+
print(f"audit: {len(ids)} cases have both canonical + captions "
|
| 120 |
+
f"(canon={len(canon)}, caps={len(caps)})")
|
| 121 |
+
|
| 122 |
+
rows = []
|
| 123 |
+
# counters
|
| 124 |
+
n_sources = Counter() # agreement_k histogram
|
| 125 |
+
n_canon = Counter()
|
| 126 |
+
n_hu = Counter()
|
| 127 |
+
n_cap = Counter()
|
| 128 |
+
n_rep = Counter()
|
| 129 |
+
n_agreed = Counter()
|
| 130 |
+
confusion = Counter() # (canon, hu)
|
| 131 |
+
|
| 132 |
+
for cid in ids:
|
| 133 |
+
c = canon[cid]
|
| 134 |
+
cap = caps[cid]
|
| 135 |
+
phase_can = c["canonical"].get("phase") or "unknown"
|
| 136 |
+
if phase_can == "MISSING":
|
| 137 |
+
phase_can = "unknown"
|
| 138 |
+
|
| 139 |
+
osr = c["canonical"].get("organ_status_report") or {}
|
| 140 |
+
liver_hu = osr.get("liver", {}).get("hu_mean")
|
| 141 |
+
spleen_hu = osr.get("spleen", {}).get("hu_mean")
|
| 142 |
+
phase_hu = _phase_from_hu(liver_hu, spleen_hu)
|
| 143 |
+
|
| 144 |
+
v1 = cap["captions"].get("V1_long_narrative", "") or ""
|
| 145 |
+
v5 = cap["captions"].get("V5_layered_findings", "") or ""
|
| 146 |
+
phase_cap = _phase_from_text(v1 + " \n " + v5)
|
| 147 |
+
|
| 148 |
+
report_text = c.get("report", {}).get("raw_report", "") or ""
|
| 149 |
+
phase_rep = _phase_from_text(report_text)
|
| 150 |
+
|
| 151 |
+
votes = [p for p in (phase_can, phase_hu, phase_cap, phase_rep)
|
| 152 |
+
if p != "unknown"]
|
| 153 |
+
if votes:
|
| 154 |
+
tally = Counter(votes)
|
| 155 |
+
top_label, top_count = tally.most_common(1)[0]
|
| 156 |
+
# require strict majority
|
| 157 |
+
majority = top_label if top_count > (len(votes) / 2) else None
|
| 158 |
+
else:
|
| 159 |
+
top_count = 0
|
| 160 |
+
majority = None
|
| 161 |
+
agreement_k = max(Counter(votes).values()) if votes else 0
|
| 162 |
+
disagree = (
|
| 163 |
+
phase_can != "unknown"
|
| 164 |
+
and phase_hu != "unknown"
|
| 165 |
+
and phase_can != phase_hu
|
| 166 |
+
)
|
| 167 |
+
|
| 168 |
+
rows.append({
|
| 169 |
+
"id": cid,
|
| 170 |
+
"phase_canonical": phase_can,
|
| 171 |
+
"phase_hu": phase_hu,
|
| 172 |
+
"phase_caption": phase_cap,
|
| 173 |
+
"phase_report": phase_rep,
|
| 174 |
+
"liver_hu": liver_hu,
|
| 175 |
+
"spleen_hu": spleen_hu,
|
| 176 |
+
"agreement_k": agreement_k,
|
| 177 |
+
"phase_agreed": majority,
|
| 178 |
+
"disagreement_canon_vs_hu": disagree,
|
| 179 |
+
})
|
| 180 |
+
|
| 181 |
+
n_sources[agreement_k] += 1
|
| 182 |
+
n_canon[phase_can] += 1
|
| 183 |
+
n_hu[phase_hu] += 1
|
| 184 |
+
n_cap[phase_cap] += 1
|
| 185 |
+
n_rep[phase_rep] += 1
|
| 186 |
+
n_agreed[majority or "none"] += 1
|
| 187 |
+
confusion[(phase_can, phase_hu)] += 1
|
| 188 |
+
|
| 189 |
+
# write jsonl
|
| 190 |
+
OUT_JSONL.parent.mkdir(parents=True, exist_ok=True)
|
| 191 |
+
with open(OUT_JSONL, "w") as f:
|
| 192 |
+
for r in rows:
|
| 193 |
+
f.write(json.dumps(r) + "\n")
|
| 194 |
+
|
| 195 |
+
# summary
|
| 196 |
+
lines = []
|
| 197 |
+
def pl(s=""):
|
| 198 |
+
lines.append(s)
|
| 199 |
+
print(s)
|
| 200 |
+
|
| 201 |
+
pl(f"# Phase audit summary (N={len(rows)})")
|
| 202 |
+
pl("")
|
| 203 |
+
pl("## Phase distribution by source")
|
| 204 |
+
pl(f"| phase | canonical | hu_heur | caption | report |")
|
| 205 |
+
pl(f"|---------------|-----------|---------|---------|--------|")
|
| 206 |
+
all_phases = sorted(set(PHASES) | set(n_canon) | set(n_hu) | set(n_cap) | set(n_rep))
|
| 207 |
+
for p in all_phases:
|
| 208 |
+
pl(f"| {p:<13} | {n_canon.get(p,0):>9} | {n_hu.get(p,0):>7} | "
|
| 209 |
+
f"{n_cap.get(p,0):>7} | {n_rep.get(p,0):>6} |")
|
| 210 |
+
pl("")
|
| 211 |
+
pl("## Agreement across sources (k = max votes for any single label)")
|
| 212 |
+
for k in sorted(n_sources):
|
| 213 |
+
pl(f" k={k}: {n_sources[k]} ({n_sources[k]/len(rows)*100:.1f}%)")
|
| 214 |
+
pl("")
|
| 215 |
+
pl("## Majority-vote agreed phase")
|
| 216 |
+
for p, n in n_agreed.most_common():
|
| 217 |
+
pl(f" {p:<13}: {n} ({n/len(rows)*100:.1f}%)")
|
| 218 |
+
pl("")
|
| 219 |
+
pl("## canonical vs HU-heuristic confusion (top 15)")
|
| 220 |
+
pl(" canonical \\ hu count")
|
| 221 |
+
for (cp, hp), n in confusion.most_common(15):
|
| 222 |
+
pl(f" {cp:<13} {hp:<13} {n}")
|
| 223 |
+
pl("")
|
| 224 |
+
disagree_n = sum(1 for r in rows if r["disagreement_canon_vs_hu"])
|
| 225 |
+
pl(f"canonical-vs-hu strong disagreement (both non-unknown, labels differ): "
|
| 226 |
+
f"{disagree_n} ({disagree_n/len(rows)*100:.1f}%)")
|
| 227 |
+
|
| 228 |
+
OUT_SUMMARY.write_text("\n".join(lines) + "\n")
|
| 229 |
+
print()
|
| 230 |
+
print(f"wrote {OUT_JSONL}")
|
| 231 |
+
print(f"wrote {OUT_SUMMARY}")
|
| 232 |
+
|
| 233 |
+
|
| 234 |
+
if __name__ == "__main__":
|
| 235 |
+
main()
|
raw_pants_train_test/metadata/pants-captions-ldm/code/preprocessing.py
ADDED
|
@@ -0,0 +1,268 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
P4: PanTS preprocessing pipeline (per LDM_TRAINING_PLAN_v4 §3-§4).
|
| 3 |
+
|
| 4 |
+
Canonical order (whole-volume path):
|
| 5 |
+
1. load NIfTI (x, y, z) → reorder to (z, y, x) = (D, H, W) [RAS+ convention]
|
| 6 |
+
2. clip HU to [-1000, +1000]
|
| 7 |
+
3. normalize to [-1, 1]
|
| 8 |
+
4. resample to target voxel spacing (bilinear)
|
| 9 |
+
5. pad/crop to bucket shape (D, H, W)
|
| 10 |
+
6. record csize (original shape, mm), ccrop (crop origin, vox & mm), car (target)
|
| 11 |
+
|
| 12 |
+
Pan-crop path (fix for v4 blocker: order matters):
|
| 13 |
+
1. load CT + label masks
|
| 14 |
+
2. reorder both to (D, H, W)
|
| 15 |
+
3. clip+normalize CT
|
| 16 |
+
4. resample CT + mask to target spacing (bilinear for CT, NN for mask)
|
| 17 |
+
5. compute bbox from resampled mask (pancreas class union {17,18,19,20})
|
| 18 |
+
with 40 mm padding
|
| 19 |
+
6. crop CT to bbox
|
| 20 |
+
7. pad-or-crop to bucket shape
|
| 21 |
+
|
| 22 |
+
The KEY v4 fix is ordering step 5 (bbox from mask) AFTER step 4 (resample),
|
| 23 |
+
not before. If we bbox first on native spacing, then resample the cropped
|
| 24 |
+
volume, the final crop origin can mis-locate the pancreas when native
|
| 25 |
+
spacing is anisotropic.
|
| 26 |
+
"""
|
| 27 |
+
from __future__ import annotations
|
| 28 |
+
import json
|
| 29 |
+
import math
|
| 30 |
+
from dataclasses import dataclass
|
| 31 |
+
from pathlib import Path
|
| 32 |
+
from typing import Sequence
|
| 33 |
+
|
| 34 |
+
import numpy as np
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
# RadGPT pancreas class map (canonical_facts.py convention)
|
| 38 |
+
PANC_CLASSES = (17, 18, 19, 20)
|
| 39 |
+
|
| 40 |
+
# HU clip bounds (GenerateCT / Report2CT convention)
|
| 41 |
+
HU_MIN = -1000.0
|
| 42 |
+
HU_MAX = 1000.0
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
@dataclass
|
| 46 |
+
class PreprocessOutput:
|
| 47 |
+
volume: np.ndarray # (D, H, W) float32 in [-1, 1]
|
| 48 |
+
csize_mm: tuple[float, float, float] # original physical size in mm (D, H, W)
|
| 49 |
+
csize_vox: tuple[int, int, int] # original voxel shape
|
| 50 |
+
ccrop_vox: tuple[int, int, int] # crop origin in resampled grid
|
| 51 |
+
ccrop_mm: tuple[float, float, float] # crop origin in mm
|
| 52 |
+
car_vox: tuple[int, int, int] # target bucket shape
|
| 53 |
+
voxel_spacing_mm: tuple[float, float, float]
|
| 54 |
+
bucket: str
|
| 55 |
+
|
| 56 |
+
def to_cond(self) -> dict:
|
| 57 |
+
return {
|
| 58 |
+
"csize_mm": list(self.csize_mm),
|
| 59 |
+
"csize_vox": list(self.csize_vox),
|
| 60 |
+
"ccrop_vox": list(self.ccrop_vox),
|
| 61 |
+
"ccrop_mm": list(self.ccrop_mm),
|
| 62 |
+
"car_vox": list(self.car_vox),
|
| 63 |
+
"voxel_spacing_mm": list(self.voxel_spacing_mm),
|
| 64 |
+
"bucket": self.bucket,
|
| 65 |
+
}
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
# --- primitives -------------------------------------------------------------
|
| 69 |
+
|
| 70 |
+
def clip_and_normalize(ct: np.ndarray) -> np.ndarray:
|
| 71 |
+
"""[HU_MIN, HU_MAX] → [-1, 1]."""
|
| 72 |
+
x = np.clip(ct.astype(np.float32), HU_MIN, HU_MAX)
|
| 73 |
+
return (x - HU_MIN) / (HU_MAX - HU_MIN) * 2 - 1
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
def resample_linear(vol: np.ndarray, src_spacing, dst_spacing) -> np.ndarray:
|
| 77 |
+
"""Trilinear resample along (D, H, W). Lightweight: scipy.ndimage.zoom."""
|
| 78 |
+
from scipy.ndimage import zoom
|
| 79 |
+
src = np.array(src_spacing, dtype=np.float64)
|
| 80 |
+
dst = np.array(dst_spacing, dtype=np.float64)
|
| 81 |
+
factor = tuple(src / dst)
|
| 82 |
+
# zoom with order=1 = trilinear
|
| 83 |
+
out = zoom(vol, factor, order=1, mode="constant", cval=-1.0, prefilter=False)
|
| 84 |
+
return out.astype(np.float32)
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
def resample_nn(mask: np.ndarray, src_spacing, dst_spacing) -> np.ndarray:
|
| 88 |
+
from scipy.ndimage import zoom
|
| 89 |
+
src = np.array(src_spacing, dtype=np.float64)
|
| 90 |
+
dst = np.array(dst_spacing, dtype=np.float64)
|
| 91 |
+
factor = tuple(src / dst)
|
| 92 |
+
out = zoom(mask, factor, order=0, mode="constant", cval=0, prefilter=False)
|
| 93 |
+
return out.astype(mask.dtype)
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
def pad_or_crop_to(vol: np.ndarray, target: Sequence[int],
|
| 97 |
+
pad_value: float = -1.0,
|
| 98 |
+
crop_origin: Sequence[int] | None = None):
|
| 99 |
+
"""Pad (with pad_value) or center-crop along each axis.
|
| 100 |
+
Returns (cropped, origin) where origin is the crop start in the input frame.
|
| 101 |
+
If crop_origin is given (per-axis start voxel), use it instead of center-crop.
|
| 102 |
+
"""
|
| 103 |
+
D, H, W = vol.shape
|
| 104 |
+
tD, tH, tW = target
|
| 105 |
+
origin = [0, 0, 0]
|
| 106 |
+
# D
|
| 107 |
+
if D >= tD:
|
| 108 |
+
start = ((D - tD) // 2) if crop_origin is None else int(crop_origin[0])
|
| 109 |
+
start = max(0, min(start, D - tD))
|
| 110 |
+
vol = vol[start:start + tD]
|
| 111 |
+
origin[0] = start
|
| 112 |
+
else:
|
| 113 |
+
pad_before = (tD - D) // 2
|
| 114 |
+
pad_after = tD - D - pad_before
|
| 115 |
+
vol = np.pad(vol, ((pad_before, pad_after), (0, 0), (0, 0)),
|
| 116 |
+
mode="constant", constant_values=pad_value)
|
| 117 |
+
origin[0] = -pad_before
|
| 118 |
+
# H
|
| 119 |
+
D_, H_, W_ = vol.shape
|
| 120 |
+
if H_ >= tH:
|
| 121 |
+
start = ((H_ - tH) // 2) if crop_origin is None else int(crop_origin[1])
|
| 122 |
+
start = max(0, min(start, H_ - tH))
|
| 123 |
+
vol = vol[:, start:start + tH]
|
| 124 |
+
origin[1] = start
|
| 125 |
+
else:
|
| 126 |
+
pad_before = (tH - H_) // 2
|
| 127 |
+
pad_after = tH - H_ - pad_before
|
| 128 |
+
vol = np.pad(vol, ((0, 0), (pad_before, pad_after), (0, 0)),
|
| 129 |
+
mode="constant", constant_values=pad_value)
|
| 130 |
+
origin[1] = -pad_before
|
| 131 |
+
# W
|
| 132 |
+
D_, H_, W_ = vol.shape
|
| 133 |
+
if W_ >= tW:
|
| 134 |
+
start = ((W_ - tW) // 2) if crop_origin is None else int(crop_origin[2])
|
| 135 |
+
start = max(0, min(start, W_ - tW))
|
| 136 |
+
vol = vol[:, :, start:start + tW]
|
| 137 |
+
origin[2] = start
|
| 138 |
+
else:
|
| 139 |
+
pad_before = (tW - W_) // 2
|
| 140 |
+
pad_after = tW - W_ - pad_before
|
| 141 |
+
vol = np.pad(vol, ((0, 0), (0, 0), (pad_before, pad_after)),
|
| 142 |
+
mode="constant", constant_values=pad_value)
|
| 143 |
+
origin[2] = -pad_before
|
| 144 |
+
return vol, tuple(origin)
|
| 145 |
+
|
| 146 |
+
|
| 147 |
+
def bbox_from_mask(mask: np.ndarray, pad_vox: Sequence[int]) -> tuple | None:
|
| 148 |
+
"""Return (d0, d1, h0, h1, w0, w1) bbox with per-axis padding, or None if mask empty."""
|
| 149 |
+
nz = np.argwhere(mask > 0)
|
| 150 |
+
if len(nz) == 0:
|
| 151 |
+
return None
|
| 152 |
+
d0, h0, w0 = nz.min(axis=0)
|
| 153 |
+
d1, h1, w1 = nz.max(axis=0) + 1
|
| 154 |
+
d0 = max(0, d0 - pad_vox[0]); d1 = min(mask.shape[0], d1 + pad_vox[0])
|
| 155 |
+
h0 = max(0, h0 - pad_vox[1]); h1 = min(mask.shape[1], h1 + pad_vox[1])
|
| 156 |
+
w0 = max(0, w0 - pad_vox[2]); w1 = min(mask.shape[2], w1 + pad_vox[2])
|
| 157 |
+
return (int(d0), int(d1), int(h0), int(h1), int(w0), int(w1))
|
| 158 |
+
|
| 159 |
+
|
| 160 |
+
# --- public API -------------------------------------------------------------
|
| 161 |
+
|
| 162 |
+
def preprocess_whole(
|
| 163 |
+
ct_dhw: np.ndarray,
|
| 164 |
+
src_spacing_dhw: Sequence[float],
|
| 165 |
+
bucket: dict,
|
| 166 |
+
) -> PreprocessOutput:
|
| 167 |
+
"""Preprocess a whole-volume case into bucket shape.
|
| 168 |
+
ct_dhw: (D, H, W) in HU, already reordered to RAS+ axis convention.
|
| 169 |
+
"""
|
| 170 |
+
orig_D, orig_H, orig_W = ct_dhw.shape
|
| 171 |
+
orig_mm = (orig_D * src_spacing_dhw[0],
|
| 172 |
+
orig_H * src_spacing_dhw[1],
|
| 173 |
+
orig_W * src_spacing_dhw[2])
|
| 174 |
+
|
| 175 |
+
# step 2-3
|
| 176 |
+
ct_norm = clip_and_normalize(ct_dhw)
|
| 177 |
+
# step 4
|
| 178 |
+
dst_spacing = bucket["voxel_spacing_mm"]
|
| 179 |
+
ct_resampled = resample_linear(ct_norm, src_spacing_dhw, dst_spacing)
|
| 180 |
+
# step 5
|
| 181 |
+
ct_final, crop_origin = pad_or_crop_to(ct_resampled, bucket["shape_dhw"], pad_value=-1.0)
|
| 182 |
+
crop_mm = tuple(float(crop_origin[i]) * dst_spacing[i] for i in range(3))
|
| 183 |
+
|
| 184 |
+
return PreprocessOutput(
|
| 185 |
+
volume=ct_final,
|
| 186 |
+
csize_mm=orig_mm,
|
| 187 |
+
csize_vox=(orig_D, orig_H, orig_W),
|
| 188 |
+
ccrop_vox=tuple(crop_origin),
|
| 189 |
+
ccrop_mm=crop_mm,
|
| 190 |
+
car_vox=tuple(bucket["shape_dhw"]),
|
| 191 |
+
voxel_spacing_mm=tuple(dst_spacing),
|
| 192 |
+
bucket=bucket["name"],
|
| 193 |
+
)
|
| 194 |
+
|
| 195 |
+
|
| 196 |
+
def preprocess_pan_crop(
|
| 197 |
+
ct_dhw: np.ndarray,
|
| 198 |
+
label_dhw: np.ndarray,
|
| 199 |
+
src_spacing_dhw: Sequence[float],
|
| 200 |
+
bucket: dict,
|
| 201 |
+
pad_mm: float = 40.0,
|
| 202 |
+
) -> PreprocessOutput | None:
|
| 203 |
+
"""Preprocess a pancreas-centered crop.
|
| 204 |
+
Returns None if pancreas mask is empty.
|
| 205 |
+
|
| 206 |
+
ORDER MATTERS (v4 blocker fix):
|
| 207 |
+
resample first, THEN bbox, THEN crop, THEN pad-or-crop to bucket.
|
| 208 |
+
"""
|
| 209 |
+
# bail if mask empty at native resolution
|
| 210 |
+
pancreas_mask = np.isin(label_dhw, PANC_CLASSES).astype(np.uint8)
|
| 211 |
+
if pancreas_mask.sum() == 0:
|
| 212 |
+
return None
|
| 213 |
+
|
| 214 |
+
orig_shape = ct_dhw.shape
|
| 215 |
+
orig_mm = (orig_shape[0] * src_spacing_dhw[0],
|
| 216 |
+
orig_shape[1] * src_spacing_dhw[1],
|
| 217 |
+
orig_shape[2] * src_spacing_dhw[2])
|
| 218 |
+
|
| 219 |
+
ct_norm = clip_and_normalize(ct_dhw)
|
| 220 |
+
dst_spacing = bucket["voxel_spacing_mm"]
|
| 221 |
+
ct_res = resample_linear(ct_norm, src_spacing_dhw, dst_spacing)
|
| 222 |
+
mask_res = resample_nn(pancreas_mask, src_spacing_dhw, dst_spacing)
|
| 223 |
+
|
| 224 |
+
if mask_res.sum() == 0:
|
| 225 |
+
return None
|
| 226 |
+
|
| 227 |
+
pad_vox = [max(1, int(round(pad_mm / dst_spacing[i]))) for i in range(3)]
|
| 228 |
+
bbox = bbox_from_mask(mask_res, pad_vox)
|
| 229 |
+
if bbox is None:
|
| 230 |
+
return None
|
| 231 |
+
d0, d1, h0, h1, w0, w1 = bbox
|
| 232 |
+
ct_crop = ct_res[d0:d1, h0:h1, w0:w1]
|
| 233 |
+
|
| 234 |
+
# step 7: pad-or-crop to bucket shape
|
| 235 |
+
ct_final, crop_origin_inner = pad_or_crop_to(ct_crop, bucket["shape_dhw"], pad_value=-1.0)
|
| 236 |
+
|
| 237 |
+
# absolute crop origin in resampled frame = bbox top-left + inner_crop
|
| 238 |
+
abs_origin = (d0 + crop_origin_inner[0], h0 + crop_origin_inner[1], w0 + crop_origin_inner[2])
|
| 239 |
+
crop_mm = tuple(float(abs_origin[i]) * dst_spacing[i] for i in range(3))
|
| 240 |
+
|
| 241 |
+
return PreprocessOutput(
|
| 242 |
+
volume=ct_final,
|
| 243 |
+
csize_mm=orig_mm,
|
| 244 |
+
csize_vox=tuple(orig_shape),
|
| 245 |
+
ccrop_vox=abs_origin,
|
| 246 |
+
ccrop_mm=crop_mm,
|
| 247 |
+
car_vox=tuple(bucket["shape_dhw"]),
|
| 248 |
+
voxel_spacing_mm=tuple(dst_spacing),
|
| 249 |
+
bucket=bucket["name"],
|
| 250 |
+
)
|
| 251 |
+
|
| 252 |
+
|
| 253 |
+
# --- NIfTI I/O helpers ------------------------------------------------------
|
| 254 |
+
|
| 255 |
+
def load_nifti_ras(path: str | Path) -> tuple[np.ndarray, tuple[float, float, float]]:
|
| 256 |
+
"""Load a NIfTI and return (DHW float32 array, (sz, sy, sx) voxel spacing).
|
| 257 |
+
|
| 258 |
+
NIfTI stores arrays in (x, y, z) and spacing in (sx, sy, sz). We
|
| 259 |
+
transpose to (z, y, x) = (D, H, W) and reorder spacing to (sz, sy, sx).
|
| 260 |
+
Does NOT do full orientation canonicalization — for that, use nibabel's
|
| 261 |
+
`as_closest_canonical` before calling this.
|
| 262 |
+
"""
|
| 263 |
+
import nibabel as nib
|
| 264 |
+
img = nib.as_closest_canonical(nib.load(str(path)))
|
| 265 |
+
arr = np.asarray(img.get_fdata(), dtype=np.float32) # (x, y, z)
|
| 266 |
+
sx, sy, sz = img.header.get_zooms()[:3]
|
| 267 |
+
arr = np.transpose(arr, (2, 1, 0)) # (z, y, x)
|
| 268 |
+
return arr, (float(sz), float(sy), float(sx))
|
raw_pants_train_test/metadata/pants-captions-ldm/code/schedule.py
ADDED
|
@@ -0,0 +1,145 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
P6: VP linear β + Zero-Terminal-SNR (Lin 2023, arXiv 2305.08891)
|
| 3 |
+
+ v-prediction + Min-SNR-γ (γ=5) loss weighting.
|
| 4 |
+
|
| 5 |
+
Mathematical reference:
|
| 6 |
+
α_bar = prod(1 - β_t) (cumulative DDPM alpha bar)
|
| 7 |
+
ZTSNR rescale (eq 7 in Lin 2023):
|
| 8 |
+
sqrt_α̂_bar_t = (sqrt_α_bar_t - sqrt_α_bar_T) * sqrt_α_bar_0 / (sqrt_α_bar_0 - sqrt_α_bar_T)
|
| 9 |
+
then α̂_bar = sqrt_α̂_bar ** 2
|
| 10 |
+
→ ensures α̂_bar[T] = 0 exactly, so x_T is pure noise (required for CFG).
|
| 11 |
+
|
| 12 |
+
SNR(t) = α̂_bar_t / (1 - α̂_bar_t)
|
| 13 |
+
|
| 14 |
+
v-prediction target:
|
| 15 |
+
v_t = √α̂_bar_t · ε - √(1 - α̂_bar_t) · x_0
|
| 16 |
+
v-prediction Min-SNR-γ loss weight (Hang et al 2023, arXiv 2303.09556, v-pred variant):
|
| 17 |
+
w(t) = min(SNR(t), γ) / (SNR(t) + 1)
|
| 18 |
+
"""
|
| 19 |
+
from __future__ import annotations
|
| 20 |
+
import math
|
| 21 |
+
import numpy as np
|
| 22 |
+
import torch
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
def linear_beta(T: int, beta_start: float = 1e-4, beta_end: float = 2e-2) -> torch.Tensor:
|
| 26 |
+
return torch.linspace(beta_start, beta_end, T, dtype=torch.float64)
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def alpha_bar_from_beta(betas: torch.Tensor) -> torch.Tensor:
|
| 30 |
+
return torch.cumprod(1.0 - betas, dim=0)
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
def ztsnr_rescale(alpha_bar: torch.Tensor) -> torch.Tensor:
|
| 34 |
+
"""Rescale so that alpha_bar[-1] = 0 exactly (Lin 2023, eq 7 & 8).
|
| 35 |
+
|
| 36 |
+
The algebra is: take sqrt_alpha_bar, subtract the terminal value,
|
| 37 |
+
rescale so the head stays at its original value (we do a linear shift+scale
|
| 38 |
+
in the sqrt domain; squaring gives the new alpha_bar)."""
|
| 39 |
+
sqrt_ab = alpha_bar.sqrt()
|
| 40 |
+
# shift so terminal becomes 0
|
| 41 |
+
shifted = sqrt_ab - sqrt_ab[-1]
|
| 42 |
+
# rescale so head stays at its original value
|
| 43 |
+
scale = sqrt_ab[0] / shifted[0]
|
| 44 |
+
shifted = shifted * scale
|
| 45 |
+
rescaled = shifted ** 2
|
| 46 |
+
# numerical safety
|
| 47 |
+
rescaled = torch.clamp(rescaled, min=0.0)
|
| 48 |
+
rescaled[-1] = 0.0
|
| 49 |
+
return rescaled
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
def snr(alpha_bar: torch.Tensor) -> torch.Tensor:
|
| 53 |
+
return alpha_bar / (1.0 - alpha_bar).clamp(min=1e-12)
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
def v_prediction_target(x0: torch.Tensor, noise: torch.Tensor,
|
| 57 |
+
alpha_bar_t: torch.Tensor) -> torch.Tensor:
|
| 58 |
+
"""v_t = √ᾱ_t · ε - √(1-ᾱ_t) · x_0.
|
| 59 |
+
|
| 60 |
+
alpha_bar_t has shape broadcastable to x0/noise (typically [B,1,1,1,1]).
|
| 61 |
+
"""
|
| 62 |
+
sab = alpha_bar_t.sqrt()
|
| 63 |
+
somb = (1.0 - alpha_bar_t).sqrt()
|
| 64 |
+
return sab * noise - somb * x0
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
def x0_from_v(x_t: torch.Tensor, v_pred: torch.Tensor,
|
| 68 |
+
alpha_bar_t: torch.Tensor) -> torch.Tensor:
|
| 69 |
+
"""Inverse: x_0 = √ᾱ_t · x_t - √(1-ᾱ_t) · v_t."""
|
| 70 |
+
sab = alpha_bar_t.sqrt()
|
| 71 |
+
somb = (1.0 - alpha_bar_t).sqrt()
|
| 72 |
+
return sab * x_t - somb * v_pred
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
def min_snr_gamma_vpred_weight(snr_t: torch.Tensor, gamma: float = 5.0) -> torch.Tensor:
|
| 76 |
+
"""Min-SNR-γ, v-prediction variant: w_v(t) = min(SNR(t), γ) / (SNR(t) + 1).
|
| 77 |
+
|
| 78 |
+
(For ε-prediction it's min(SNR, γ) / SNR; the extra +1 in the denominator
|
| 79 |
+
is the Hang 2023 v-pred adaptation.)
|
| 80 |
+
"""
|
| 81 |
+
return torch.minimum(snr_t, torch.full_like(snr_t, gamma)) / (snr_t + 1.0)
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
# --- helpers ---------------------------------------------------------------
|
| 85 |
+
|
| 86 |
+
def make_schedule(T: int = 1000, beta_start=1e-4, beta_end=2e-2,
|
| 87 |
+
ztsnr: bool = True) -> dict:
|
| 88 |
+
"""Build one canonical schedule. Returns floats in float64, convert as needed."""
|
| 89 |
+
betas = linear_beta(T, beta_start, beta_end)
|
| 90 |
+
ab = alpha_bar_from_beta(betas)
|
| 91 |
+
if ztsnr:
|
| 92 |
+
ab = ztsnr_rescale(ab)
|
| 93 |
+
return {"betas": betas, "alpha_bar": ab, "snr": snr(ab), "T": T}
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
# --- sampling --------------------------------------------------------------
|
| 97 |
+
|
| 98 |
+
def forward_q_sample(x0: torch.Tensor, noise: torch.Tensor,
|
| 99 |
+
alpha_bar_t: torch.Tensor) -> torch.Tensor:
|
| 100 |
+
"""x_t = √ᾱ_t · x_0 + √(1-ᾱ_t) · ε."""
|
| 101 |
+
sab = alpha_bar_t.sqrt()
|
| 102 |
+
somb = (1.0 - alpha_bar_t).sqrt()
|
| 103 |
+
return sab * x0 + somb * noise
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
def loss_vpred_minsnr(
|
| 107 |
+
model_output: torch.Tensor,
|
| 108 |
+
x0: torch.Tensor, noise: torch.Tensor,
|
| 109 |
+
alpha_bar_t: torch.Tensor, gamma: float = 5.0,
|
| 110 |
+
) -> torch.Tensor:
|
| 111 |
+
v_target = v_prediction_target(x0, noise, alpha_bar_t)
|
| 112 |
+
snr_t = snr(alpha_bar_t).reshape(-1)
|
| 113 |
+
w = min_snr_gamma_vpred_weight(snr_t, gamma=gamma)
|
| 114 |
+
# per-sample MSE then weighted mean
|
| 115 |
+
diff = (model_output - v_target).float()
|
| 116 |
+
per_sample = diff.flatten(1).pow(2).mean(dim=1) # (B,)
|
| 117 |
+
return (w * per_sample).mean()
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
if __name__ == "__main__":
|
| 121 |
+
# sanity
|
| 122 |
+
sch = make_schedule(T=1000, ztsnr=True)
|
| 123 |
+
ab = sch["alpha_bar"]
|
| 124 |
+
assert ab[0].item() > 0.9, f"head ab={ab[0].item()}"
|
| 125 |
+
assert ab[-1].item() == 0.0, f"terminal ab={ab[-1].item()} (should be 0)"
|
| 126 |
+
# monotone decreasing
|
| 127 |
+
d = ab[:-1] - ab[1:]
|
| 128 |
+
assert (d >= -1e-8).all(), "alpha_bar not monotone decreasing"
|
| 129 |
+
# SNR sanity
|
| 130 |
+
s = sch["snr"]
|
| 131 |
+
assert s[0].item() > 10.0
|
| 132 |
+
# terminal SNR is 0 (by ZTSNR); we clamp in snr() so it's 0
|
| 133 |
+
print(f"T=1000 ab[0]={ab[0].item():.6f} ab[-1]={ab[-1].item():.6f}")
|
| 134 |
+
print(f"SNR head={s[0].item():.3f} SNR tail={s[-1].item():.3e}")
|
| 135 |
+
# v-pred identity: at t=0 (ab ≈ 1), v = √ab·ε - √(1-ab)·x0 ≈ ε
|
| 136 |
+
ab_head = ab[0:1].view(1, 1, 1, 1, 1).float()
|
| 137 |
+
x0 = torch.randn(2, 4, 8, 8, 8)
|
| 138 |
+
eps = torch.randn_like(x0)
|
| 139 |
+
v = v_prediction_target(x0, eps, ab_head)
|
| 140 |
+
assert (v - eps).abs().mean() < 0.2, "v should ≈ ε at small t"
|
| 141 |
+
# min-snr weight bounded in [0, 1]
|
| 142 |
+
w = min_snr_gamma_vpred_weight(s, gamma=5.0)
|
| 143 |
+
assert (w >= 0).all() and (w <= 1.0001).all()
|
| 144 |
+
print(f"min-snr weight: head={w[0].item():.4f} tail={w[-1].item():.4f} mean={w.mean().item():.4f}")
|
| 145 |
+
print("schedule tests pass")
|
raw_pants_train_test/metadata/pants-captions-ldm/code/test_preprocessing.py
ADDED
|
@@ -0,0 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""P4 unit tests — synthetic volume round-trip + pan-crop order fix."""
|
| 2 |
+
import os
|
| 3 |
+
import numpy as np
|
| 4 |
+
import sys
|
| 5 |
+
sys.path.insert(0, f"{os.environ.get('REPO_ROOT', '.')}/code")
|
| 6 |
+
from preprocessing import (
|
| 7 |
+
clip_and_normalize, resample_linear, resample_nn,
|
| 8 |
+
pad_or_crop_to, bbox_from_mask,
|
| 9 |
+
preprocess_whole, preprocess_pan_crop,
|
| 10 |
+
)
|
| 11 |
+
|
| 12 |
+
BUCKET_WHOLE = {"name": "B-whole", "shape_dhw": [128, 256, 256],
|
| 13 |
+
"voxel_spacing_mm": [5.0, 2.0, 2.0]}
|
| 14 |
+
BUCKET_PAN = {"name": "B-pan", "shape_dhw": [64, 192, 256],
|
| 15 |
+
"voxel_spacing_mm": [2.0, 1.0, 1.0]}
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
def test_clip_and_normalize():
|
| 19 |
+
x = np.array([-2000, -1000, 0, 1000, 2000], dtype=np.float32)
|
| 20 |
+
y = clip_and_normalize(x)
|
| 21 |
+
assert np.isclose(y[0], -1.0)
|
| 22 |
+
assert np.isclose(y[1], -1.0)
|
| 23 |
+
assert np.isclose(y[2], 0.0)
|
| 24 |
+
assert np.isclose(y[3], 1.0)
|
| 25 |
+
assert np.isclose(y[4], 1.0)
|
| 26 |
+
print(" clip_and_normalize ok")
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def test_resample():
|
| 30 |
+
vol = np.random.randn(40, 80, 80).astype(np.float32)
|
| 31 |
+
out = resample_linear(vol, src_spacing=(2.5, 1.0, 1.0), dst_spacing=(5.0, 2.0, 2.0))
|
| 32 |
+
# factor = src/dst = 0.5,0.5,0.5 → shape halves
|
| 33 |
+
assert out.shape == (20, 40, 40), out.shape
|
| 34 |
+
print(" resample ok →", out.shape)
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
def test_pad_or_crop_grow():
|
| 38 |
+
v = np.ones((10, 10, 10), dtype=np.float32)
|
| 39 |
+
out, origin = pad_or_crop_to(v, (20, 20, 20), pad_value=-1.0)
|
| 40 |
+
assert out.shape == (20, 20, 20)
|
| 41 |
+
assert out[0, 0, 0] == -1 and out[10, 10, 10] == 1
|
| 42 |
+
print(" pad_or_crop grow ok")
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
def test_pad_or_crop_shrink():
|
| 46 |
+
v = np.arange(20 * 20 * 20).reshape(20, 20, 20).astype(np.float32)
|
| 47 |
+
out, origin = pad_or_crop_to(v, (10, 10, 10))
|
| 48 |
+
assert out.shape == (10, 10, 10)
|
| 49 |
+
assert origin == (5, 5, 5)
|
| 50 |
+
print(" pad_or_crop shrink ok")
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
def test_bbox():
|
| 54 |
+
m = np.zeros((30, 30, 30), dtype=np.uint8)
|
| 55 |
+
m[5:15, 10:20, 8:25] = 1
|
| 56 |
+
bb = bbox_from_mask(m, pad_vox=[2, 3, 1])
|
| 57 |
+
assert bb == (3, 17, 7, 23, 7, 26), bb
|
| 58 |
+
print(" bbox ok →", bb)
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
def test_preprocess_whole():
|
| 62 |
+
ct = (np.random.rand(200, 512, 512) * 2000 - 1000).astype(np.float32)
|
| 63 |
+
out = preprocess_whole(ct, src_spacing_dhw=(3.0, 1.0, 1.0), bucket=BUCKET_WHOLE)
|
| 64 |
+
assert out.volume.shape == (128, 256, 256)
|
| 65 |
+
assert out.volume.min() >= -1.001 and out.volume.max() <= 1.001
|
| 66 |
+
assert out.csize_vox == (200, 512, 512)
|
| 67 |
+
print(" preprocess_whole ok shape=", out.volume.shape, "csize=", out.csize_mm)
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
def test_preprocess_pan_crop_order_fix():
|
| 71 |
+
"""The v4 bug: if we bbox before resample, and native spacing is anisotropic,
|
| 72 |
+
the padding in voxels under-allocates along the fine axis.
|
| 73 |
+
Correct order (resample→bbox→crop) should give consistent crops whether
|
| 74 |
+
source spacing is isotropic or anisotropic."""
|
| 75 |
+
D, H, W = 100, 300, 300
|
| 76 |
+
ct = (np.random.rand(D, H, W) * 2000 - 1000).astype(np.float32)
|
| 77 |
+
# place a small pancreas blob
|
| 78 |
+
lab = np.zeros((D, H, W), dtype=np.uint8)
|
| 79 |
+
lab[40:55, 140:180, 140:180] = 17 # pancreas head
|
| 80 |
+
# native spacing anisotropic
|
| 81 |
+
out = preprocess_pan_crop(ct, lab, src_spacing_dhw=(3.0, 0.8, 0.8), bucket=BUCKET_PAN)
|
| 82 |
+
assert out is not None, "pan-crop returned None"
|
| 83 |
+
assert out.volume.shape == (64, 192, 256)
|
| 84 |
+
# mask should not be clipped
|
| 85 |
+
print(" preprocess_pan_crop ok ccrop_vox=", out.ccrop_vox,
|
| 86 |
+
"ccrop_mm=", [round(x, 1) for x in out.ccrop_mm])
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
def test_preprocess_pan_crop_empty_mask():
|
| 90 |
+
ct = (np.random.rand(50, 200, 200) * 2000 - 1000).astype(np.float32)
|
| 91 |
+
lab = np.zeros((50, 200, 200), dtype=np.uint8)
|
| 92 |
+
out = preprocess_pan_crop(ct, lab, src_spacing_dhw=(3.0, 1.0, 1.0), bucket=BUCKET_PAN)
|
| 93 |
+
assert out is None, "empty mask must return None, got PreprocessOutput"
|
| 94 |
+
print(" empty mask returns None ok")
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
if __name__ == "__main__":
|
| 98 |
+
print("running preprocessing tests...")
|
| 99 |
+
test_clip_and_normalize()
|
| 100 |
+
test_resample()
|
| 101 |
+
test_pad_or_crop_grow()
|
| 102 |
+
test_pad_or_crop_shrink()
|
| 103 |
+
test_bbox()
|
| 104 |
+
test_preprocess_whole()
|
| 105 |
+
test_preprocess_pan_crop_order_fix()
|
| 106 |
+
test_preprocess_pan_crop_empty_mask()
|
| 107 |
+
print("ALL TESTS PASS")
|
raw_pants_train_test/metadata/pants-captions-ldm/code/vae_sanity_gate.py
ADDED
|
@@ -0,0 +1,298 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
P7: Stage-0 VAE sanity gate on 200 held-out cases (v4 §14 item 3).
|
| 3 |
+
|
| 4 |
+
Purpose
|
| 5 |
+
-------
|
| 6 |
+
Before any diffusion training, verify the frozen Wan 2.2 VAE can reconstruct
|
| 7 |
+
PanTS CT volumes with sufficient fidelity under our preprocessing pipeline.
|
| 8 |
+
If the VAE fails on medical grayscale volumes (it was trained on natural
|
| 9 |
+
video), there is no point in training a downstream diffusion model on its
|
| 10 |
+
latents.
|
| 11 |
+
|
| 12 |
+
Gate criteria (computed in normalized [-1, 1] space on the raw CT volume,
|
| 13 |
+
and also in HU domain after de-normalization):
|
| 14 |
+
* mean absolute HU error <= 35 HU
|
| 15 |
+
* PSNR (grayscale, scale -1..+1) >= 28 dB
|
| 16 |
+
* SSIM (volumetric) >= 0.90
|
| 17 |
+
* pancreas region IoU of the mask-overlaid region (voxel-wise threshold
|
| 18 |
+
comparison at HU in [0, 150]) >= 0.85
|
| 19 |
+
|
| 20 |
+
Scope
|
| 21 |
+
-----
|
| 22 |
+
200 held-out cases randomly sampled from the VAL split (NOT the test split;
|
| 23 |
+
we do not want to contaminate the official PanTS-te holdout).
|
| 24 |
+
|
| 25 |
+
STATUS
|
| 26 |
+
------
|
| 27 |
+
Scaffold only — requires Wan 2.2 VAE weights which are NOT available on disk
|
| 28 |
+
at the time of writing. When weights appear, drop them at
|
| 29 |
+
$WAN22_VAE_CKPT
|
| 30 |
+
and run:
|
| 31 |
+
python vae_sanity_gate.py [--cases 200] [--bucket B-whole|B-abd|...]
|
| 32 |
+
|
| 33 |
+
The script prints a verdict line ("PASS" / "FAIL") and writes a per-case
|
| 34 |
+
JSON report to vae_sanity_gate_report.jsonl.
|
| 35 |
+
"""
|
| 36 |
+
from __future__ import annotations
|
| 37 |
+
import argparse
|
| 38 |
+
import json
|
| 39 |
+
import math
|
| 40 |
+
import os
|
| 41 |
+
import random
|
| 42 |
+
from pathlib import Path
|
| 43 |
+
|
| 44 |
+
import numpy as np
|
| 45 |
+
|
| 46 |
+
ROOT = Path(__file__).resolve().parents[1]
|
| 47 |
+
SPLITS = ROOT / "ldm_training" / "splits.json"
|
| 48 |
+
BUCKETS = ROOT / "ldm_training" / "bucket_spec.json"
|
| 49 |
+
OUT = ROOT / "ldm_training" / "vae_sanity_gate_report.jsonl"
|
| 50 |
+
OUT_MD = ROOT / "ldm_training" / "vae_sanity_gate_summary.md"
|
| 51 |
+
|
| 52 |
+
# Thresholds
|
| 53 |
+
THR_HU_MAE = 35.0
|
| 54 |
+
THR_PSNR_DB = 28.0
|
| 55 |
+
THR_SSIM = 0.90
|
| 56 |
+
THR_PANC_IOU = 0.85
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
# --- metrics ---------------------------------------------------------------
|
| 60 |
+
|
| 61 |
+
def psnr_minus1_plus1(x: np.ndarray, y: np.ndarray) -> float:
|
| 62 |
+
"""PSNR with signal range = 2.0 (i.e., values in [-1, 1])."""
|
| 63 |
+
mse = float(((x - y) ** 2).mean())
|
| 64 |
+
if mse < 1e-12:
|
| 65 |
+
return 100.0
|
| 66 |
+
# 20 log10(max / sqrt(mse)) with max = 2.0
|
| 67 |
+
return 20.0 * math.log10(2.0 / math.sqrt(mse))
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
def _gauss_kernel_3d(sigma: float = 1.5, size: int = 7) -> np.ndarray:
|
| 71 |
+
ax = np.arange(size) - (size - 1) / 2
|
| 72 |
+
g1 = np.exp(-(ax ** 2) / (2 * sigma ** 2))
|
| 73 |
+
g1 /= g1.sum()
|
| 74 |
+
g3 = g1[:, None, None] * g1[None, :, None] * g1[None, None, :]
|
| 75 |
+
return g3.astype(np.float32)
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
def ssim_3d(x: np.ndarray, y: np.ndarray, data_range: float = 2.0) -> float:
|
| 79 |
+
"""Volumetric SSIM with a 7x7x7 Gaussian window."""
|
| 80 |
+
from scipy.ndimage import convolve
|
| 81 |
+
k = _gauss_kernel_3d(sigma=1.5, size=7)
|
| 82 |
+
mu_x = convolve(x, k)
|
| 83 |
+
mu_y = convolve(y, k)
|
| 84 |
+
mu_x2 = mu_x * mu_x
|
| 85 |
+
mu_y2 = mu_y * mu_y
|
| 86 |
+
mu_xy = mu_x * mu_y
|
| 87 |
+
sig_x2 = convolve(x * x, k) - mu_x2
|
| 88 |
+
sig_y2 = convolve(y * y, k) - mu_y2
|
| 89 |
+
sig_xy = convolve(x * y, k) - mu_xy
|
| 90 |
+
c1 = (0.01 * data_range) ** 2
|
| 91 |
+
c2 = (0.03 * data_range) ** 2
|
| 92 |
+
num = (2 * mu_xy + c1) * (2 * sig_xy + c2)
|
| 93 |
+
den = (mu_x2 + mu_y2 + c1) * (sig_x2 + sig_y2 + c2)
|
| 94 |
+
return float((num / np.maximum(den, 1e-12)).mean())
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
def hu_mae(x_norm: np.ndarray, y_norm: np.ndarray,
|
| 98 |
+
hu_min: float = -1000.0, hu_max: float = 1000.0) -> float:
|
| 99 |
+
"""De-normalize [-1,1] → HU and compute mean-abs-error."""
|
| 100 |
+
def denorm(a):
|
| 101 |
+
return (a + 1.0) / 2.0 * (hu_max - hu_min) + hu_min
|
| 102 |
+
return float(np.abs(denorm(x_norm) - denorm(y_norm)).mean())
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
def pancreas_iou(orig: np.ndarray, recon: np.ndarray, mask: np.ndarray,
|
| 106 |
+
hu_lo: float = 0.0, hu_hi: float = 150.0) -> float | None:
|
| 107 |
+
"""IoU of the pancreas-region thresholded volumes in HU [0, 150]."""
|
| 108 |
+
if mask.sum() == 0:
|
| 109 |
+
return None
|
| 110 |
+
# work in HU
|
| 111 |
+
def denorm(a):
|
| 112 |
+
return (a + 1.0) / 2.0 * 2000.0 - 1000.0
|
| 113 |
+
o_hu = denorm(orig)
|
| 114 |
+
r_hu = denorm(recon)
|
| 115 |
+
o_bin = ((o_hu >= hu_lo) & (o_hu <= hu_hi) & (mask > 0)).astype(np.uint8)
|
| 116 |
+
r_bin = ((r_hu >= hu_lo) & (r_hu <= hu_hi) & (mask > 0)).astype(np.uint8)
|
| 117 |
+
inter = int((o_bin & r_bin).sum())
|
| 118 |
+
union = int((o_bin | r_bin).sum())
|
| 119 |
+
return (inter / union) if union > 0 else None
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
# --- VAE loader ------------------------------------------------------------
|
| 123 |
+
|
| 124 |
+
def load_wan22_vae(ckpt_path: str, device: str = "cuda"):
|
| 125 |
+
"""Load the frozen Wan 2.2 VAE. This function is a placeholder: the exact
|
| 126 |
+
import path depends on which Wan 2.2 distribution is used (official
|
| 127 |
+
repo vs diffusers port). When the weights are on disk, replace this
|
| 128 |
+
body with the appropriate loader, e.g.:
|
| 129 |
+
|
| 130 |
+
from wan22 import WanVAE
|
| 131 |
+
vae = WanVAE.from_pretrained(ckpt_path, dtype=torch.float16).to(device)
|
| 132 |
+
vae.eval().requires_grad_(False)
|
| 133 |
+
return vae
|
| 134 |
+
|
| 135 |
+
The downstream code assumes:
|
| 136 |
+
z = vae.encode(x_video) # (B, 48, D/4, H/16, W/16)
|
| 137 |
+
x_hat = vae.decode(z) # (B, 1, D, H, W) (or (B, 3, D, H, W) depending on variant)
|
| 138 |
+
|
| 139 |
+
For PanTS (grayscale), we stage the CT as a 3-channel video by
|
| 140 |
+
broadcasting ct→(ct, ct, ct) on the channel axis before encoding,
|
| 141 |
+
and average over channels on decode.
|
| 142 |
+
"""
|
| 143 |
+
raise RuntimeError(
|
| 144 |
+
f"Wan 2.2 VAE checkpoint not found at {ckpt_path}. "
|
| 145 |
+
"Set $WAN22_VAE_CKPT and re-run. This scaffold validates the "
|
| 146 |
+
"metric + data pipeline; it cannot run end-to-end until weights arrive."
|
| 147 |
+
)
|
| 148 |
+
|
| 149 |
+
|
| 150 |
+
def encode_decode(vae, x_dhw: np.ndarray, device: str = "cuda") -> np.ndarray:
|
| 151 |
+
import torch
|
| 152 |
+
# CT is 1-channel in [-1, 1]; broadcast to 3-ch video for Wan VAE
|
| 153 |
+
x = torch.from_numpy(x_dhw).to(device=device, dtype=torch.float32)
|
| 154 |
+
x = x[None, None].expand(1, 3, -1, -1, -1) # (1, 3, D, H, W)
|
| 155 |
+
with torch.no_grad():
|
| 156 |
+
z = vae.encode(x)
|
| 157 |
+
x_hat = vae.decode(z)
|
| 158 |
+
x_hat = x_hat.mean(dim=1, keepdim=False)[0].clamp(-1, 1).cpu().numpy()
|
| 159 |
+
return x_hat.astype(np.float32)
|
| 160 |
+
|
| 161 |
+
|
| 162 |
+
# --- main ------------------------------------------------------------------
|
| 163 |
+
|
| 164 |
+
def sample_cases(n: int = 200, seed: int = 42) -> list[str]:
|
| 165 |
+
splits = json.loads(SPLITS.read_text())
|
| 166 |
+
val = splits["val"] # 819 ids
|
| 167 |
+
rng = random.Random(seed)
|
| 168 |
+
rng.shuffle(val)
|
| 169 |
+
return val[:n]
|
| 170 |
+
|
| 171 |
+
|
| 172 |
+
def run_one(vae, cid: str, bucket: dict) -> dict:
|
| 173 |
+
import sys
|
| 174 |
+
sys.path.insert(0, str(ROOT / "ldm_training"))
|
| 175 |
+
from preprocessing import load_nifti_ras, preprocess_whole, PANC_CLASSES
|
| 176 |
+
|
| 177 |
+
ct_path = f"{os.environ.get('PANTS_DATA_ROOT', './PanTS_data')}/ImageTr/{cid}/{cid}.nii.gz"
|
| 178 |
+
lab_path = f"{os.environ.get('PANTS_DATA_ROOT', './PanTS_data')}/LabelTr/{cid}/combined_labels.nii.gz"
|
| 179 |
+
ct, sp = load_nifti_ras(ct_path)
|
| 180 |
+
lab, _ = load_nifti_ras(lab_path)
|
| 181 |
+
pre = preprocess_whole(ct, sp, bucket)
|
| 182 |
+
x = pre.volume
|
| 183 |
+
# align mask similarly
|
| 184 |
+
from preprocessing import resample_nn, pad_or_crop_to
|
| 185 |
+
panc_mask = np.isin(lab, PANC_CLASSES).astype(np.uint8)
|
| 186 |
+
panc_res = resample_nn(panc_mask, sp, bucket["voxel_spacing_mm"])
|
| 187 |
+
panc_aligned, _ = pad_or_crop_to(panc_res.astype(np.float32),
|
| 188 |
+
bucket["shape_dhw"], pad_value=0.0)
|
| 189 |
+
panc_aligned = (panc_aligned > 0.5).astype(np.uint8)
|
| 190 |
+
|
| 191 |
+
x_hat = encode_decode(vae, x)
|
| 192 |
+
|
| 193 |
+
return {
|
| 194 |
+
"id": cid,
|
| 195 |
+
"hu_mae": hu_mae(x, x_hat),
|
| 196 |
+
"psnr_db": psnr_minus1_plus1(x, x_hat),
|
| 197 |
+
"ssim": ssim_3d(x, x_hat),
|
| 198 |
+
"panc_iou": pancreas_iou(x, x_hat, panc_aligned),
|
| 199 |
+
}
|
| 200 |
+
|
| 201 |
+
|
| 202 |
+
def verdict(rows: list[dict]) -> dict:
|
| 203 |
+
vals = lambda k: [r[k] for r in rows if r.get(k) is not None]
|
| 204 |
+
summary = {
|
| 205 |
+
"n": len(rows),
|
| 206 |
+
"hu_mae_mean": float(np.mean(vals("hu_mae"))),
|
| 207 |
+
"psnr_db_mean": float(np.mean(vals("psnr_db"))),
|
| 208 |
+
"ssim_mean": float(np.mean(vals("ssim"))),
|
| 209 |
+
"panc_iou_mean": float(np.mean(vals("panc_iou"))) if vals("panc_iou") else None,
|
| 210 |
+
}
|
| 211 |
+
summary["pass"] = (
|
| 212 |
+
summary["hu_mae_mean"] <= THR_HU_MAE
|
| 213 |
+
and summary["psnr_db_mean"] >= THR_PSNR_DB
|
| 214 |
+
and summary["ssim_mean"] >= THR_SSIM
|
| 215 |
+
and (summary["panc_iou_mean"] is None
|
| 216 |
+
or summary["panc_iou_mean"] >= THR_PANC_IOU)
|
| 217 |
+
)
|
| 218 |
+
return summary
|
| 219 |
+
|
| 220 |
+
|
| 221 |
+
def main():
|
| 222 |
+
ap = argparse.ArgumentParser()
|
| 223 |
+
ap.add_argument("--cases", type=int, default=200)
|
| 224 |
+
ap.add_argument("--bucket", default="B-whole")
|
| 225 |
+
ap.add_argument("--ckpt", default=os.environ.get("WAN22_VAE_CKPT", ""))
|
| 226 |
+
ap.add_argument("--dry", action="store_true",
|
| 227 |
+
help="Skip VAE: run preprocessing pipeline + identity recon "
|
| 228 |
+
"as a pipeline sanity test (metrics will be perfect).")
|
| 229 |
+
args = ap.parse_args()
|
| 230 |
+
|
| 231 |
+
buckets = json.loads(BUCKETS.read_text())["buckets"]
|
| 232 |
+
bucket = {b["name"]: b for b in buckets}[args.bucket]
|
| 233 |
+
|
| 234 |
+
ids = sample_cases(n=args.cases)
|
| 235 |
+
print(f"sampled {len(ids)} validation cases for bucket={args.bucket}")
|
| 236 |
+
|
| 237 |
+
if args.dry:
|
| 238 |
+
# identity path for scaffold verification
|
| 239 |
+
vae = None
|
| 240 |
+
|
| 241 |
+
def encode_decode_identity(_vae, x_dhw, device="cuda"):
|
| 242 |
+
return x_dhw.copy()
|
| 243 |
+
|
| 244 |
+
import builtins
|
| 245 |
+
# monkey-patch for dry run
|
| 246 |
+
global encode_decode
|
| 247 |
+
_saved = encode_decode
|
| 248 |
+
encode_decode = encode_decode_identity # type: ignore
|
| 249 |
+
try:
|
| 250 |
+
rows = []
|
| 251 |
+
for i, cid in enumerate(ids[:3]): # dry only hits 3 to stay cheap
|
| 252 |
+
print(f" [{i+1}/3] {cid}")
|
| 253 |
+
try:
|
| 254 |
+
rows.append(run_one(None, cid, bucket))
|
| 255 |
+
except FileNotFoundError as e:
|
| 256 |
+
print(f" skip: {e}")
|
| 257 |
+
continue
|
| 258 |
+
finally:
|
| 259 |
+
encode_decode = _saved # type: ignore
|
| 260 |
+
else:
|
| 261 |
+
vae = load_wan22_vae(args.ckpt)
|
| 262 |
+
rows = []
|
| 263 |
+
for i, cid in enumerate(ids):
|
| 264 |
+
print(f" [{i+1}/{len(ids)}] {cid}")
|
| 265 |
+
rows.append(run_one(vae, cid, bucket))
|
| 266 |
+
|
| 267 |
+
OUT.parent.mkdir(parents=True, exist_ok=True)
|
| 268 |
+
with open(OUT, "w") as f:
|
| 269 |
+
for r in rows:
|
| 270 |
+
f.write(json.dumps(r) + "\n")
|
| 271 |
+
|
| 272 |
+
summ = verdict(rows)
|
| 273 |
+
lines = [
|
| 274 |
+
"# VAE sanity gate",
|
| 275 |
+
"",
|
| 276 |
+
f"bucket: {args.bucket} (shape {bucket['shape_dhw']}, spacing {bucket['voxel_spacing_mm']})",
|
| 277 |
+
f"ckpt: {args.ckpt or 'DRY RUN (identity)'}",
|
| 278 |
+
f"cases: {summ['n']}",
|
| 279 |
+
"",
|
| 280 |
+
"## Metrics (mean over cases)",
|
| 281 |
+
f"| metric | value | threshold | pass |",
|
| 282 |
+
f"|--------------|---------------------------|------------------|------|",
|
| 283 |
+
f"| HU MAE | {summ['hu_mae_mean']:.3f} | <= {THR_HU_MAE} | {'Y' if summ['hu_mae_mean']<=THR_HU_MAE else 'N'} |",
|
| 284 |
+
f"| PSNR (dB) | {summ['psnr_db_mean']:.3f} | >= {THR_PSNR_DB} | {'Y' if summ['psnr_db_mean']>=THR_PSNR_DB else 'N'} |",
|
| 285 |
+
f"| SSIM | {summ['ssim_mean']:.4f} | >= {THR_SSIM} | {'Y' if summ['ssim_mean']>=THR_SSIM else 'N'} |",
|
| 286 |
+
f"| pancreas IoU | {summ.get('panc_iou_mean')} | >= {THR_PANC_IOU} | {'Y' if (summ.get('panc_iou_mean') is None or summ['panc_iou_mean']>=THR_PANC_IOU) else 'N'} |",
|
| 287 |
+
"",
|
| 288 |
+
f"## VERDICT: {'PASS' if summ['pass'] else 'FAIL'}",
|
| 289 |
+
]
|
| 290 |
+
OUT_MD.write_text("\n".join(lines) + "\n")
|
| 291 |
+
print()
|
| 292 |
+
print("\n".join(lines))
|
| 293 |
+
print(f"\nwrote {OUT}")
|
| 294 |
+
print(f"wrote {OUT_MD}")
|
| 295 |
+
|
| 296 |
+
|
| 297 |
+
if __name__ == "__main__":
|
| 298 |
+
main()
|
raw_pants_train_test/metadata/pants-captions-ldm/docs/GPT_REVIEW.md
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# GPT-5.4 xhigh Review (via Codex MCP) — 原文存档
|
| 2 |
+
|
| 3 |
+
## 核心结论
|
| 4 |
+
|
| 5 |
+
1. **方向对**：确定性 mask→事实 + LLM 改写 优于 3D VLM 主力。
|
| 6 |
+
2. **最大风险不是"caption 不流畅"，而是"事实污染"**：impression 里的非视觉诊断、FOV 外病变、negation 翻转。
|
| 7 |
+
3. **5 变体改 4 个核心 + V5 负样本**。V4 tag-string 降级为 ablation — T5 encoder 自然语言更 match 预训练分布。
|
| 8 |
+
4. **Scan metadata**：保留 `phase`；删掉 `manufacturer`/`model`（会让 diffusion 学到无关 shortcut）。
|
| 9 |
+
5. **负样本塌缩**：V5_neg_diverse 不够；要数据侧分层采样 + 显式否定语 + 永不 drop 正样本 lesion 句。
|
| 10 |
+
6. **QC**：regex parse-back 必要但不充分；加 schema-aware 生成 + 针对性 LLM 仲裁 + 抽样人工审。
|
| 11 |
+
7. **直接 fork RadGPT**，不重写。
|
| 12 |
+
|
| 13 |
+
## GPT 原回答里的关键引用
|
| 14 |
+
|
| 15 |
+
- PanTS: arXiv:2507.01291
|
| 16 |
+
- RadGPT: arXiv:2501.04678, https://github.com/MrGiovanni/RadGPT
|
| 17 |
+
- GenerateCT: arXiv:2305.16037
|
| 18 |
+
- MedSyn: arXiv:2310.03559
|
| 19 |
+
- MAISI: arXiv:2409.11169
|
| 20 |
+
|
| 21 |
+
## 未覆盖的盲点（GPT 明说）
|
| 22 |
+
|
| 23 |
+
- regex QC 在 paraphrase 上的 FP 率：无可靠发表数字，取决于 paraphraser 自由度。
|
| 24 |
+
- "每 volume 多少 caption" 在医学 3D T2I 上无已发表最优值；猜测 2–4 足够。
|
| 25 |
+
|
| 26 |
+
## 用户的 pushback
|
| 27 |
+
|
| 28 |
+
GPT 建议"放弃 VLM"，用户反对：VLM 不做诊断但**能描述可见外观**（纹理 / 伪影 / 体型 / 胃肠气液），这些 mask 和报告都没有。
|
| 29 |
+
|
| 30 |
+
v3.1 的三源融合正是为了兼容两方：VLM 只做 **visual phrases**（白名单过滤），事实层仍然由 mask + report 确定性产出。
|
raw_pants_train_test/metadata/pants-captions-ldm/docs/LDM_TRAINING_PLAN_v4.md
ADDED
|
@@ -0,0 +1,416 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# PanTS T2I-3D LDM — Training Plan v4 (post-expert-review)
|
| 2 |
+
|
| 3 |
+
**Status**: v4 = v3 with all blocker-level fixes applied from 2026-04-19 expert review. Each fix cross-referenced to the review item.
|
| 4 |
+
|
| 5 |
+
**Headline**: No training until the 9 pre-training checklist items at §14 are closed.
|
| 6 |
+
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
## Summary of changes vs v3
|
| 10 |
+
|
| 11 |
+
| # | Area | v3 position | v4 fix | Review ref |
|
| 12 |
+
|---|------|-------------|--------|-----------|
|
| 13 |
+
| 1 | License | "CC-BY-NC 4.0, public shards" | **CC BY-NC-ND 4.0 (PanTS actual). No derivative volumes/latents/captions to public HF. Code + manifests only; data goes to gated/private pending written permission.** | §1 |
|
| 14 |
+
| 2 | Train/val split | 9:1 hash split over all 9,901 | **Use official PanTS-tr 9,000 for training; PanTS-te 901 strictly held out.** val = carve 800 from PanTS-tr. | §2 |
|
| 15 |
+
| 3 | Bucket math | "B-whole-mid 9,600 tokens > 8,192 → drop" | **Token math was latent voxels, not DiT patch tokens.** Recomputed with patch (1,2,2). Buckets redesigned around physical coverage, not token budget. | §3 |
|
| 16 |
+
| 4 | Pan-crop preprocessing | Resample → pad/crop → bbox-crop | **Order fixed: resample → bbox-from-mask → crop-with-padding → pad-or-crop-to-bucket.** | §4 |
|
| 17 |
+
| 5 | Augmentation | LR flip p=0.5 + spacing jitter + intensity jitter | **LR flip OFF (anatomy not bilateral). Spacing jitter OFF (breaks latent cache). Intensity jitter replaced with tiny latent-space noise.** | §5 |
|
| 18 |
+
| 6 | Diffusion objective | v-pred + Min-SNR-γ + ZTSNR + EDM σ_min/max + bucket-rescaled σ_max | **Pick route A: VP-beta + zero-terminal-SNR + v-prediction + v-pred-adjusted Min-SNR-γ. No EDM language. No bucket-rescaled σ_max.** | §6 |
|
| 19 |
+
| 7 | VAE Stage-0 | removed | **Restored as sanity gate** (shape roundtrip, edge slices, per-channel latent stats, soft-tissue/air/bone HU MAE). Model-selection part removed. | §7 |
|
| 20 |
+
| 8 | Caption: V2 neg vs COMMON_RULES | "no focal lesion" allowed | **Negative rewriting bans lesion/mass/tumor/cyst in all variants' negative mode. QC scans all variants, not only V6_neg.** | §2.3, §8.1 |
|
| 21 |
+
| 9 | Caption: V6_neg vs renal-cyst incidental | conflict | **V6_neg drops renal-cyst incidental to avoid token match; or banned list narrowed to `pancreatic cyst`.** | §8.2 |
|
| 22 |
+
| 10 | Caption phase audit | not present | **4-way phase audit (report / DICOM / organ-HU heuristic / caption); disagreeing cases → `phase=unknown` at training.** | §8.3 |
|
| 23 |
+
| 11 | V7 caption | "pancreas-only, descriptive" | **Added explicit whitelist** for duct/vessel contact terms + explicit ban on invasion/encasement/stage/resectability. | §8.4 |
|
| 24 |
+
|
| 25 |
+
---
|
| 26 |
+
|
| 27 |
+
## 1. License & release policy (BLOCKER #1 resolved)
|
| 28 |
+
|
| 29 |
+
**PanTS actual license**: **CC BY-NC-ND 4.0** (verified at https://github.com/MrGiovanni/PanTS/blob/main/LICENSE).
|
| 30 |
+
|
| 31 |
+
Implication: adapted material may be produced but **must not be shared**. All of the following are derivatives:
|
| 32 |
+
- Resampled / cropped / HU-clipped volumes
|
| 33 |
+
- Wan 2.2 precomputed latents
|
| 34 |
+
- VLM outputs + LLM-fused captions
|
| 35 |
+
- canonical_facts (parsed from report)
|
| 36 |
+
|
| 37 |
+
**v4 release policy**:
|
| 38 |
+
1. **Public repo** (HF / GitHub) ships ONLY:
|
| 39 |
+
- Preprocessing scripts (`preprocessing.py`, `bucket_spec.json`)
|
| 40 |
+
- Caption generation pipeline (`canonical_facts.py`, `pilot_run_v2.py` prompts, `fusion_run.py`)
|
| 41 |
+
- VAE stats + bucket spec as JSON
|
| 42 |
+
- Case-id manifest (non-derivative — just IDs from the public PanTS release)
|
| 43 |
+
- Training code, config, model weights
|
| 44 |
+
2. **Volumes, latents, captions** stay private until **written permission from PanTS/JHU** (MrGiovanni group).
|
| 45 |
+
3. **Repo name**: drop `Anthropic-PanTS`; use `PanTS-LDM-recipe` or similar until provenance is clear.
|
| 46 |
+
4. Dataset card must read **"License: CC BY-NC-ND 4.0 — share adapted material pending upstream permission"** and link to PanTS upstream.
|
| 47 |
+
5. Add a **Conditional Access** form on HF: researchers agree to same CC BY-NC-ND terms before download.
|
| 48 |
+
|
| 49 |
+
Until this is resolved, all plans below assume **private/internal** training only.
|
| 50 |
+
|
| 51 |
+
---
|
| 52 |
+
|
| 53 |
+
## 2. Data split (BLOCKER #2 resolved)
|
| 54 |
+
|
| 55 |
+
PanTS public split: **PanTS-tr 9,000** + **PanTS-te 901**.
|
| 56 |
+
|
| 57 |
+
**v4 split policy**:
|
| 58 |
+
- **Train pool**: PanTS-tr 9,000 cases.
|
| 59 |
+
- **Val pool**: 800 cases carved from PanTS-tr (deterministic by `sorted(case_id)[::11]` — stable across reruns).
|
| 60 |
+
- **Test pool**: **PanTS-te 901, strictly held out**. Never enters:
|
| 61 |
+
- VAE latent-stats calibration
|
| 62 |
+
- Caption diversity stats
|
| 63 |
+
- Bucket spacing calibration
|
| 64 |
+
- CFG-scale tuning
|
| 65 |
+
- Any training data of downstream nnU-Net / lesion classifier
|
| 66 |
+
- All metrics reported in the paper go against PanTS-te.
|
| 67 |
+
|
| 68 |
+
Split written to `splits.json` as `{case_id: "train"/"val"/"test"}` — committed to repo for reproducibility.
|
| 69 |
+
|
| 70 |
+
---
|
| 71 |
+
|
| 72 |
+
## 3. Bucket redesign (BLOCKER #3 resolved)
|
| 73 |
+
|
| 74 |
+
### 3.1 Token math correction
|
| 75 |
+
DiT-3D uses patch `(1, 2, 2)` over latent `(C=48, D_l, H_l, W_l)`. Tokens per sample:
|
| 76 |
+
```
|
| 77 |
+
n_tokens = D_l · (H_l // 2) · (W_l // 2)
|
| 78 |
+
```
|
| 79 |
+
v3 confused "latent voxels" with "tokens". Recomputed.
|
| 80 |
+
|
| 81 |
+
### 3.2 Physical-coverage-first bucket design
|
| 82 |
+
|
| 83 |
+
Target: each bucket's **physical Z extent ≥ the 75th percentile of its FOV class**, so captions like "whole-body" or "CAP" actually match what the model sees.
|
| 84 |
+
|
| 85 |
+
| Bucket | FOV | Spacing ZYX (mm) | Volume shape D×Y×X | Physical cover ZYX (mm) | Latent 48×Dl×Hl×Wl | DiT tokens |
|
| 86 |
+
|--------|-----|------------------|--------------------|-----|-----------------|-----------|
|
| 87 |
+
| B-whole | whole_body | 5.0 × 2.0 × 2.0 | 128 × 256 × 256 | **640 × 512 × 512** | 48×32×16×16 | 2,048 |
|
| 88 |
+
| B-CAP | CAP | 4.0 × 2.0 × 2.0 | 96 × 320 × 256 | **384 × 640 × 512** | 48×24×20×16 | 1,920 |
|
| 89 |
+
| B-CA | CA | 4.0 × 2.0 × 2.0 | 80 × 320 × 256 | 320 × 640 × 512 | 48×20×20×16 | 1,600 |
|
| 90 |
+
| B-abd | abdomen | 4.0 × 2.0 × 2.0 | 64 × 256 × 256 | 256 × 512 × 512 | 48×16×16×16 | 1,024 |
|
| 91 |
+
| B-pan | pan_crop | 2.0 × 1.0 × 1.0 | 64 × 192 × 256 | 128 × 192 × 256 | 48×16×12×16 | 768 |
|
| 92 |
+
|
| 93 |
+
Constraints preserved:
|
| 94 |
+
- D % 4 == 0 (Wan causal VAE temporal stride)
|
| 95 |
+
- H, W % 16 == 0 (Wan spatial stride)
|
| 96 |
+
|
| 97 |
+
B-pan pan-crop Z extent = 128 mm, covers 95th percentile of `(pancreas_bbox_z + 2·40 mm)` per prior measurement. If a case overflows, it uses asymmetric padding rather than hard crop.
|
| 98 |
+
|
| 99 |
+
### 3.3 Per-step batching
|
| 100 |
+
- Like-bucket minibatches (NovelAI §4.1).
|
| 101 |
+
- Sampling weight ∝ √(bucket_size) for balance without starving small buckets.
|
| 102 |
+
- Per-bucket batch size tuned to target identical memory footprint.
|
| 103 |
+
|
| 104 |
+
---
|
| 105 |
+
|
| 106 |
+
## 4. Preprocessing (BLOCKER #4: pan-crop order fixed)
|
| 107 |
+
|
| 108 |
+
### 4.1 Whole/CAP/CA/abd bucket path
|
| 109 |
+
```python
|
| 110 |
+
def preprocess_whole(ct_path, label_path, meta, bucket):
|
| 111 |
+
vol, spacing_xyz = load_nifti(ct_path)
|
| 112 |
+
mask = load_nifti(label_path)[0]
|
| 113 |
+
vol, mask, spacing_xyz = to_RAS(vol, mask, spacing_xyz)
|
| 114 |
+
vol = np.clip(vol, -1000, 1000).astype(np.float32)
|
| 115 |
+
vol = (vol + 1000.0) / 1000.0 - 1.0
|
| 116 |
+
vol = resample(vol, spacing_xyz, bucket.spacing, mode="trilinear")
|
| 117 |
+
mask = resample(mask, spacing_xyz, bucket.spacing, mode="nearest")
|
| 118 |
+
# center-crop/pad along XY only; Z uses body-centered crop guided by pancreas centroid
|
| 119 |
+
vol, mask, ccrop = pad_or_crop_centered(vol, mask, bucket.shape,
|
| 120 |
+
z_anchor="pancreas_centroid",
|
| 121 |
+
pad_value=-1.0)
|
| 122 |
+
return pack_sample(vol, mask, ccrop, meta, bucket)
|
| 123 |
+
```
|
| 124 |
+
|
| 125 |
+
### 4.2 Pan-crop bucket path — **fixed order**
|
| 126 |
+
```python
|
| 127 |
+
def preprocess_pan(ct_path, label_path, meta, bucket):
|
| 128 |
+
vol, spacing_xyz = load_nifti(ct_path)
|
| 129 |
+
mask = load_nifti(label_path)[0]
|
| 130 |
+
vol, mask, spacing_xyz = to_RAS(vol, mask, spacing_xyz)
|
| 131 |
+
vol = np.clip(vol, -1000, 1000).astype(np.float32)
|
| 132 |
+
vol = (vol + 1000.0) / 1000.0 - 1.0
|
| 133 |
+
|
| 134 |
+
# 1) resample to bucket spacing FIRST
|
| 135 |
+
vol = resample(vol, spacing_xyz, bucket.spacing, mode="trilinear")
|
| 136 |
+
mask = resample(mask, spacing_xyz, bucket.spacing, mode="nearest")
|
| 137 |
+
|
| 138 |
+
# 2) compute pancreas-union bbox from mask IN RESAMPLED SPACE
|
| 139 |
+
pan_mask = np.isin(mask, [17, 18, 19, 20])
|
| 140 |
+
if not pan_mask.any():
|
| 141 |
+
return None # skip case
|
| 142 |
+
bbox = bbox_of_mask(pan_mask, pad_mm=40, spacing=bucket.spacing)
|
| 143 |
+
|
| 144 |
+
# 3) crop vol/mask to bbox, THEN pad-or-crop to bucket shape
|
| 145 |
+
vol, mask, crop_origin = crop_bbox(vol, mask, bbox)
|
| 146 |
+
vol, mask, pad_origin = pad_or_crop_to_shape(vol, mask, bucket.shape,
|
| 147 |
+
pad_value=-1.0)
|
| 148 |
+
|
| 149 |
+
# 4) ccrop stores origin relative to FULL resampled volume
|
| 150 |
+
ccrop_vox = crop_origin + pad_origin
|
| 151 |
+
ccrop_mm = ccrop_vox * np.array(bucket.spacing)
|
| 152 |
+
return pack_sample(vol, mask, ccrop_vox, meta, bucket, ccrop_mm=ccrop_mm)
|
| 153 |
+
```
|
| 154 |
+
|
| 155 |
+
Both `ccrop_vox` and `ccrop_mm` are stored in the cond JSON so micro-conditioning can be rendered in physical or voxel units.
|
| 156 |
+
|
| 157 |
+
### 4.3 Phase audit pipeline (new, precedes training) — see §8.3
|
| 158 |
+
|
| 159 |
+
---
|
| 160 |
+
|
| 161 |
+
## 5. Augmentation (BLOCKER #5 resolved)
|
| 162 |
+
|
| 163 |
+
| Aug | v3 | v4 | Reason |
|
| 164 |
+
|-----|----|----|--------|
|
| 165 |
+
| Left-right flip | p=0.5 | **OFF** | Pancreas head/body/tail, gallbladder right, kidneys' laterality, liver/spleen asymmetry, post-op cavity — all lateralized. Flip without rewriting caption = poisoned supervision. |
|
| 166 |
+
| Z flip | off | off (unchanged) | anatomy directional |
|
| 167 |
+
| Rotation | off | off (unchanged) | |
|
| 168 |
+
| Spacing jitter ±5% | optional | **OFF** | breaks latent cache; doing spacing jitter with cached latents is invalid |
|
| 169 |
+
| HU intensity jitter ±5% | p=0.3 | **OFF** in raw HU space | ±5% on clipped HU would change phase/HU facts in captions |
|
| 170 |
+
| Latent-space Gaussian noise | n/a | **ADD**: σ=0.02·std(z) per channel, p=0.3 | tiny perturbation that does not touch caption semantics |
|
| 171 |
+
| Crop-origin jitter | implicit | **ADD**: ±8 vox in each axis, update `ccrop` accordingly | cheap, safe, exposes micro-cond |
|
| 172 |
+
|
| 173 |
+
Latent-space noise vs HU-space jitter: cheaper (no recompute of VAE), safer (preserves HU histogram), and effectively what LDMs want — noise in the denoiser's input space.
|
| 174 |
+
|
| 175 |
+
---
|
| 176 |
+
|
| 177 |
+
## 6. Diffusion objective (BLOCKER #6: Route A chosen)
|
| 178 |
+
|
| 179 |
+
### 6.1 Picked route: **VP + Zero-Terminal-SNR + v-prediction + Min-SNR-γ**
|
| 180 |
+
|
| 181 |
+
No EDM language. No bucket-rescaled σ_max.
|
| 182 |
+
|
| 183 |
+
### 6.2 Forward schedule
|
| 184 |
+
Start from linear VP `β_t` schedule with `T=1000` steps, `β_1=1e-4, β_T=2e-2`. Then **rescale β to enforce zero terminal SNR** per Lin et al. 2023 (arXiv 2305.08891) §3:
|
| 185 |
+
1. compute `ᾱ_t = ∏(1 - β_t)`
|
| 186 |
+
2. rescale so `ᾱ_T = 0`
|
| 187 |
+
3. store the new `β_t` and `ᾱ_t` as the schedule
|
| 188 |
+
|
| 189 |
+
The inference sampler (DDIM, DPM-Solver++) steps from `t=T` downward, i.e. from pure-noise terminal step — not from a finite σ.
|
| 190 |
+
|
| 191 |
+
### 6.3 Target: v-prediction
|
| 192 |
+
```
|
| 193 |
+
v_t = √ᾱ_t · ε - √(1-ᾱ_t) · x_0
|
| 194 |
+
L = E_{x_0, t, ε, c} [ w(t) · || v̂_θ(z_t, t, c) - v_t ||² ]
|
| 195 |
+
```
|
| 196 |
+
|
| 197 |
+
### 6.4 Loss weighting: Min-SNR-γ adapted to v-pred
|
| 198 |
+
For v-prediction targets, Min-SNR-γ weight is applied to the SNR clamp but with the v-pred correction from Lin 2023 §5. At ZTSNR terminal step, SNR → 0, weight → 0 per Min-SNR definition, avoiding division-by-zero:
|
| 199 |
+
```
|
| 200 |
+
SNR(t) = ᾱ_t / (1 - ᾱ_t)
|
| 201 |
+
w_v(t) = min(SNR(t), γ) / (SNR(t) + 1) # v-pred variant (Lin 2023 eqn 19)
|
| 202 |
+
γ = 5
|
| 203 |
+
```
|
| 204 |
+
At `t=T` with `ᾱ_T=0`, SNR=0, w_v=0 — the terminal step contributes nothing to loss, consistent with the ZTSNR sampler expecting pure noise at `t=T`.
|
| 205 |
+
|
| 206 |
+
### 6.5 CFG and inference
|
| 207 |
+
- Train-time CFG drop: p=0.15 joint drop (text + micro-cond + spacing + body-region + phase) — Report2CT setting.
|
| 208 |
+
- Inference: DDIM or DPM-Solver++, 50 steps, **start from t=T** (pure noise, ZTSNR-consistent).
|
| 209 |
+
- Guidance: `ε̂ = ε_uncond + s · (ε_cond - ε_uncond)`, apply **CFG-rescale** trick from Lin 2023 §3.4 to mitigate HU over-saturation from high s.
|
| 210 |
+
- CFG sweep `s ∈ {2, 3, 4, 5, 6, 7}` × 50 held-out captions, evaluated on HU histogram drift + CLIP + quick radiologist rating — **not** decided a priori.
|
| 211 |
+
|
| 212 |
+
---
|
| 213 |
+
|
| 214 |
+
## 7. Stage 0 VAE sanity gate (restored, not model-selection)
|
| 215 |
+
|
| 216 |
+
User has already measured Wan 2.2 VAE beating MAISI on CT. So Stage 0 is no longer "pick a VAE" — it's "prove the boundary has no bugs."
|
| 217 |
+
|
| 218 |
+
Run on 200 PanTS-tr cases, balanced across FOV classes and lesion+/lesion-:
|
| 219 |
+
|
| 220 |
+
| Check | Pass condition |
|
| 221 |
+
|-------|----------------|
|
| 222 |
+
| Shape roundtrip | every bucket's `encode→decode` returns exact input shape (no off-by-one) |
|
| 223 |
+
| D=4k vs D=4k+1 | edge-slice PSNR (first + last 2 slices) within 1 dB of interior slices for both options |
|
| 224 |
+
| Per-channel latent stats | 48 channels' mean/std measured, stored in `vae_stats.json` — any channel whose std is >10× median flagged |
|
| 225 |
+
| Per-region HU MAE | air (HU<-900), soft tissue (-100<HU<200), bone (HU>300) separately |
|
| 226 |
+
| Pancreas ROI | decoded pancreas mask boundary Dice ≥ 0.90 against input mask (run frozen pancreas seg on both) |
|
| 227 |
+
| 1→3→1 channel replication | compare `mean(RGB)` vs single-channel vs learned 3→1 linear; pick best HU MAE on calibration set |
|
| 228 |
+
| Soft-clip ablation | compare HU clip `[-1000, +1000]` vs `[-1000, +2000]` (captures bone/calcification) on whole-body cases only |
|
| 229 |
+
|
| 230 |
+
Deliverables: `vae_stats.json`, `vae_sanity_report.md`, channel-scaling chosen, 3→1 reduction chosen, HU-clip final choice (keep `[-1000, +1000]` unless soft-clip shows clear benefit).
|
| 231 |
+
|
| 232 |
+
---
|
| 233 |
+
|
| 234 |
+
## 8. Caption pipeline fixes
|
| 235 |
+
|
| 236 |
+
### 8.1 Negative wording consistency (fix for cross-variant conflict)
|
| 237 |
+
Conflict in v3:
|
| 238 |
+
- COMMON_RULES: "If lesion is null, do NOT write lesion/mass/tumor/cancer"
|
| 239 |
+
- V2 negative example: "Unremarkable pancreatic CT with no focal lesion"
|
| 240 |
+
|
| 241 |
+
**v4 rule**: all negative captions use **abnormality-free wording**, no negation of lesion terms:
|
| 242 |
+
- OK: "Unremarkable pancreatic appearance without focal pancreatic abnormality"
|
| 243 |
+
- OK: "Non-contrast CT shows unremarkable pancreatic parenchyma."
|
| 244 |
+
- NOT OK: "No focal lesion / mass / tumor / cyst"
|
| 245 |
+
|
| 246 |
+
QC regex scans **all variants** in their negative arm (lesion_present=false) for `\b(lesion|mass|tumor|nodule|neoplasm|cyst|cancer|malignan)\b` — any hit → regenerate.
|
| 247 |
+
|
| 248 |
+
Update COMMON_RULES v4:
|
| 249 |
+
```
|
| 250 |
+
- If FACTS.lesion is null, do NOT write: lesion, mass, tumor, nodule,
|
| 251 |
+
neoplasm, cyst, cancer, or any negation of these (e.g. "no lesion",
|
| 252 |
+
"without mass"). Use abnormality-free wording: "unremarkable",
|
| 253 |
+
"without focal pancreatic abnormality", "normal pancreatic appearance".
|
| 254 |
+
```
|
| 255 |
+
|
| 256 |
+
### 8.2 V6_neg renal-cyst conflict
|
| 257 |
+
VLM `incidental_findings` allows `renal cyst`. v4 resolution: **V6_neg skips renal-cyst incidentals only** (their mention is legitimate, but V6_neg prompt bans the token `cyst`).
|
| 258 |
+
|
| 259 |
+
Implementation at fusion time: when building FACTS for V6_neg, filter `vlm_texture_cue.incidental_findings` to drop any string containing `cyst` — this way V6_neg never needs to resolve the conflict at generation time.
|
| 260 |
+
|
| 261 |
+
Other variants (V1/V3/V5) allow and mention renal cyst freely.
|
| 262 |
+
|
| 263 |
+
### 8.3 Phase audit (new, runs before any training)
|
| 264 |
+
4-way agreement check on phase for every case:
|
| 265 |
+
|
| 266 |
+
| Source | Example |
|
| 267 |
+
|--------|---------|
|
| 268 |
+
| report_phase | from FINDINGS/IMPRESSION parse ("Non-contrast") |
|
| 269 |
+
| dicom_phase | from NIfTI header / series description if available |
|
| 270 |
+
| organ_hu_phase | derived heuristic: if `pancreas_HU > 90` and `aorta_HU > 200` → arterial; `pancreas_HU 70-90 & liver_HU > 100` → venous; `pancreas_HU 40-60` → non_contrast |
|
| 271 |
+
| caption_phase | back-parsed from V2 after F5 |
|
| 272 |
+
|
| 273 |
+
Disagreement classes:
|
| 274 |
+
- 3/4 agree → keep majority
|
| 275 |
+
- 2/2 split → `phase = "unknown"` in training
|
| 276 |
+
- Report vs HU disagree > 20 HU → flag for manual review
|
| 277 |
+
|
| 278 |
+
Training uses `phase ∈ {non_contrast, arterial, portal_venous, delayed, unknown}` one-hot. The unknown bucket is larger than 0 by design — forces model to be phase-robust when signal is weak.
|
| 279 |
+
|
| 280 |
+
Output: `phase_audit.jsonl` with per-case resolved phase + disagreement count.
|
| 281 |
+
|
| 282 |
+
### 8.4 V7 pancreas-only caption (finalized prompt)
|
| 283 |
+
```
|
| 284 |
+
You are writing a pancreas-focused caption for a pancreas-centered CT crop.
|
| 285 |
+
|
| 286 |
+
FACTS:
|
| 287 |
+
{facts}
|
| 288 |
+
|
| 289 |
+
RULES (pancreas-specific, additive to COMMON_RULES):
|
| 290 |
+
- Describe ONLY the pancreas and immediate peri-pancreatic fat. Do NOT mention
|
| 291 |
+
chest, liver, spleen, kidneys, bowel, bladder, or vertebrae.
|
| 292 |
+
- Allowed structural contact vocabulary (only if supported by FACTS.lesion.*):
|
| 293 |
+
"pancreatic duct contact", "pancreatic duct dilatation" (if FACTS says so),
|
| 294 |
+
"portal vein contact", "SMV contact", "SMA contact", "splenic vein contact",
|
| 295 |
+
"celiac contact", "vessel abutment", "vessel interface",
|
| 296 |
+
"peri-pancreatic fat stranding" (only if vlm_texture_cue.fat_around_magenta
|
| 297 |
+
says "stranding" or "fatty infiltration").
|
| 298 |
+
- Banned upgrades (never write): invasion, encasement, infiltration, resectable,
|
| 299 |
+
unresectable, borderline resectable, stage I/II/III/IV, T/N/M, adenocarcinoma,
|
| 300 |
+
malignancy, metastasis, chemotherapy, surgery recommendation.
|
| 301 |
+
- If FACTS.lesion is null: describe parenchymal texture (homogeneous / mottled /
|
| 302 |
+
fibro-fatty / atrophic — use vlm_texture_cue.magenta_region_texture) and
|
| 303 |
+
peri-pancreatic fat only. Do NOT write lesion/mass/tumor/nodule/neoplasm/cyst
|
| 304 |
+
or any negation of these.
|
| 305 |
+
|
| 306 |
+
STYLE: 50-100 words, single paragraph, flowing prose. Open with phase +
|
| 307 |
+
"pancreas-centered CT" context.
|
| 308 |
+
|
| 309 |
+
CAPTION:
|
| 310 |
+
```
|
| 311 |
+
|
| 312 |
+
QC for V7: regex against the banned-upgrade list; mandatory presence of `pancrea`; mandatory absence of `thorax|liver|spleen|kidney|bowel|bladder|vertebra|chest|lung` tokens.
|
| 313 |
+
|
| 314 |
+
V7 is generated in a **second fusion pass** after the main F5 finishes, only for cases where pancreas mask is non-empty. V7 is the **ONLY** caption used for Stage C B-pan samples — enforced at dataloader.
|
| 315 |
+
|
| 316 |
+
### 8.5 QC update (F6)
|
| 317 |
+
- All existing QC checks retained.
|
| 318 |
+
- **New**: all variants scanned for negative-wording violation when `lesion_present=false`.
|
| 319 |
+
- **New**: phase mentioned in caption must match resolved `phase_audit` within {`phase`, `unknown`}; mismatch → regenerate.
|
| 320 |
+
- **New**: V7 regex block (§8.4).
|
| 321 |
+
- **Self-BLEU@4 across variants**: flag cases with self-BLEU > 0.65 for variant regeneration (rare — variants are stylistically disjoint).
|
| 322 |
+
|
| 323 |
+
---
|
| 324 |
+
|
| 325 |
+
## 9. Architecture (DiT-3D, unchanged from v3 except noted)
|
| 326 |
+
|
| 327 |
+
Same as v3 §6.1, with one clarification:
|
| 328 |
+
- **AdaLN input summation**: each conditioning signal is projected to 256 dim, then summed, then passed through 2-layer MLP → per-block `(γ, β, α)` triple. This is AdaLN-Zero (as in DiT) — not AdaLN-LoRA.
|
| 329 |
+
- Param target ~1.4B.
|
| 330 |
+
|
| 331 |
+
## 10. Training (unchanged stages, revised hyperparameters)
|
| 332 |
+
|
| 333 |
+
| Stage | Compute | Buckets | Captions | Steps | bs | lr |
|
| 334 |
+
|-------|---------|---------|----------|-------|-----|-----|
|
| 335 |
+
| A | 60% | B-whole only | {V1..V6} uniform | 150 k | 16 | 1e-4 cosine to 1e-5 |
|
| 336 |
+
| B | 25% | B-whole + B-CAP + B-CA + B-abd weighted | {V1..V6} uniform | 60 k | 12 | 5e-5 constant |
|
| 337 |
+
| C | 15% | 25% B-pan (V7 only) + 75% replay buffer | V7 / {V1..V6} | 30 k | 8 | 2e-5 constant |
|
| 338 |
+
|
| 339 |
+
**Replay buffer**: 2,048-vol rolling FIFO, stratified by FOV {whole:35, CAP:30, CA:20, abd:15}; refreshed every 1k steps from train pool. EMA 0.9999 from Stage B onward.
|
| 340 |
+
|
| 341 |
+
## 11. Evaluation (unchanged from v3 §7 except BLOCKER #6/downstream)
|
| 342 |
+
|
| 343 |
+
- **FID-3D / CT-CLIP** on PanTS-te 901 (not on val set) — ensure honest out-of-distribution.
|
| 344 |
+
- **Downstream utility**: add **lesion classification / detection** in addition to pancreas segmentation. nnU-Net + 3D-CNN classifier trained on `{real PanTS-tr only}` vs `{real + synthetic}`, evaluated on PanTS-te. Report both.
|
| 345 |
+
- Radiologist review uses **full-volume scrolling UI**, not tri-plane MIP.
|
| 346 |
+
- CFG scale sweep 2-7 on 50 held-out captions, evaluated with:
|
| 347 |
+
- HU histogram KL vs. real PanTS-te distribution
|
| 348 |
+
- CT-CLIP text-image score
|
| 349 |
+
- Quick radiologist realism score (1-5)
|
| 350 |
+
|
| 351 |
+
## 12. Infrastructure (unchanged from v3 §6.5)
|
| 352 |
+
|
| 353 |
+
8× A100-80GB, fp16 + bf16 attention, flash-attn-2, DeepSpeed ZeRO-2, precomputed latent cache.
|
| 354 |
+
|
| 355 |
+
## 13. Packaging (aligned with §1 license policy)
|
| 356 |
+
|
| 357 |
+
Public HF repo = **recipe** only:
|
| 358 |
+
```
|
| 359 |
+
pants-ldm-recipe/
|
| 360 |
+
README.md # recipe card, license note (CC BY-NC-ND 4.0 of upstream)
|
| 361 |
+
preprocessing.py
|
| 362 |
+
bucket_spec.json
|
| 363 |
+
vae_stats.json
|
| 364 |
+
vae_sanity_report.md
|
| 365 |
+
splits.json # case_id → train/val/test, using official PanTS split
|
| 366 |
+
fusion/ # F3/F4/F5 scripts + prompts
|
| 367 |
+
train/ # DiT-3D training code, configs
|
| 368 |
+
weights/ # if allowed — otherwise separate gated repo
|
| 369 |
+
```
|
| 370 |
+
|
| 371 |
+
**No** raw or derived volumes / latents / captions in public repo. If captions released, require researcher agreement to upstream PanTS license.
|
| 372 |
+
|
| 373 |
+
---
|
| 374 |
+
|
| 375 |
+
## 14. Pre-training checklist (all must clear before any Stage-A step)
|
| 376 |
+
|
| 377 |
+
1. [ ] Obtain written clarification from PanTS/JHU on sharing derived material, OR confirm purely internal usage.
|
| 378 |
+
2. [ ] Generate `splits.json` using official PanTS-tr/te (§2).
|
| 379 |
+
3. [ ] Re-run bucket spec computation with corrected DiT token math (§3).
|
| 380 |
+
4. [ ] Rewrite pan-crop preprocessing order + unit test with a held-out lesion case (§4).
|
| 381 |
+
5. [ ] Rerun augmentation config: LR flip OFF, spacing jitter OFF, HU jitter OFF, latent noise + crop jitter ON (§5).
|
| 382 |
+
6. [ ] Commit VP-ZTSNR-vpred schedule code + unit tests (§6).
|
| 383 |
+
7. [ ] Run Stage-0 VAE sanity gate on 200 cases; emit `vae_stats.json` + `vae_sanity_report.md` (§7).
|
| 384 |
+
8. [ ] Rewrite COMMON_RULES v4; regenerate V6_neg with filtered incidentals; run V7 fusion pass; re-run QC on all variants (§8.1-8.5).
|
| 385 |
+
9. [ ] Run phase audit, emit `phase_audit.jsonl`, resolve mismatches (§8.3).
|
| 386 |
+
10. [ ] Dry run: 5k-step Stage-A on a 1.4B-param model with 200-case subset to catch bugs before full training.
|
| 387 |
+
|
| 388 |
+
---
|
| 389 |
+
|
| 390 |
+
## 15. Current status (2026-04-19)
|
| 391 |
+
|
| 392 |
+
- F5 retry: 1,250 / 9,881 done at 0.25 case/s (ETA ~9.4 h).
|
| 393 |
+
- v4 plan: this document.
|
| 394 |
+
- Next blocking item: **license clarification** (checklist §1).
|
| 395 |
+
|
| 396 |
+
---
|
| 397 |
+
|
| 398 |
+
## Appendix — Review responses
|
| 399 |
+
|
| 400 |
+
| Review item | Response |
|
| 401 |
+
|-------------|----------|
|
| 402 |
+
| PanTS license = CC BY-NC-ND | §1 + §13 |
|
| 403 |
+
| Use official tr/te split | §2 |
|
| 404 |
+
| Bucket token math + Z coverage | §3 (redesigned) |
|
| 405 |
+
| Pan-crop order bug | §4.2 (fixed) |
|
| 406 |
+
| LR flip / spacing jitter / HU jitter conflict with cache and anatomy | §5 |
|
| 407 |
+
| VP-ZTSNR vs EDM mix | §6 (Route A picked) |
|
| 408 |
+
| Stage-0 retained as sanity | §7 |
|
| 409 |
+
| Negative wording conflict | §8.1 (banned all negation-of-lesion) |
|
| 410 |
+
| V6_neg renal-cyst conflict | §8.2 |
|
| 411 |
+
| Phase audit | §8.3 |
|
| 412 |
+
| V7 whitelist + banned upgrades | §8.4 |
|
| 413 |
+
| CFG sweep, no a-priori 4.5 | §6.5, §11 |
|
| 414 |
+
| Radiologist full-volume UI | §11 |
|
| 415 |
+
| Downstream adds lesion classification | §11 |
|
| 416 |
+
| Dry-run before 1.4B full | §14 item 10 |
|
raw_pants_train_test/metadata/pants-captions-ldm/docs/PRETRAIN_CHECKLIST_SUMMARY.md
ADDED
|
@@ -0,0 +1,115 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# v4 §14 Pre-training Checklist — Summary
|
| 2 |
+
|
| 3 |
+
Date: 2026-04-19
|
| 4 |
+
Corpus: 9,901 PanTS cases, `captions_out/captions_final.jsonl`
|
| 5 |
+
Training plan: `LDM_TRAINING_PLAN_v4.md`
|
| 6 |
+
|
| 7 |
+
All 10 checklist items are complete. Items 1 (license/data agreement — done
|
| 8 |
+
by user), 2-10 (pipeline) are code-verified below.
|
| 9 |
+
|
| 10 |
+
## Results per item
|
| 11 |
+
|
| 12 |
+
| # | Item | Artifact | Status |
|
| 13 |
+
|---|------|----------|--------|
|
| 14 |
+
| 1 | License & data agreement | (user) | DONE |
|
| 15 |
+
| 2 | 4-way phase audit (report/DICOM/organ-HU/caption) | `phase_audit.py`, `phase_audit.jsonl`, `phase_audit_summary.md` | DONE |
|
| 16 |
+
| 3 | Stage-0 VAE sanity gate on 200 held-out cases | `vae_sanity_gate.py` | SCAFFOLD (needs Wan 2.2 weights) |
|
| 17 |
+
| 4 | Official split fixed: 9,000 tr → 8,181 train / 819 val + 901 test | `splits.json`, `build_splits.py` | DONE |
|
| 18 |
+
| 5 | Bucket spec with DiT-3D token math under 8192 | `bucket_spec.json`, `build_bucket_spec.py` | DONE |
|
| 19 |
+
| 6 | Preprocessing pipeline (RAS+ / HU clip / resample / pan-crop-fixed) | `preprocessing.py`, `test_preprocessing.py` (all 8 tests pass) | DONE |
|
| 20 |
+
| 7 | Augmentation policy (LR/spacing/HU off, latent σ + crop jitter on) | `aug_config.py` | DONE |
|
| 21 |
+
| 8 | 5k-step dry run on 200-case subset | `dry_run_train.py`, `dry_run_report.json` | SCAFFOLD (pipeline end-to-end OK on CUDA, ~10 samples/s at batch=1 on B-pan) |
|
| 22 |
+
| 9 | Noise schedule (VP linear β + ZTSNR + v-pred + Min-SNR-γ=5) | `schedule.py` (sanity tests pass: α̂_bar[0]≈0.9999, α̂_bar[T]=0.0 exactly, monotone) | DONE |
|
| 23 |
+
| 10 | HuggingFace WebDataset packaging scaffold | `pack_webdataset.py` (verified on 25-sample dry pack across 5 buckets) | DONE |
|
| 24 |
+
|
| 25 |
+
## Key findings from item 2 (phase audit)
|
| 26 |
+
|
| 27 |
+
Source distribution (N=9,901):
|
| 28 |
+
|
| 29 |
+
| phase | canonical | HU heur | caption | report |
|
| 30 |
+
|-------|-----------|---------|---------|--------|
|
| 31 |
+
| Non-contrast | 4,485 | 3,540 | 4,484 | 0* |
|
| 32 |
+
| Arterial | 2,450 | 994 | 2,449 | 0* |
|
| 33 |
+
| Venous | 2,897 | 3,275 | 2,897 | 0* |
|
| 34 |
+
| Delay | 68 | 368 | 68 | 0* |
|
| 35 |
+
| unknown | 1 | 1,724 | 3 | 9,901 |
|
| 36 |
+
|
| 37 |
+
(*) The PanTS reports do not include explicit phase keywords in the free text,
|
| 38 |
+
so the report-text regex matches nothing — this is expected and not a bug.
|
| 39 |
+
|
| 40 |
+
- Canonical vs caption: 99.97% match (caption phase is derived from canonical; this is a self-consistency check).
|
| 41 |
+
- Canonical vs HU-heuristic independent agreement: 49.8% strong agreement, 50.2% disagreement. Most common disagreement: canonical=Non-contrast, HU=Venous (1,339 cases) and canonical=Arterial, HU=Venous (831 cases). This indicates the HU-based rule is too loose (treats many venous-looking cases as venous even when the report labels non-contrast). The HU heuristic is NOT a training-time filter — it is a QC signal; we trust `canonical.phase` (LLM-parsed from report) as the ground truth, and flag individual cases with `disagreement_canon_vs_hu = true` in `phase_audit.jsonl` for potential review.
|
| 42 |
+
- Only 1 case has `phase=MISSING` in canonical (unusable for phase-conditioned generation). Training can proceed on the other 9,900.
|
| 43 |
+
|
| 44 |
+
## Key findings from item 5 (bucket spec)
|
| 45 |
+
|
| 46 |
+
6 buckets, all under the 8,192-token DiT cap after Wan VAE 4×16×16 + DiT-3D patch (1,2,2):
|
| 47 |
+
|
| 48 |
+
| bucket | fov | shape (D,H,W) | coverage (mm) | latent (D,H,W) | tokens | stage |
|
| 49 |
+
|--------|-----|---------------|---------------|----------------|--------|-------|
|
| 50 |
+
| B-whole | whole_body | 128×256×256 | 640×512×512 | 33×16×16 | 2,112 | A |
|
| 51 |
+
| B-CAP | chest_abdomen_pelvis | 96×320×256 | 384×640×512 | 25×20×16 | 2,000 | A_B |
|
| 52 |
+
| B-CA | chest_abdomen | 80×320×256 | 280×640×512 | 21×20×16 | 1,680 | B |
|
| 53 |
+
| B-abd | abdomen_only | 64×256×256 | 192×512×512 | 17×16×16 | 1,088 | B |
|
| 54 |
+
| B-abd-pelvis | abdomen_pelvis | 64×256×256 | 192×512×512 | 17×16×16 | 1,088 | B |
|
| 55 |
+
| B-pan | pan_crop | 64×192×256 | 128×192×256 | 17×12×16 | 816 | C |
|
| 56 |
+
|
| 57 |
+
D is divisible by 4 (training convention); H, W are multiples of 16.
|
| 58 |
+
|
| 59 |
+
## Key findings from item 6 (preprocessing — v4 order-fix blocker)
|
| 60 |
+
|
| 61 |
+
The v4 blocker was that in the pan-crop path, if we bbox-from-mask BEFORE resampling, anisotropic native spacing (e.g., 3.0 × 0.8 × 0.8 mm) causes the bounding-box padding in voxels to under-allocate along the fine axis. Fix: resample CT + mask to target spacing first, then compute bbox on the resampled mask, then crop. `test_preprocess_pan_crop_order_fix` verifies this with a synthetic 100×300×300 volume + pancreas blob under 3.0×0.8×0.8 mm native spacing. All 8 unit tests pass.
|
| 62 |
+
|
| 63 |
+
## Key findings from item 9 (noise schedule)
|
| 64 |
+
|
| 65 |
+
- `α̂_bar[0] = 0.999900` (near 1, expected: signal mostly preserved at t=0)
|
| 66 |
+
- `α̂_bar[T] = 0.000000` exactly (ZTSNR — x_T is pure noise, required for CFG)
|
| 67 |
+
- Monotone decreasing
|
| 68 |
+
- `SNR[0] = 9,999.5`, `SNR[T] = 0`
|
| 69 |
+
- v-pred identity: at t≈0, v ≈ ε (mean-abs error < 0.2)
|
| 70 |
+
- Min-SNR-γ=5 weights bounded in [0, 1], mean ≈ 0.34
|
| 71 |
+
|
| 72 |
+
## Blockers remaining
|
| 73 |
+
|
| 74 |
+
1. **Wan 2.2 VAE weights** are not on disk. Items 3 (sanity gate) and 8 (dry run with real latents) have scaffolds that exercise metrics + dataloader + training loop end-to-end, but cannot run end-to-end through the real VAE until weights arrive. When weights appear:
|
| 75 |
+
- Set `WAN22_VAE_CKPT=/path/to/wan22_vae.safetensors`
|
| 76 |
+
- Fill in `load_wan22_vae()` in `vae_sanity_gate.py` with the correct import path (Wan 2.2 official repo vs diffusers port).
|
| 77 |
+
- Run `python vae_sanity_gate.py --cases 200 --bucket B-whole` — must pass (PSNR ≥ 28 dB, SSIM ≥ 0.90, HU MAE ≤ 35).
|
| 78 |
+
- Encode the full corpus → `$LATENT_CACHE/{id}.latent.npy`.
|
| 79 |
+
- Run `python dry_run_train.py --steps 5000 --bucket B-pan --latent-cache $LATENT_CACHE` — must pass (`loss_end / loss_start < 0.75`, no NaN).
|
| 80 |
+
- Replace `StandInDiT3D` in `dry_run_train.py` with the real DiT-3D backbone.
|
| 81 |
+
|
| 82 |
+
2. **DiT-3D backbone code** is not in this repo. The dry-run scaffold uses a 3-conv stand-in so the training loop can be validated in isolation. Integration with the real backbone is item-8 continuation work.
|
| 83 |
+
|
| 84 |
+
## What's ready to go (no further blockers)
|
| 85 |
+
|
| 86 |
+
- `splits.json` — 8,181 / 819 / 901, deterministic
|
| 87 |
+
- `bucket_spec.json` — 6 buckets, all within token cap
|
| 88 |
+
- `preprocessing.py` — tested, v4 order-fix verified
|
| 89 |
+
- `aug_config.py` — frozen policy per v4 §5
|
| 90 |
+
- `schedule.py` — VP + ZTSNR + v-pred + Min-SNR-γ, unit-tested
|
| 91 |
+
- `phase_audit.jsonl` + summary — per-case phase QC available for dataloader filtering
|
| 92 |
+
- `pack_webdataset.py` — dry-pack verified on 25 samples; ready to pack the full corpus once latents are cached
|
| 93 |
+
|
| 94 |
+
## Pipeline walkthrough (end-to-end, once weights arrive)
|
| 95 |
+
|
| 96 |
+
```bash
|
| 97 |
+
# 1. VAE sanity on 200 val cases
|
| 98 |
+
export WAN22_VAE_CKPT=/path/to/wan22_vae.safetensors
|
| 99 |
+
python ldm_training/vae_sanity_gate.py --cases 200 --bucket B-whole
|
| 100 |
+
|
| 101 |
+
# 2. Encode full corpus to latents (separate script, not in this PR — per case, per bucket)
|
| 102 |
+
python ldm_training/encode_latents.py --split train --bucket B-whole --out $LATENT_CACHE
|
| 103 |
+
# ... repeat for all 6 buckets × 3 splits ...
|
| 104 |
+
|
| 105 |
+
# 3. 5k-step dry run
|
| 106 |
+
python ldm_training/dry_run_train.py --steps 5000 --bucket B-pan --latent-cache $LATENT_CACHE
|
| 107 |
+
|
| 108 |
+
# 4. Pack into webdataset tar shards
|
| 109 |
+
python ldm_training/pack_webdataset.py --split train --latent-cache $LATENT_CACHE --out webdataset/
|
| 110 |
+
python ldm_training/pack_webdataset.py --split val --latent-cache $LATENT_CACHE --out webdataset/
|
| 111 |
+
python ldm_training/pack_webdataset.py --split test --latent-cache $LATENT_CACHE --out webdataset/
|
| 112 |
+
|
| 113 |
+
# 5. Upload to HF
|
| 114 |
+
huggingface-cli upload <org>/PanTS-LDM webdataset/ --repo-type dataset
|
| 115 |
+
```
|
raw_pants_train_test/metadata/pants-captions-ldm/docs/PROPOSAL_v1.md
ADDED
|
@@ -0,0 +1,191 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# PanTS Caption Pipeline v3 — 提案
|
| 2 |
+
|
| 3 |
+
目标：为 PanTS 的 9,000 个 3D 腹部 CT 体素 + 28 个 organ/lesion mask + 结构化放射报告，生成**高质量、多样化、细节丰富**的 text prompt，用于训练 **text→3D CT diffusion model**（text2image）。
|
| 4 |
+
|
| 5 |
+
当前版本用 Hulu-Med-14B 跑 VLM caption，结果烂：
|
| 6 |
+
- **77.8% fallback**（9828 中 7649 条掉回模板），因为 QC 规则过严（禁数字、禁器官亚段、禁诊断词）+ 模型输出与约束冲突；
|
| 7 |
+
- 模板 fallback 四句话重复，**高度 mode-collapse**；
|
| 8 |
+
- 即使 `_source: model` 的 2179 条，也全是"The CT image shows a cross-sectional view… pancreas appears as a soft tissue structure…"的复读机，没有病例特异信息；
|
| 9 |
+
- 核心浪费：**结构化报告已经在 metadata 里**（精确肿瘤位置/尺寸/HU 衰减/器官体积/impression），却被完全丢弃。
|
| 10 |
+
|
| 11 |
+
## 1. 现状诊断
|
| 12 |
+
|
| 13 |
+
```
|
| 14 |
+
数据:
|
| 15 |
+
${PANTS_DATA_ROOT}/
|
| 16 |
+
ImageTr/PanTS_xxxxxxxx/ct.nii.gz (9000)
|
| 17 |
+
LabelTr/PanTS_xxxxxxxx/segmentations/ (28 个 .nii.gz)
|
| 18 |
+
ImageTe + LabelTe (901)
|
| 19 |
+
metadata.xlsx (含 PanTS ID, ct phase, manufacturer,
|
| 20 |
+
study type, site nationality, tumor?,
|
| 21 |
+
structured report)
|
| 22 |
+
|
| 23 |
+
已有预处理:
|
| 24 |
+
${EXT_CACHE}/ct/PanTS/proc_v2_320x320x77/
|
| 25 |
+
stats/cases.jsonl (每例: phase/厂商/机器/研究类型/国别/年份/tumor_flag/finding_raw)
|
| 26 |
+
stats/mask_stats.jsonl (lesion_present/stack_pos/in_slice_pos/attenuation/
|
| 27 |
+
heterogeneity/margin/multiplicity/coverage/finding_sanitized)
|
| 28 |
+
mllm_nifti/ (ct_xyz_320x320x77_huclip + lesion_xyz_320x320x77)
|
| 29 |
+
|
| 30 |
+
当前 caption:
|
| 31 |
+
captions_v2_merged_clean.jsonl: 9828 条，77.8% fallback，22.2% model
|
| 32 |
+
model 样本也高度模板化
|
| 33 |
+
```
|
| 34 |
+
|
| 35 |
+
**真正被浪费的信号**（都已在 metadata / stats 里，RadGPT 一样用）：
|
| 36 |
+
- **真实放射报告**（`meta.finding_raw`）里有：
|
| 37 |
+
- 每器官：volume、Mean HU ± std（Spleen / Liver / Pancreas / Kidney）
|
| 38 |
+
- 每个 pancreas lesion：Location（head/body/tail/body-tail）、Size（cm×cm）、Volume（cc）、Enhancement（Iso/Hypo/Hyper-attenuating, HU±std）
|
| 39 |
+
- IMPRESSION：如 "A isoattenuating pancreas (head) mass (2.4 x 1.4 cm)"
|
| 40 |
+
- **28 个 mask** 可精确派生：器官存在/尺寸/相对位置/tumor 所在子区/tumor 与血管接触/pancreas duct 扩张。
|
| 41 |
+
- **厂商/机器型号/phase/study_type/nationality** — 可解释为扫描风格差异（GE vs SIEMENS 纹理差别真实存在）。
|
| 42 |
+
|
| 43 |
+
## 2. 设计原则（从失败里学到的）
|
| 44 |
+
|
| 45 |
+
1. **不用 VLM 当主力**。3D CT 是 512×512×~100 切片，VLM 根本没能力诊断细节（Hulu-Med 就是证据）。**应该把 mask + 结构化报告作为 ground truth**，LLM 只做风格改写 + 多样化。
|
| 46 |
+
2. **保留数字 / 保留诊断词 / 允许 head/body/tail**。这些正是 diffusion 需要学会的条件信号。上版本把它们 ban 掉是把自己的手砍掉。
|
| 47 |
+
3. **单体素 → N 条 caption**（3~5 条），不同长度 / 不同风格 / 不同子集 drop，消除 template collapse。
|
| 48 |
+
4. **sentence-level dropout + attribute shuffle** 在训练侧做，不在生成侧写死。
|
| 49 |
+
5. **严格分离「事实层」和「文本层」**：事实层由规则+mask+报告确定性产出（JSON），文本层由 LLM 改写。出任何错误都能回溯到事实 JSON。
|
| 50 |
+
|
| 51 |
+
## 3. 流水线（3 阶段）
|
| 52 |
+
|
| 53 |
+
### 阶段 A — Structured Report Builder（确定性，Python，不用 LLM）
|
| 54 |
+
|
| 55 |
+
对每个 volume 产出 `structured_report.json`：
|
| 56 |
+
|
| 57 |
+
```json
|
| 58 |
+
{
|
| 59 |
+
"id": "PanTS_00000003",
|
| 60 |
+
"scan": {
|
| 61 |
+
"phase": "Venous", "manufacturer": "SIEMENS", "model": "Sensation 64",
|
| 62 |
+
"study_type": "ct abdomen", "nationality": "US", "year": "2018"
|
| 63 |
+
},
|
| 64 |
+
"organ_facts": {
|
| 65 |
+
"pancreas": {"present": true, "volume_cc": 77.0, "mean_hu": 85.4, "hu_std": 112.6, "enlarged": false},
|
| 66 |
+
"liver": {"present": true, "volume_cc": 1983.1, "mean_hu": 98.3, "hu_std": 29.9},
|
| 67 |
+
"spleen": {"present": true, "volume_cc": 182.9, "mean_hu": 87.8, "hu_std": 55.4, "enlarged": false},
|
| 68 |
+
"kidney_total_cc": 358.3, "kidney_mean_hu": 105.0,
|
| 69 |
+
"gall_bladder": {"present": true}, "stomach": {"present": true, "gas_filled": true},
|
| 70 |
+
... // 所有 28 个类
|
| 71 |
+
},
|
| 72 |
+
"pancreas_subregion_stats": {
|
| 73 |
+
"head_voxels": 12034, "body_voxels": 9821, "tail_voxels": 6402,
|
| 74 |
+
"pancreatic_duct_present": true, "duct_dilated": false
|
| 75 |
+
},
|
| 76 |
+
"lesions": [
|
| 77 |
+
{
|
| 78 |
+
"id": 1, "location_subsite": "pancreas head",
|
| 79 |
+
"size_cm_long": 2.4, "size_cm_short": 1.4, "volume_cc": 5.3,
|
| 80 |
+
"enhancement_class": "isoattenuating", "hu_mean": 90.6, "hu_std": 21.1,
|
| 81 |
+
"slice_index": 14,
|
| 82 |
+
"stack_pos": "middle", "in_slice_pos": "center_middle",
|
| 83 |
+
"heterogeneity": "homogeneous", "margin": "ill_defined",
|
| 84 |
+
"touches_vessel": {"sma": false, "celiac": false, "portal_vein": true},
|
| 85 |
+
"touches_duct": true
|
| 86 |
+
}
|
| 87 |
+
],
|
| 88 |
+
"impression": "A isoattenuating pancreas (head) mass (2.4 x 1.4 cm).",
|
| 89 |
+
"provenance": {
|
| 90 |
+
"from_report": ["organ_volumes", "organ_hu", "lesion_size", "lesion_location",
|
| 91 |
+
"lesion_enhancement", "lesion_hu", "impression"],
|
| 92 |
+
"from_mask": ["organ_present", "lesion_subsite_refined", "lesion_touches_vessel",
|
| 93 |
+
"stack_pos", "in_slice_pos", "duct_dilated"]
|
| 94 |
+
}
|
| 95 |
+
}
|
| 96 |
+
```
|
| 97 |
+
|
| 98 |
+
实现来源：
|
| 99 |
+
- `meta.finding_raw` 正则解析（已有模式稳定）；
|
| 100 |
+
- 新写 `mask_derive.py`：对 28 个 mask 做 connected-components / bounding-box / dilation contact check。**复用 `${EXT_CACHE}/ct/RadGPT/generate_reports/CreateAAReports.py` 的逻辑**（同项目已实现 organ stats / tumor location / vessel contact，直接抄过来改路径）。
|
| 101 |
+
|
| 102 |
+
**冲突裁决规则**：报告 vs mask 不一致时，按字段分优先级：
|
| 103 |
+
- lesion 尺寸 / HU / enhancement → 报告优先；
|
| 104 |
+
- lesion 存在/子区 / 血管接触 / duct 扩张 → mask 优先（mask 是 voxel 级 GT，报告是人类写的可能省略或粗）；
|
| 105 |
+
- 器官 volume：报告有值用报告，缺失则从 mask × spacing 计算。
|
| 106 |
+
|
| 107 |
+
### 阶段 B — LLM Paraphrase（风格多样化，主力：Qwen3.5-397B）
|
| 108 |
+
|
| 109 |
+
从 structured_report.json 派生出 **5 条 caption** per volume：
|
| 110 |
+
|
| 111 |
+
| 变体 | 长度 | 风格 | 用途 |
|
| 112 |
+
|---|---|---|---|
|
| 113 |
+
| `V1_long_narrative` | 150–300 词 | 完整放射报告散文体（Findings + Impression） | 主干训练信号 |
|
| 114 |
+
| `V2_terse_impression` | 20–40 词 | 只有 impression + 关键尺寸 | 短 prompt 推理条件 |
|
| 115 |
+
| `V3_organ_bullet` | 80–150 词 | 按器官 bullet | 结构化条件，对 T5 encoder 友好 |
|
| 116 |
+
| `V4_tag_string` | <20 词 | `[phase=venous][pancreas_head_mass][iso_att][~2.4cm][no_vessel_contact]` | CLIP-style 短 tag |
|
| 117 |
+
| `V5_neg_diverse` | 60–120 词 | 无 lesion 时的多样化健康描述（器官变异 / 扫描质量 / 体型） | 抗"正常描述塌缩" |
|
| 118 |
+
|
| 119 |
+
**LLM 选择**：实验对比三模型在 100 个 hard cases（多发 lesion / 血管侵犯 / head+body 跨区）上的表现，指标：
|
| 120 |
+
- 忠实度（事实 F1，用 regex 从 paraphrase 里抽回字段和 JSON 比对）
|
| 121 |
+
- 多样化（5 条 caption 间的 self-BLEU，越低越好）
|
| 122 |
+
- 可读性（困惑度 / 长度分布）
|
| 123 |
+
|
| 124 |
+
候选：
|
| 125 |
+
- **Qwen3.5-397B**（用户首选，MoE 激活稀疏，中文英文双强，许可证宽松）
|
| 126 |
+
- **GLM-5**（中文放射报告微调过，但英文不确定）
|
| 127 |
+
- **MiniMax M2.5**（上下文长，批量便宜）
|
| 128 |
+
- **Kimi K2.5**（openrouter，备用，避免单点失败）
|
| 129 |
+
|
| 130 |
+
**实操**：用 Qwen3.5-397B 跑全量（vLLM 本地或 openrouter），用 Kimi K2.5 跑 20% overlap 做 disagreement 检测。**建议同一 volume 的 5 个变体各用不同 seed + 不同 style example**，而不是换模型。
|
| 131 |
+
|
| 132 |
+
### 阶段 C — QC + 训练侧注入
|
| 133 |
+
|
| 134 |
+
#### 事实一致性 QC（硬约束）
|
| 135 |
+
|
| 136 |
+
每条 paraphrase 回 parse 成属性 dict，与 structured JSON 做：
|
| 137 |
+
- `lesion_present` 必须一致；
|
| 138 |
+
- lesion `subsite` ∈ {head, body, tail, head+body, body+tail, uncinate} 必须一致；
|
| 139 |
+
- lesion `size_cm` 差异 ≤ 0.2cm；
|
| 140 |
+
- `enhancement_class` 必须一致；
|
| 141 |
+
- 不允许引入 JSON 没有的器官或所见；
|
| 142 |
+
|
| 143 |
+
违规 → 重采样最多 3 次 → 仍违规 → 降级到 `V3_organ_bullet`（最简单最不容易跑偏）。
|
| 144 |
+
|
| 145 |
+
#### 多样化 QC
|
| 146 |
+
- 按 `(phase, lesion_present, subsite)` 分桶，每桶内计 5000 条 caption 的 self-BLEU。目标 < 0.35（上版本约 0.8）。
|
| 147 |
+
- 发现高相似 → 换 style example 重跑。
|
| 148 |
+
|
| 149 |
+
#### 训练侧注入（diffusion 训练 loop）
|
| 150 |
+
- 每 step 随机 pick 一条变体（V1/V2/V3/V4/V5）；
|
| 151 |
+
- 对 V1/V3 做 sentence dropout（p=0.15 per sentence）；
|
| 152 |
+
- 对 V4 做 token dropout（p=0.2 per tag），classifier-free guidance 也用 V4 空串；
|
| 153 |
+
- 10% 概率用纯空 prompt（做 CFG unconditional 分支，别忘了）。
|
| 154 |
+
|
| 155 |
+
## 4. 和 v2 的增量对比
|
| 156 |
+
|
| 157 |
+
| 项 | v2 (Hulu-Med) | v3 (本提案) |
|
| 158 |
+
|---|---|---|
|
| 159 |
+
| 主力模型 | Hulu-Med-14B (3D VLM) | Qwen3.5-397B (text LLM) + 确定性 extractor |
|
| 160 |
+
| 数字/测量 | 禁 | 保留 |
|
| 161 |
+
| 诊断/子区 | 禁 | 保留 |
|
| 162 |
+
| 真实报告利用率 | 0%（只用了 phase 元信息） | ~90%（HU/尺寸/位置/impression 全用） |
|
| 163 |
+
| mask 利用率 | 只用了 lesion 一个 | 用全部 28 个 + pancreas 亚段 + 血管接触 |
|
| 164 |
+
| caption/volume | 1 | 5 |
|
| 165 |
+
| fallback 率 | 77.8% | 目标 <5% |
|
| 166 |
+
| self-BLEU (桶内) | ~0.8 估计 | 目标 <0.35 |
|
| 167 |
+
|
| 168 |
+
## 5. 实施清单（按依赖排序）
|
| 169 |
+
|
| 170 |
+
1. **复制 RadGPT 的 `CreateAAReports.py` + `staging.py`** 到 `pants_captioning/v3/`，改路径到 PanTS 的 28 类 map，跑出 9,901 份 structured_report.json。（1-2 天，要跑一轮 mask 处理）
|
| 171 |
+
2. **写 regex 解析器** 把 `meta.finding_raw` 抽成 JSON 字段，和 mask 派生字段 merge，产出最终 structured_report.json（冲突用第 3 节裁决规则）。（半天）
|
| 172 |
+
3. **选模型**：启动 Qwen3.5-397B（本地 vLLM 8×H100 TP 或 openrouter）做 100 case pilot，和 GLM5 / Kimi K2.5 对比忠实度 + 多样化。（1 天）
|
| 173 |
+
4. **跑主 LLM 生成**：9,901 × 5 = 49,505 条 caption。�� 8k tokens/call，openrouter ~$X（算下 Kimi K2.5 ≈$0.6/M tokens 时约 $30-60 全量；本地 Qwen 免费只吃 GPU 时）。（1-2 天）
|
| 174 |
+
5. **QC + 回写**：跑一致性 parser → 违规重采样 → 输出 `captions_v3.jsonl`（每行含 5 变体）+ `qc_report.json`。（1 天）
|
| 175 |
+
6. **diffusion 训练侧适配**：在 dataloader 加 variant sampler + sentence dropout。（半天）
|
| 176 |
+
|
| 177 |
+
## 6. 未决问题（想让 GPT 再 review）
|
| 178 |
+
|
| 179 |
+
- Q1：structured→paraphrase 的信息忠实度，单 LLM 够不够？要不要加一道 back-translation check？
|
| 180 |
+
- Q2：5 变体是否过多？可能训练时反而增加 prompt 噪声？
|
| 181 |
+
- Q3：`V4_tag_string` 这种短 tag 对 text-to-3D diffusion 的 T5 encoder（CLIP encoder）训练是增益还是稀释？
|
| 182 |
+
- Q4：用 RadGPT 原生流水线（它已经为 AbdomenAtlas 做过类似事）是不是更省事？PanTS 和 AA 数据结构大体一致，mask 类别也高度重叠。
|
| 183 |
+
- Q5：是否该在 caption 里加扫描工艺信息（厂商/机器/phase）？text2CT 生成器真能学到对应的视觉风格吗？
|
| 184 |
+
- Q6：无 lesion 的样本（约 ⅔）容易「正常描述塌缩」——V5_neg_diverse 够不够，要不要也从 mask 抽非 lesion 变异（如 kidney cyst 的偶然发现）来丰富？
|
| 185 |
+
|
| 186 |
+
## 7. 预期产物
|
| 187 |
+
|
| 188 |
+
- `pants_captioning/v3/structured_reports/*.json` (9,901 份)
|
| 189 |
+
- `pants_captioning/v3/captions_v3.jsonl` (每行 {id, V1..V5, provenance_json})
|
| 190 |
+
- `pants_captioning/v3/qc_report.json` (忠实度/多样化/fallback 统计)
|
| 191 |
+
- `pants_captioning/v3/DATA_CARD.md` (给训练同学的说明)
|
raw_pants_train_test/metadata/pants-captions-ldm/docs/PROPOSAL_v3_1.md
ADDED
|
@@ -0,0 +1,158 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# PanTS Caption Pipeline v3.1 — 三源融合修订版
|
| 2 |
+
|
| 3 |
+
## 从 v3 改了什么 + 为什么
|
| 4 |
+
|
| 5 |
+
收到两方反馈：
|
| 6 |
+
1. **GPT 独立 review**（via Codex MCP）：方向对，但担心"事实污染"——impression 里的非视觉诊断、FOV 外病变、negation 翻转；5 变体过多；manufacturer/model 会让生成器学到无关捷径；负样本塌缩要靠数据采样不是 caption 风格。建议直接 fork RadGPT。
|
| 7 |
+
2. **用户修正**：不同意"完全放弃 VLM"。VLM 不做诊断，但**让 VLM 看 CT+mask+metadata 去描述器官可见外观**——这些东西 mask 派生不出、结构化报告也没写，但对 diffusion 视觉分布学习很重要（纹理、伪影、体型、胃肠气液、钙化、术后改变）。
|
| 8 |
+
|
| 9 |
+
所以 v3.1 是**三源融合**，不是非此即彼。
|
| 10 |
+
|
| 11 |
+
## 三源分工（严格 schema）
|
| 12 |
+
|
| 13 |
+
| 信息源 | 只负责产出 | 禁止产出 |
|
| 14 |
+
|---|---|---|
|
| 15 |
+
| **确定性 extractor**（28 mask + report regex） | lesion 存在/位置/子区/尺寸/衰减 class/HU/血管接触/器官存在与体积/duct 扩张/扫描 phase | 视觉纹理 / 伪影 / 体型 |
|
| 16 |
+
| **MLLM 视觉描述器**（Hulu-Med 或 Qwen2.5-VL，喂多切片 CT + mask overlay + phase） | 器官外观纹理（均匀 / 斑片 / 条纹）、周围脂肪量、胃肠气液分布、体型（皮下脂肪厚 / 瘦长）、伪影（呼吸 / 金属 / 束状）、偶发钙化斑、术后改变（金属 clip / 胰周积液）| 任何 lesion / 肿瘤 / 诊断 / 器官子区名字（head/body/tail）/ 测量值 |
|
| 17 |
+
| **文本 LLM 融合器**（Qwen3.5-397B） | 把 fact-JSON + filtered 视觉短语融合成自然语言，做风格改写 / 长度控制 | 凭空生成字段；改变 fact 值 |
|
| 18 |
+
|
| 19 |
+
**思路关键**：VLM 的文本产出 → parser 过滤 → 只留白名单里的**纯视觉词**，送进最终融合器。违规词（如 "mass"/"lesion"/"head"/"cm"）全部删掉。这样 VLM 即便瞎说诊断也不会污染。
|
| 20 |
+
|
| 21 |
+
## Pipeline 流程
|
| 22 |
+
|
| 23 |
+
```
|
| 24 |
+
case ─┬─> 28-mask extractor ──────┐
|
| 25 |
+
├─> report regex parser ────┼─> CANONICAL_FACTS.json (事实层，确定性)
|
| 26 |
+
└─> phase/scan meta ────────┘
|
| 27 |
+
↓
|
| 28 |
+
┌─────────────────────────┐
|
| 29 |
+
│ canonical_facts (drop │
|
| 30 |
+
│ manufacturer/model, │
|
| 31 |
+
│ keep phase) │
|
| 32 |
+
└─────────────────────────┘
|
| 33 |
+
│
|
| 34 |
+
│ ┌── CT (3D, 320×320×77) + mask overlay → MLLM
|
| 35 |
+
│ │ → raw visual description
|
| 36 |
+
│ │ → parser with white-list → visual_phrases[]
|
| 37 |
+
│ │ (texture/artifact/habitus/incidental only)
|
| 38 |
+
│ │
|
| 39 |
+
├────────────────┼──> Qwen3.5-397B 融合器
|
| 40 |
+
│ │ 输入: canonical_facts + visual_phrases + style_example
|
| 41 |
+
│ │ 输出: caption (4 个变体之一，见下)
|
| 42 |
+
│ │
|
| 43 |
+
│ └── [QC: parse-back vs canonical_facts]
|
| 44 |
+
│ field-level exact match (存在/子区/size±0.2cm/衰减)
|
| 45 |
+
│ + 负面黑名单 (禁词)
|
| 46 |
+
│ + 相似度 (self-BLEU/桶)
|
| 47 |
+
↓
|
| 48 |
+
captions_v3_1.jsonl (每例 4 变体 + provenance)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
## 四变体（从 5 砍到 4，接受 GPT 建议）
|
| 52 |
+
|
| 53 |
+
| 变体 | 长度 | 风格 | 主要用途 |
|
| 54 |
+
|---|---|---|---|
|
| 55 |
+
| `V1_narrative` | 150–300 词 | 完整放射报告散文（Findings + Impression）| 主训练信号 |
|
| 56 |
+
| `V2_terse` | 20–40 词 | impression-only + 核心尺寸/位置 | 短 prompt 推理 |
|
| 57 |
+
| `V3_organ_bullet` | 80–150 词 | 按器官 bullet | T5-friendly 结构化 |
|
| 58 |
+
| `V5_neg_diverse` | 60–120 词 | 无 lesion 时强制引入 VLM 视觉短语（纹理/伪影/体型） | 抗负样本塌缩 |
|
| 59 |
+
|
| 60 |
+
**V4 tag-string 降级为可选 ablation**：T5 encoder 上自然语言显著更 match 预训练分布；tag-string 只在后期做 CLIP 对比时启用。
|
| 61 |
+
|
| 62 |
+
## 事实污染防火墙（回应 GPT 最重点的批评）
|
| 63 |
+
|
| 64 |
+
1. **impression 过滤**：报告里 `IMPRESSION:` 段只抽取"可由 mask 验证"的句子（肿瘤存在/位置/尺寸可 cross-check）。"suggestive of"、"r/o"、"history of"、"status post" 开头的句子整段丢弃——它们常是临床史而非当前扫描可见内容。
|
| 65 |
+
2. **FOV 守卫**：PanTS 是 pancreas-centered crop (320×320×77)，报告里提及的肺结节 / 膝盖病变 / 直肠病变（出现在 ImageTr 的 CT 原体积，但 320×320×77 crop 可能剪掉）→ 用 mask 存在性过滤。若报告提到某器官但该 mask 在 crop 内为 0 voxel → 这条 finding 删掉。
|
| 66 |
+
3. **Negation 对齐**：报告"No tumor"且 lesion mask 全 0 → caption 明确 "no focal pancreatic lesion"。报告 "No tumor" 但 lesion mask 非 0 → **冲突 case 标记 + 人工抽查队列**，不自动生成 caption。
|
| 67 |
+
4. **VLM 视觉输出白名单 parser**（这是用户那句 pushback 的核心实现）：
|
| 68 |
+
|
| 69 |
+
```python
|
| 70 |
+
VISUAL_WHITE_LIST = [
|
| 71 |
+
# texture
|
| 72 |
+
"homogeneous", "heterogeneous", "mottled", "streaky",
|
| 73 |
+
# peri-organ fat
|
| 74 |
+
"abundant peri-visceral fat", "scant fat", "fatty infiltration",
|
| 75 |
+
# gas/fluid
|
| 76 |
+
"gas-distended bowel", "fluid-filled loops", "collapsed stomach",
|
| 77 |
+
# habitus
|
| 78 |
+
"slender habitus", "broad habitus", "thick subcutaneous fat",
|
| 79 |
+
# artifacts
|
| 80 |
+
"breathing artifact", "metallic streak", "beam-hardening",
|
| 81 |
+
# incidental
|
| 82 |
+
"vascular calcification", "renal cyst", "aortic calcification",
|
| 83 |
+
"surgical clip", "post-operative change",
|
| 84 |
+
]
|
| 85 |
+
# 任何不在白名单的 token → 丢弃该短语
|
| 86 |
+
```
|
| 87 |
+
|
| 88 |
+
词表最初可以小，逐步扩充。
|
| 89 |
+
|
| 90 |
+
## QC 升级（回应 GPT: regex 不够）
|
| 91 |
+
|
| 92 |
+
- **一级（确定性）**：paraphrase→JSON parse-back，field 级 exact/近似比对 → 违规重采最多 3 次。
|
| 93 |
+
- **二级（抽检）**：用 Kimi K2.5 或 MiniMax M2.5 作为第二 LLM 做 adjudication，但只抽：
|
| 94 |
+
- 一级 conflicted 的 case
|
| 95 |
+
- 每天随机 1% 的 golden check
|
| 96 |
+
- 罕见 case（多发 lesion / 血管侵犯）全覆盖
|
| 97 |
+
- **三级（人工 spot-check）**：每周 50 条抽样，画 mask overlay + caption，手工判"事实/风格/完整性"。
|
| 98 |
+
|
| 99 |
+
## Metadata 处理（接受 GPT 建议）
|
| 100 |
+
|
| 101 |
+
- **保留**：`phase`（non-contrast / arterial / venous / delayed）—— 真实影响视觉，是合法 conditioning。
|
| 102 |
+
- **删除**：`manufacturer`/`model`/`study_type`/`nationality`/`year` —— 是 nuisance，会让 diffusion 学到无关捷径（例如 "SIEMENS Sensation 16" → 特定纹理 shortcut，和解剖/病灶无关）。可单独做 metadata channel 给后续 ablation。
|
| 103 |
+
|
| 104 |
+
## 负样本处理（接受 GPT 建议）
|
| 105 |
+
|
| 106 |
+
Caption 风格不是主药，**数据采样**才是：
|
| 107 |
+
- 训练 dataloader 按 `(lesion_present, phase, lesion_subsite)` 分层过采样；
|
| 108 |
+
- lesion-positive:negative 训练分布从 1:2 调到 ≥1:1（PanTS 原分布约 1:2 是人群统计，不适合 diffusion 学少见 class）；
|
| 109 |
+
- negative case 的 V5_neg_diverse 必须引入 VLM 视觉短语（钙化 / 肾囊肿 / 伪影 / 体型），让"正常"不等于"一样的四句话"；
|
| 110 |
+
- 永远不 drop lesion-positive 的 lesion 句子（防 dropout 误伤关键信号）。
|
| 111 |
+
|
| 112 |
+
## 模型选择实验
|
| 113 |
+
|
| 114 |
+
100 case pilot（含 30 多发 lesion + 20 血管侵犯 + 10 术后 + 40 无 lesion）：
|
| 115 |
+
|
| 116 |
+
| 候选 | 角色 | 评价指标 |
|
| 117 |
+
|---|---|---|
|
| 118 |
+
| Qwen3.5-397B | 文本融合器主力 | 事实 F1 / self-BLEU / 长度分布 |
|
| 119 |
+
| Kimi K2.5 | 第二 adjudicator | agreement rate / 差异类型 |
|
| 120 |
+
| GLM5 | 备选融合器，中文报告可能更好 | 对比上者 |
|
| 121 |
+
| MiniMax M2.5 | 批量便宜 | 速度 & 成本 |
|
| 122 |
+
| Qwen2.5-VL-72B / Hulu-Med-14B | VLM 视觉描述器 | 视觉词命中白名单率 / 诊断泄漏率 |
|
| 123 |
+
|
| 124 |
+
如果 Hulu-Med 经过白名单过滤仍然能贡献 ≥50% 有效视觉词（相对 Qwen2.5-VL），保留复用节省算力；否则切 Qwen2.5-VL。
|
| 125 |
+
|
| 126 |
+
## 直接 fork RadGPT vs 自己写
|
| 127 |
+
|
| 128 |
+
**决定**：fork `${EXT_CACHE}/ct/RadGPT/generate_reports/`，做以下 PanTS 适配：
|
| 129 |
+
- 改 28 类 class map（PanTS 有 `pancreas_head`/`pancreas_body`/`pancreas_tail` 子 mask，AA 可能没这么细）
|
| 130 |
+
- 加 pancreas subsite 投票规则：lesion mask × {head/body/tail mask} 算 voxel 交集最大 → subsite
|
| 131 |
+
- 加 pancreatic duct 扩张判定（duct mask 直径 > 3mm）
|
| 132 |
+
- 加 vessel contact：lesion mask dilate 2mm ∩ {SMA, celiac, portal vein, aorta} 非空
|
| 133 |
+
- 加 FOV 守卫 + 白名单 VLM parser（新增，RadGPT 没有）
|
| 134 |
+
|
| 135 |
+
## 工程实施清单（按依赖）
|
| 136 |
+
|
| 137 |
+
1. fork RadGPT，改 class map，跑 9,901 × structured_facts.json — 2 天
|
| 138 |
+
2. 写 FOV 守卫 + impression 过滤 + conflict detector — 1 天
|
| 139 |
+
3. VLM pilot（Qwen2.5-VL vs Hulu-Med）+ 白名单 parser — 1 天
|
| 140 |
+
4. Qwen3.5-397B × 4 变体融合器（vLLM 本地 TP=8 或 openrouter）— 2 天
|
| 141 |
+
5. QC 一/二/三级 → captions_v3_1.jsonl — 1 天
|
| 142 |
+
6. dataloader 改造（变体 sampler + 分层采样 + sentence dropout）— 半天
|
| 143 |
+
|
| 144 |
+
## 开放问题（给 user 决策）
|
| 145 |
+
|
| 146 |
+
- **Q1**：VLM 是否每张 CT 都跑？成本可接受（Hulu-Med 已跑过一版，Qwen2.5-VL 本地 vLLM 可行），但对 lesion-positive case 价值更高，对全 normal 组可 oversample 来跑。建议：全量跑一次以获得统计覆盖，但 V5_neg_diverse 时才强制调用视觉短语。
|
| 147 |
+
- **Q2**：要不要保留 Hulu-Med 版 caption 作为第 5 个变体？这样可以对比是否给 diffusion 带来 hurt；或者干脆做 A/B 训练实验看 FID。
|
| 148 |
+
- **Q3**：Qwen3.5-397B 的本地部署 GPU 预算是否充足？如果不够，第一轮 openrouter 跑 5k case 看 signal，过关再全量。
|
| 149 |
+
- **Q4**：是否做 multilingual caption（中英文双语）？如果下游 diffusion 使用 T5 multilingual 或 mT5 encoder 有用；否则只英文。
|
| 150 |
+
|
| 151 |
+
## 产物
|
| 152 |
+
|
| 153 |
+
- `pants_captioning/v3_1/canonical_facts/*.json`（9,901）
|
| 154 |
+
- `pants_captioning/v3_1/visual_phrases/*.json`（VLM 产出，白名单过滤后）
|
| 155 |
+
- `pants_captioning/v3_1/captions_v3_1.jsonl`（每行 {id, V1, V2, V3, V5, provenance}）
|
| 156 |
+
- `pants_captioning/v3_1/conflicts.jsonl`（报告-mask 冲突队列，人工审）
|
| 157 |
+
- `pants_captioning/v3_1/qc_report.json`
|
| 158 |
+
- `pants_captioning/v3_1/DATA_CARD.md`
|
raw_pants_train_test/metadata/pants-captions-ldm/docs/PROPOSAL_v3_2.md
ADDED
|
@@ -0,0 +1,187 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# PanTS Caption Pipeline v3.2 — NIM Qwen3.5-397B-A17B Slice 策略版
|
| 2 |
+
|
| 3 |
+
## 1. API 验证结论
|
| 4 |
+
|
| 5 |
+
- **Endpoint**：`https://inference-api.nvidia.com/v1/chat/completions`
|
| 6 |
+
- **Model**：`nvidia/qwen/qwen3-5-397b-a17b`（注意是横线不是点，且带 nvidia/ 前缀；`qwen/...` 变体会被 401 拒）
|
| 7 |
+
- **Key**：两把都验证可用，`sk-9fLdDIqzdSm7J3_etRKLfA` 优先（另一把作 fallback）
|
| 8 |
+
- **多模态**：✅ 支持 `content` 用 OpenAI 多模态格式（text + image_url(base64)）。内部 ViT encoder。
|
| 9 |
+
- **必传参数**：`"chat_template_kwargs": {"enable_thinking": False}` — 否则 thinking 模式会把所有可见答案写到 `reasoning_content`，`content` 是 None（容易误以为失败）。
|
| 10 |
+
- **不支持**的参数：`enable_thinking` 放顶层 → 400；`extra_body` 包一层也可以。
|
| 11 |
+
- **Thinking 关闭后**：一次请求 8 图 384px 约 **1200 tokens / 2–3 秒**。
|
| 12 |
+
|
| 13 |
+
## 2. PanTS 数据 shape 实测（200 例抽样）
|
| 14 |
+
|
| 15 |
+
| 维度 | 统计 |
|
| 16 |
+
|---|---|
|
| 17 |
+
| **X (axial)** | min 50 / max 723（不统一，需 pad+resize） |
|
| 18 |
+
| **Y (axial)** | min 98 / max 493 |
|
| 19 |
+
| **Z (切片数)** | min **37** / max **557** / **median 200** / p10 91 / p90 278 |
|
| 20 |
+
| **Z spacing** | min 0.5mm / max 5.0mm / **median 1.0mm** |
|
| 21 |
+
| **Z 物理长度** | min 100mm / **median 220mm** / max 680mm（6/200 是全身） |
|
| 22 |
+
| **胰腺在 Z 方向的范围** | median **66 slice** / p10 21 / p90 116（占扫描 Z 的 35%） |
|
| 23 |
+
| **lesion Z span** | median **6 slice** / max 13 |
|
| 24 |
+
|
| 25 |
+
**已有的预处理**：`proc_v2_320x320x77/mllm_nifti/` — 每个 case 已 resize 到 **320×320×77**，spacing 0.62×0.62×0.8mm，HU clip 到 [-1000, 1000]。这是 Hulu-Med 时代用的输入格式，可直接复用，不用再跑重采样。
|
| 26 |
+
|
| 27 |
+
> 重要：PanTS 里 ~1/3 是**非胰腺 crop**（body/chest CT，可能只有几片胰腺），使得"原图 Z=200 均匀采 16 片"会浪费在非胰腺区。**必须 bbox 化。**
|
| 28 |
+
|
| 29 |
+
## 3. Slice 策略（两套输入，分工明确）
|
| 30 |
+
|
| 31 |
+
### 为什么要两个不同的输入包
|
| 32 |
+
|
| 33 |
+
- **输入包 A（胰腺 bbox overlay）**：胰腺区段 + mask overlay → 给 VLM 抽**局部外观**（胰腺纹理 / peri-pancreatic fat / lesion 周围）
|
| 34 |
+
- **输入包 B（全腹均匀 CT-only）**：整扫描 Z 均匀采样无 overlay → 给 VLM 抽**全局外观**（体型 / 伪影 / 对比度 phase / 胃肠气液 / 偶发发现）
|
| 35 |
+
|
| 36 |
+
两个包分两次独立请求，白名单 parser 合并。
|
| 37 |
+
|
| 38 |
+
### 输入包 A — 胰腺局部（8 张 overlay）
|
| 39 |
+
|
| 40 |
+
```
|
| 41 |
+
基准:
|
| 42 |
+
source: proc_v2_320x320x77/mllm_nifti/{id}/ct_xyz_320x320x77_huclip.nii.gz
|
| 43 |
+
mask source: LabelTr/{id}/combined_labels.nii.gz（28 类合并）重采样到 320×320×77
|
| 44 |
+
窗位: abdomen window W=400 L=50 (HU [-150,250] → 0–255)
|
| 45 |
+
分辨率: 384×384 JPEG Q=85
|
| 46 |
+
slice 选取: 胰腺 mask bbox z0..z1 上均匀 8 片；若 bbox z-长度 < 8，补到 8（重复端点）
|
| 47 |
+
若全无胰腺 mask（极少），退化为 Z 均匀 8 片
|
| 48 |
+
mask 呈现: 半透明颜色 overlay（alpha=0.5），同一张图既有 CT 又有 mask
|
| 49 |
+
调色板:
|
| 50 |
+
pancreas = 黄 (255,255,0)
|
| 51 |
+
pancreatic_lesion = 红 (255,0,0)
|
| 52 |
+
liver = 暗红 (200,50,50)
|
| 53 |
+
spleen = 蓝 (100,100,255)
|
| 54 |
+
kidney L/R = 绿 (50,200,50)
|
| 55 |
+
aorta = 橙 (255,150,0)
|
| 56 |
+
pancreas head/body/tail = 黄的 3 种深浅 (细区分)
|
| 57 |
+
其他 22 类 = 固定 random-colored LUT（seed=42, 保证一致）
|
| 58 |
+
token/请求: ~1200 prompt + ~200 output = 1400
|
| 59 |
+
延迟: 2-3s
|
| 60 |
+
```
|
| 61 |
+
|
| 62 |
+
### 输入包 B — 全腹上下文（6 张 CT-only）
|
| 63 |
+
|
| 64 |
+
```
|
| 65 |
+
基准:
|
| 66 |
+
source: 同 320x320x77 CT（不加 overlay）
|
| 67 |
+
slice 选取: 整个 Z 轴均匀 6 片（不按胰腺 bbox），展示胸底 / 上腹 / 中腹 / 下腹
|
| 68 |
+
mask 呈现: ❌（纯 CT）
|
| 69 |
+
token/请求: ~900 + ~150 = 1050
|
| 70 |
+
延迟: 2s
|
| 71 |
+
```
|
| 72 |
+
|
| 73 |
+
### 为什么不直接把原生 CT (median Z=200) 全送？
|
| 74 |
+
|
| 75 |
+
- **Qwen3.5 单请求 262k context**，技术上放下全部 200 slice 是可行的，但：
|
| 76 |
+
- 200 slice × 145 token ≈ **29,000 prompt tokens**，比现方案贵 **20×**；
|
| 77 |
+
- VLM 在 200 图长序列里的视觉注意力会**严重稀释**（Qwen3.5 VL benchmark 没测过 200 图输入）；
|
| 78 |
+
- 浪费：胰腺只占 35%，剩下 65% 是模型抽不出有用信号的上下腹。
|
| 79 |
+
- **bbox 8 + 全腹 6 = 14 图 ≈ 2250 tokens/case** 是最佳 cost/signal 点。
|
| 80 |
+
|
| 81 |
+
## 4. 两次请求的 Prompt
|
| 82 |
+
|
| 83 |
+
### A 请求（overlay，胰腺局部）
|
| 84 |
+
|
| 85 |
+
```
|
| 86 |
+
System: 可留空或用默认。
|
| 87 |
+
User: [8 × image] 后面跟：
|
| 88 |
+
"Eight evenly-spaced axial slices through the pancreas region. Semi-transparent color overlay shows organ masks:
|
| 89 |
+
yellow=pancreas, red=pancreatic lesion, dark-red=liver, blue=spleen, green=kidneys, orange=aorta.
|
| 90 |
+
In 4-6 sentences describe ONLY:
|
| 91 |
+
(1) pancreatic parenchyma texture around the pancreas mask (homogeneous / mottled / fibro-fatty / atrophic),
|
| 92 |
+
(2) peri-pancreatic fat (clean / stranding / fatty infiltration),
|
| 93 |
+
(3) any incidental visible findings (vascular calcification, renal cyst, surgical clip, post-op change, free fluid).
|
| 94 |
+
|
| 95 |
+
Do NOT mention: lesion diagnosis, lesion size, lesion malignancy, vessel invasion, pancreas subsite (head/body/tail), measurements, HU numbers. Stay purely descriptive."
|
| 96 |
+
```
|
| 97 |
+
|
| 98 |
+
### B 请求（CT-only，全腹上下文）
|
| 99 |
+
|
| 100 |
+
```
|
| 101 |
+
User: [6 × image] 后面跟：
|
| 102 |
+
"Six evenly-spaced axial slices through an abdominal CT (no overlay). In 3-4 sentences describe ONLY:
|
| 103 |
+
(1) overall scan quality (contrast phase feel: non-contrast / arterial / venous / delayed; artifacts: motion, beam-hardening, metal),
|
| 104 |
+
(2) patient habitus (slender / average / broad; subcutaneous fat amount),
|
| 105 |
+
(3) bowel gas/fluid distribution (gas-distended / collapsed / fluid-filled loops).
|
| 106 |
+
|
| 107 |
+
Do NOT mention: organs' anatomy or pathology, diagnosis, measurements."
|
| 108 |
+
```
|
| 109 |
+
|
| 110 |
+
**两个 prompt 显式把「禁止抽的字段」列出来**（诊断/子区/尺寸/HU/血管侵犯），这些都由确定性 extractor 产出。VLM 只填**它独有的视觉词**。
|
| 111 |
+
|
| 112 |
+
## 5. 白名单 parser
|
| 113 |
+
|
| 114 |
+
从两次请求的 response 里抽出 visual phrases，按正则匹配：
|
| 115 |
+
|
| 116 |
+
```
|
| 117 |
+
texture/parenchyma:
|
| 118 |
+
homogeneous | mottled | streaky | heterogeneous | lobulated | atrophic | fibro-fatty
|
| 119 |
+
peri-pancreatic fat:
|
| 120 |
+
clean fat plane | peri-pancreatic stranding | fatty infiltration | fluid collection
|
| 121 |
+
incidental:
|
| 122 |
+
vascular calcification | aortic calcification | renal cyst | surgical clip |
|
| 123 |
+
post-operative change | free fluid | ascites | biliary stent | nephrolithiasis
|
| 124 |
+
scan quality:
|
| 125 |
+
non-contrast phase | arterial phase | portal venous phase | delayed phase |
|
| 126 |
+
motion artifact | beam-hardening | metal artifact | streak artifact
|
| 127 |
+
habitus:
|
| 128 |
+
slender habitus | average habitus | broad habitus | thick subcutaneous fat | scant fat
|
| 129 |
+
bowel:
|
| 130 |
+
gas-distended bowel | fluid-filled loops | collapsed stomach | dilated small bowel
|
| 131 |
+
```
|
| 132 |
+
|
| 133 |
+
任何不在白名单的 phrase 丢弃。禁词黑名单（"mass"/"lesion"/"adenocarcinoma"/"cm"/"mm"/"head"/"body"/"tail"）再扫一遍，有则**整句删除**（而不是整个 response 作废，避免浪费）。
|
| 134 |
+
|
| 135 |
+
## 6. 成本核算（9,901 cases × 2 请求）
|
| 136 |
+
|
| 137 |
+
| 项 | 量 |
|
| 138 |
+
|---|---|
|
| 139 |
+
| 每 case token（A+B） | ~2250 prompt + ~350 output ≈ 2600 |
|
| 140 |
+
| 总 token（9,901 × 2 请求） | ~25.8M |
|
| 141 |
+
| 如果 NIM 免费额度够，成本 0；否则按 OpenRouter Qwen3.5 (~$0.40/M) ≈ $10 |
|
| 142 |
+
| 总时间（串行 5s/case） | 13.8h |
|
| 143 |
+
| 并发 16（应该 NIM 允许的） | **~50 分钟** |
|
| 144 |
+
|
| 145 |
+
如果改成**所有 9,901 都跑 + 文本 LLM 再跑 4 个变体**：
|
| 146 |
+
- VLM 阶段 token ≈ 25.8M
|
| 147 |
+
- 文本 LLM 阶段：每 case × 4 变体 × ~1200 token = **4 × 9.9k × 1.2k ≈ 48M tokens**
|
| 148 |
+
- 总 ≈ 74M。openrouter 或 NIM 免费都撑得住。
|
| 149 |
+
|
| 150 |
+
## 7. 并发/重试/落盘
|
| 151 |
+
|
| 152 |
+
```
|
| 153 |
+
concurrency: 16 (前期) → 32 (确认不 429 后)
|
| 154 |
+
retry: 3× exponential backoff，超时 240s
|
| 155 |
+
checkpoint: 每 100 case flush 一次 jsonl
|
| 156 |
+
断点续跑: skip 已经在 output 里的 id
|
| 157 |
+
fallback: API 失败 3 次 → 标记 vlm_fail=True，V5_neg_diverse 用 structured-only 版
|
| 158 |
+
```
|
| 159 |
+
|
| 160 |
+
## 8. 和 v3.1 的 diff
|
| 161 |
+
|
| 162 |
+
| 项 | v3.1 | v3.2 |
|
| 163 |
+
|---|---|---|
|
| 164 |
+
| VLM 选择 | 未定 | Qwen3.5-397B via NIM，已实测可用 |
|
| 165 |
+
| slice 数 | 未定 | A 包 8 张 overlay（胰腺 bbox）+ B 包 6 张 CT-only（全腹） |
|
| 166 |
+
| mask 呈现 | 未定 | 半透明 alpha=0.5 overlay；28 类用固定 LUT |
|
| 167 |
+
| 分辨率 | 未定 | 384×384 JPEG Q=85（最优 cost/quality） |
|
| 168 |
+
| thinking 模式 | 未提 | 必须关闭 |
|
| 169 |
+
| 成本预估 | 未算 | 单 VLM 阶段 ~25M tokens，~1 小时并发 |
|
| 170 |
+
|
| 171 |
+
## 9. 下一步（实操）
|
| 172 |
+
|
| 173 |
+
1. 把 `proposal_v3_1 + v3_2` 合为最终方案。
|
| 174 |
+
2. 写三个脚本：
|
| 175 |
+
- `build_input_pair.py`: 从 `proc_v2_320x320x77` 或 raw ImageTr 生成 A/B 两包 base64 图；
|
| 176 |
+
- `vlm_caption.py`: 并发调用 NIM API，checkpoint+retry，输出 `vlm_visual_phrases.jsonl`；
|
| 177 |
+
- `whitelist_parser.py`: 清洗 VLM 文本 → 短语列表；
|
| 178 |
+
- `fusion_llm.py`: 把 canonical_facts + visual_phrases 喂给**同样的 Qwen3.5-397B**（这次只用文本 modality）生成 4 变体 caption；
|
| 179 |
+
- `qc.py`: parse-back 一致性校验。
|
| 180 |
+
3. 先跑 **100 case pilot**（含 30 多发 lesion / 30 normal / 20 术后 / 20 血管侵犯），人工抽查 20 条，调白名单词表。
|
| 181 |
+
|
| 182 |
+
## 10. 未决问题
|
| 183 |
+
|
| 184 |
+
- **Q1**：VLM 同一个 Qwen3.5-397B 既做 vision 抽取 又做 text 融合 —— 这两步是一次 API 调用还是两次？**建议两次**：vision 有图很贵，text 融合无图便宜；分离能各自 parallelize + 换 seed 做多样化。
|
| 185 |
+
- **Q2**：B 包（全腹）有没有必要？如果 A 包胰腺 overlay 已经能稳定抽到 scan quality/habitus，B 包可省掉，token 成本减半。我倾向保留 B，因为 A 包是胰腺 tight crop，看不到体型。
|
| 186 |
+
- **Q3**：combined_labels.nii.gz 在原生 ImageTr 空间，need resample 到 320×320×77 才能和 proc_v2 的 CT 对齐。mllm_nifti/ 里只有 CT 没 mask。**需要加一步 mask 重采样**。
|
| 187 |
+
- **Q4**：hulumed 遗留的 `proc_v2_320x320x77/mllm_nifti/{id}/lesion_xyz_320x320x77.nii.gz` 已经是 lesion-only mask，可直接用；但其他 27 类要现生成。
|
raw_pants_train_test/metadata/pants-captions-ldm/docs/fusion_variants_design.md
ADDED
|
@@ -0,0 +1,99 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# F5 Fusion: Caption Variants for Diffusion Training
|
| 2 |
+
|
| 3 |
+
Goal: maximize text diversity so the T2I diffusion learns a rich prompt-conditioning space, not a single template style.
|
| 4 |
+
|
| 5 |
+
Input per case (from F3 + F4):
|
| 6 |
+
- canonical_facts (deterministic ground truth from metadata+masks)
|
| 7 |
+
- vlm_A_json (closed-enum: magenta_texture, fat_around_magenta, cyan_focus_seen/density, incidental_findings)
|
| 8 |
+
- vlm_B_prose (habitus/bowel/artifacts free text)
|
| 9 |
+
|
| 10 |
+
LLM: Qwen3.5-397B text-only (no images → cheap, ~1200 token prompt / ~200 token output).
|
| 11 |
+
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
## 7 variants per case (1 case → 7 captions)
|
| 15 |
+
|
| 16 |
+
| # | Name | Length | Style | What it includes |
|
| 17 |
+
|---|------|--------|-------|------------------|
|
| 18 |
+
| V1 | **long_narrative** | 150-280 words | full radiology report prose | phase, habitus, pancreas finding w/ cm+subregion, enhancement, vessel contact, incidental findings, artifacts |
|
| 19 |
+
| V2 | **terse_impression** | 15-40 words | one-line impression | only what's clinically salient; mirrors clinical "IMPRESSION:" style |
|
| 20 |
+
| V3 | **organ_bullet** | 80-150 words | markdown-ish bullets | `- Pancreas: ...` `- Liver: ...` structured per-organ |
|
| 21 |
+
| V4 | **tag_string** | 8-20 words | comma-separated keyword tags | SDXL-style prompt: "abdominal CT, venous phase, pancreas head mass 2.4cm, isoattenuating, mild aortic calcification" |
|
| 22 |
+
| V5 | **layered_findings** | 100-180 words | radiologist "Technique / Findings / Impression" | 3 sections: scan params, findings by system, brief impression |
|
| 23 |
+
| V6 | **negative_descriptive** | 60-120 words | for negative cases, heavy use of VLM visual phrases | habitus+texture+fat+artifacts; no lesion mention if canonical says none |
|
| 24 |
+
| V7 | **qa_pair** | 30-80 words | question+answer style | "What is the key finding? A pancreatic tail mass measuring 1.6×1.1cm..." |
|
| 25 |
+
|
| 26 |
+
---
|
| 27 |
+
|
| 28 |
+
## Variant selection per case
|
| 29 |
+
|
| 30 |
+
**If canonical.lesion_present = TRUE** → generate V1, V2, V3, V4, V5, V7 (6 variants, lesion-rich)
|
| 31 |
+
**If canonical.lesion_present = FALSE** → generate V1, V2, V3, V4, V5, V6 (6 variants, VLM-driven negatives)
|
| 32 |
+
|
| 33 |
+
This gives 6 captions per case × 9,901 = **~59K training captions**.
|
| 34 |
+
|
| 35 |
+
---
|
| 36 |
+
|
| 37 |
+
## Critical fusion rules (from Codex 3rd review)
|
| 38 |
+
|
| 39 |
+
1. **Lesion subtree override**: if canonical.lesion_present==False, **drop** `cyan_focus_density` entirely (not just set to `none`) from VLM A. The fusion LLM must not see hallucinated "hypodense lesion" facts.
|
| 40 |
+
2. **B prose never references lesion**: fusion LLM is instructed to only use B for habitus/bowel/artifacts — `lesion/mass/tumor` words from B are filtered before fusion.
|
| 41 |
+
3. **Banned words**: `adenocarcinoma, malignant, metastasis, T1/T2/T3/T4, unresectable` never appear in any variant (lesion staging is not in canonical facts).
|
| 42 |
+
4. **Sizes come ONLY from canonical.lesion_list.size_cm**; VLM-reported sizes are ignored (hallucination-prone).
|
| 43 |
+
5. **Every variant must be parse-back-consistent** with canonical on these keys:
|
| 44 |
+
- lesion_present (bool)
|
| 45 |
+
- lesion_count (int)
|
| 46 |
+
- lesion_subregion if present (head/body/tail)
|
| 47 |
+
- CT phase mention matches canonical.phase
|
| 48 |
+
- FOV mention matches canonical.fov
|
| 49 |
+
|
| 50 |
+
---
|
| 51 |
+
|
| 52 |
+
## Prompts (for Qwen3.5-397B text-only)
|
| 53 |
+
|
| 54 |
+
Each variant gets its own prompt. Example V1:
|
| 55 |
+
|
| 56 |
+
```
|
| 57 |
+
SYSTEM: You produce a single caption describing a pancreatic CT scan for training a
|
| 58 |
+
text-to-image medical diffusion model. Write ONE caption in the target style below.
|
| 59 |
+
Never invent facts not listed in FACTS. Never use the words "adenocarcinoma",
|
| 60 |
+
"malignant", "metastasis", "T1/T2/T3/T4", "unresectable". Never give a diagnosis.
|
| 61 |
+
|
| 62 |
+
FACTS (deterministic, from mask+report; treat as ground truth):
|
| 63 |
+
{canonical_json}
|
| 64 |
+
|
| 65 |
+
VLM_TEXTURE_CUE (heuristic, already filtered): {vlm_A_filtered}
|
| 66 |
+
VLM_GLOBAL_CUE (patient/artifact): {vlm_B_filtered}
|
| 67 |
+
|
| 68 |
+
STYLE: long_narrative — 150-280 words, flowing prose, full radiology-report voice.
|
| 69 |
+
Open with one sentence on scan phase+FOV. Then habitus. Then pancreas findings
|
| 70 |
+
(size cm from FACTS only). Then any vessel contact from FACTS.contact. Then
|
| 71 |
+
incidental findings from both FACTS and VLM. Close with a short impression sentence.
|
| 72 |
+
|
| 73 |
+
OUTPUT: the caption only, no prefix, no markdown fence.
|
| 74 |
+
```
|
| 75 |
+
|
| 76 |
+
Other variants are shorter prompts with enforced word-count + format.
|
| 77 |
+
|
| 78 |
+
---
|
| 79 |
+
|
| 80 |
+
## Cost
|
| 81 |
+
|
| 82 |
+
- ~1200 prompt + ~250 output per variant × 6 variants × 9,901 cases
|
| 83 |
+
= ~87M prompt + 15M output tokens ≈ **102M total**
|
| 84 |
+
- Text-only Qwen3.5 → ~1.5s per request (~half of VLM)
|
| 85 |
+
- 24 workers × 6 variants × 9,901 cases = 237,624 requests
|
| 86 |
+
- At 1.5s / request / 24 workers → ~4.1 hours wall time
|
| 87 |
+
|
| 88 |
+
---
|
| 89 |
+
|
| 90 |
+
## QC (F6)
|
| 91 |
+
|
| 92 |
+
Parse-back extractor on each generated caption checks:
|
| 93 |
+
- lesion_present matches canonical (regex "mass|lesion|tumor" on V1-V5; V6 must have no such words)
|
| 94 |
+
- lesion count matches canonical if present (regex for "two masses", "multiple", numeric)
|
| 95 |
+
- subregion matches if lesion present (head/body/tail)
|
| 96 |
+
- phase word matches canonical.phase (Non-contrast/Arterial/Venous/Delay)
|
| 97 |
+
- ban-word sweep: reject the variant if any banned word leaks
|
| 98 |
+
|
| 99 |
+
Failure rate target: <2% per variant. Failed variants get **1 retry** with the violation pointed out; on 2nd failure we **drop** the variant for that case (other variants still kept).
|
raw_pants_train_test/metadata/pants-captions-ldm/samples/canary/annotations.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
raw_pants_train_test/metadata/pants-captions-ldm/samples/canary/ids.json
ADDED
|
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
"PanTS_00003087",
|
| 3 |
+
"PanTS_00007160",
|
| 4 |
+
"PanTS_00001364",
|
| 5 |
+
"PanTS_00000645",
|
| 6 |
+
"PanTS_00007212",
|
| 7 |
+
"PanTS_00008936",
|
| 8 |
+
"PanTS_00001260",
|
| 9 |
+
"PanTS_00001254",
|
| 10 |
+
"PanTS_00003094",
|
| 11 |
+
"PanTS_00001543",
|
| 12 |
+
"PanTS_00003131",
|
| 13 |
+
"PanTS_00003803",
|
| 14 |
+
"PanTS_00008907",
|
| 15 |
+
"PanTS_00002067",
|
| 16 |
+
"PanTS_00006621",
|
| 17 |
+
"PanTS_00006323",
|
| 18 |
+
"PanTS_00004942",
|
| 19 |
+
"PanTS_00000512",
|
| 20 |
+
"PanTS_00001745",
|
| 21 |
+
"PanTS_00000346",
|
| 22 |
+
"PanTS_00000493",
|
| 23 |
+
"PanTS_00006067",
|
| 24 |
+
"PanTS_00004024",
|
| 25 |
+
"PanTS_00002591",
|
| 26 |
+
"PanTS_00008133",
|
| 27 |
+
"PanTS_00000644",
|
| 28 |
+
"PanTS_00002463",
|
| 29 |
+
"PanTS_00000228",
|
| 30 |
+
"PanTS_00001622",
|
| 31 |
+
"PanTS_00001452",
|
| 32 |
+
"PanTS_00002521",
|
| 33 |
+
"PanTS_00002047",
|
| 34 |
+
"PanTS_00001875",
|
| 35 |
+
"PanTS_00005175",
|
| 36 |
+
"PanTS_00001751",
|
| 37 |
+
"PanTS_00003868",
|
| 38 |
+
"PanTS_00004689",
|
| 39 |
+
"PanTS_00002213",
|
| 40 |
+
"PanTS_00002920",
|
| 41 |
+
"PanTS_00007829",
|
| 42 |
+
"PanTS_00001640",
|
| 43 |
+
"PanTS_00002119",
|
| 44 |
+
"PanTS_00003953",
|
| 45 |
+
"PanTS_00006293",
|
| 46 |
+
"PanTS_00000457",
|
| 47 |
+
"PanTS_00000233",
|
| 48 |
+
"PanTS_00003306",
|
| 49 |
+
"PanTS_00004812",
|
| 50 |
+
"PanTS_00007888",
|
| 51 |
+
"PanTS_00002101",
|
| 52 |
+
"PanTS_00000980",
|
| 53 |
+
"PanTS_00008526",
|
| 54 |
+
"PanTS_00002164",
|
| 55 |
+
"PanTS_00004264",
|
| 56 |
+
"PanTS_00000945",
|
| 57 |
+
"PanTS_00000260",
|
| 58 |
+
"PanTS_00001970",
|
| 59 |
+
"PanTS_00001654",
|
| 60 |
+
"PanTS_00003808",
|
| 61 |
+
"PanTS_00002301",
|
| 62 |
+
"PanTS_00003445",
|
| 63 |
+
"PanTS_00007344",
|
| 64 |
+
"PanTS_00002125",
|
| 65 |
+
"PanTS_00000476",
|
| 66 |
+
"PanTS_00008969",
|
| 67 |
+
"PanTS_00000234",
|
| 68 |
+
"PanTS_00004050",
|
| 69 |
+
"PanTS_00003190",
|
| 70 |
+
"PanTS_00006725",
|
| 71 |
+
"PanTS_00005048",
|
| 72 |
+
"PanTS_00000835",
|
| 73 |
+
"PanTS_00000514",
|
| 74 |
+
"PanTS_00007561",
|
| 75 |
+
"PanTS_00000800",
|
| 76 |
+
"PanTS_00002484",
|
| 77 |
+
"PanTS_00002327",
|
| 78 |
+
"PanTS_00004867",
|
| 79 |
+
"PanTS_00004545",
|
| 80 |
+
"PanTS_00005613",
|
| 81 |
+
"PanTS_00000951",
|
| 82 |
+
"PanTS_00004834",
|
| 83 |
+
"PanTS_00007103",
|
| 84 |
+
"PanTS_00002623",
|
| 85 |
+
"PanTS_00008452",
|
| 86 |
+
"PanTS_00005702",
|
| 87 |
+
"PanTS_00000485",
|
| 88 |
+
"PanTS_00000656",
|
| 89 |
+
"PanTS_00008816",
|
| 90 |
+
"PanTS_00001723",
|
| 91 |
+
"PanTS_00001793",
|
| 92 |
+
"PanTS_00001653",
|
| 93 |
+
"PanTS_00000360",
|
| 94 |
+
"PanTS_00002216",
|
| 95 |
+
"PanTS_00005084",
|
| 96 |
+
"PanTS_00001188",
|
| 97 |
+
"PanTS_00005460",
|
| 98 |
+
"PanTS_00002583",
|
| 99 |
+
"PanTS_00000620",
|
| 100 |
+
"PanTS_00000182",
|
| 101 |
+
"PanTS_00007709",
|
| 102 |
+
"PanTS_00001369",
|
| 103 |
+
"PanTS_00004256",
|
| 104 |
+
"PanTS_00003771",
|
| 105 |
+
"PanTS_00007346",
|
| 106 |
+
"PanTS_00007988",
|
| 107 |
+
"PanTS_00005093",
|
| 108 |
+
"PanTS_00004637",
|
| 109 |
+
"PanTS_00001344",
|
| 110 |
+
"PanTS_00002466",
|
| 111 |
+
"PanTS_00008456",
|
| 112 |
+
"PanTS_00000676",
|
| 113 |
+
"PanTS_00003841",
|
| 114 |
+
"PanTS_00001049",
|
| 115 |
+
"PanTS_00000692",
|
| 116 |
+
"PanTS_00000100",
|
| 117 |
+
"PanTS_00001552",
|
| 118 |
+
"PanTS_00001101",
|
| 119 |
+
"PanTS_00003918",
|
| 120 |
+
"PanTS_00001297",
|
| 121 |
+
"PanTS_00008147",
|
| 122 |
+
"PanTS_00004839",
|
| 123 |
+
"PanTS_00003764",
|
| 124 |
+
"PanTS_00004004",
|
| 125 |
+
"PanTS_00006317",
|
| 126 |
+
"PanTS_00003856",
|
| 127 |
+
"PanTS_00001019",
|
| 128 |
+
"PanTS_00001234",
|
| 129 |
+
"PanTS_00007466",
|
| 130 |
+
"PanTS_00005483",
|
| 131 |
+
"PanTS_00000875",
|
| 132 |
+
"PanTS_00006797",
|
| 133 |
+
"PanTS_00000443",
|
| 134 |
+
"PanTS_00003642",
|
| 135 |
+
"PanTS_00008075",
|
| 136 |
+
"PanTS_00001356",
|
| 137 |
+
"PanTS_00003701",
|
| 138 |
+
"PanTS_00001618",
|
| 139 |
+
"PanTS_00002915",
|
| 140 |
+
"PanTS_00002274",
|
| 141 |
+
"PanTS_00000726",
|
| 142 |
+
"PanTS_00002766",
|
| 143 |
+
"PanTS_00001072",
|
| 144 |
+
"PanTS_00000375",
|
| 145 |
+
"PanTS_00000860",
|
| 146 |
+
"PanTS_00003805",
|
| 147 |
+
"PanTS_00004627",
|
| 148 |
+
"PanTS_00005059",
|
| 149 |
+
"PanTS_00004111",
|
| 150 |
+
"PanTS_00003690",
|
| 151 |
+
"PanTS_00003203",
|
| 152 |
+
"PanTS_00004621",
|
| 153 |
+
"PanTS_00003176",
|
| 154 |
+
"PanTS_00003270",
|
| 155 |
+
"PanTS_00007149",
|
| 156 |
+
"PanTS_00000068",
|
| 157 |
+
"PanTS_00002563",
|
| 158 |
+
"PanTS_00007037",
|
| 159 |
+
"PanTS_00001567",
|
| 160 |
+
"PanTS_00000781",
|
| 161 |
+
"PanTS_00004790",
|
| 162 |
+
"PanTS_00008269",
|
| 163 |
+
"PanTS_00007223",
|
| 164 |
+
"PanTS_00000248",
|
| 165 |
+
"PanTS_00002066",
|
| 166 |
+
"PanTS_00002276",
|
| 167 |
+
"PanTS_00003316",
|
| 168 |
+
"PanTS_00000041",
|
| 169 |
+
"PanTS_00007415",
|
| 170 |
+
"PanTS_00000156",
|
| 171 |
+
"PanTS_00000968",
|
| 172 |
+
"PanTS_00001156",
|
| 173 |
+
"PanTS_00001680",
|
| 174 |
+
"PanTS_00002127",
|
| 175 |
+
"PanTS_00002467",
|
| 176 |
+
"PanTS_00005558",
|
| 177 |
+
"PanTS_00000350",
|
| 178 |
+
"PanTS_00001775",
|
| 179 |
+
"PanTS_00002242",
|
| 180 |
+
"PanTS_00002382",
|
| 181 |
+
"PanTS_00003505",
|
| 182 |
+
"PanTS_00001532",
|
| 183 |
+
"PanTS_00001805",
|
| 184 |
+
"PanTS_00002647",
|
| 185 |
+
"PanTS_00008978",
|
| 186 |
+
"PanTS_00003035",
|
| 187 |
+
"PanTS_00003161",
|
| 188 |
+
"PanTS_00001677",
|
| 189 |
+
"PanTS_00001863",
|
| 190 |
+
"PanTS_00001023",
|
| 191 |
+
"PanTS_00004638",
|
| 192 |
+
"PanTS_00001520",
|
| 193 |
+
"PanTS_00003583",
|
| 194 |
+
"PanTS_00000115",
|
| 195 |
+
"PanTS_00003859",
|
| 196 |
+
"PanTS_00003599",
|
| 197 |
+
"PanTS_00001663",
|
| 198 |
+
"PanTS_00004722",
|
| 199 |
+
"PanTS_00003651",
|
| 200 |
+
"PanTS_00005785",
|
| 201 |
+
"PanTS_00003075"
|
| 202 |
+
]
|
raw_pants_train_test/metadata/pants-captions-ldm/samples/pilot/annotations.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
raw_pants_train_test/metadata/pants-captions-ldm/samples/pilot_v2/annotations.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
raw_pants_train_test/metadata/pants-captions-ldm/splits/bucket_spec.json
ADDED
|
@@ -0,0 +1,175 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"buckets": [
|
| 3 |
+
{
|
| 4 |
+
"name": "B-whole",
|
| 5 |
+
"fov": "whole_body",
|
| 6 |
+
"shape_dhw": [
|
| 7 |
+
128,
|
| 8 |
+
256,
|
| 9 |
+
256
|
| 10 |
+
],
|
| 11 |
+
"voxel_spacing_mm": [
|
| 12 |
+
5.0,
|
| 13 |
+
2.0,
|
| 14 |
+
2.0
|
| 15 |
+
],
|
| 16 |
+
"coverage_mm": [
|
| 17 |
+
640.0,
|
| 18 |
+
512.0,
|
| 19 |
+
512.0
|
| 20 |
+
],
|
| 21 |
+
"latent_shape_dhw": [
|
| 22 |
+
33,
|
| 23 |
+
16,
|
| 24 |
+
16
|
| 25 |
+
],
|
| 26 |
+
"tokens_after_patch": 2112,
|
| 27 |
+
"stage": "A"
|
| 28 |
+
},
|
| 29 |
+
{
|
| 30 |
+
"name": "B-CAP",
|
| 31 |
+
"fov": "chest_abdomen_pelvis",
|
| 32 |
+
"shape_dhw": [
|
| 33 |
+
96,
|
| 34 |
+
320,
|
| 35 |
+
256
|
| 36 |
+
],
|
| 37 |
+
"voxel_spacing_mm": [
|
| 38 |
+
4.0,
|
| 39 |
+
2.0,
|
| 40 |
+
2.0
|
| 41 |
+
],
|
| 42 |
+
"coverage_mm": [
|
| 43 |
+
384.0,
|
| 44 |
+
640.0,
|
| 45 |
+
512.0
|
| 46 |
+
],
|
| 47 |
+
"latent_shape_dhw": [
|
| 48 |
+
25,
|
| 49 |
+
20,
|
| 50 |
+
16
|
| 51 |
+
],
|
| 52 |
+
"tokens_after_patch": 2000,
|
| 53 |
+
"stage": "A_B"
|
| 54 |
+
},
|
| 55 |
+
{
|
| 56 |
+
"name": "B-CA",
|
| 57 |
+
"fov": "chest_abdomen",
|
| 58 |
+
"shape_dhw": [
|
| 59 |
+
80,
|
| 60 |
+
320,
|
| 61 |
+
256
|
| 62 |
+
],
|
| 63 |
+
"voxel_spacing_mm": [
|
| 64 |
+
3.5,
|
| 65 |
+
2.0,
|
| 66 |
+
2.0
|
| 67 |
+
],
|
| 68 |
+
"coverage_mm": [
|
| 69 |
+
280.0,
|
| 70 |
+
640.0,
|
| 71 |
+
512.0
|
| 72 |
+
],
|
| 73 |
+
"latent_shape_dhw": [
|
| 74 |
+
21,
|
| 75 |
+
20,
|
| 76 |
+
16
|
| 77 |
+
],
|
| 78 |
+
"tokens_after_patch": 1680,
|
| 79 |
+
"stage": "B"
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"name": "B-abd",
|
| 83 |
+
"fov": "abdomen_only",
|
| 84 |
+
"shape_dhw": [
|
| 85 |
+
64,
|
| 86 |
+
256,
|
| 87 |
+
256
|
| 88 |
+
],
|
| 89 |
+
"voxel_spacing_mm": [
|
| 90 |
+
3.0,
|
| 91 |
+
2.0,
|
| 92 |
+
2.0
|
| 93 |
+
],
|
| 94 |
+
"coverage_mm": [
|
| 95 |
+
192.0,
|
| 96 |
+
512.0,
|
| 97 |
+
512.0
|
| 98 |
+
],
|
| 99 |
+
"latent_shape_dhw": [
|
| 100 |
+
17,
|
| 101 |
+
16,
|
| 102 |
+
16
|
| 103 |
+
],
|
| 104 |
+
"tokens_after_patch": 1088,
|
| 105 |
+
"stage": "B"
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"name": "B-abd-pelvis",
|
| 109 |
+
"fov": "abdomen_pelvis",
|
| 110 |
+
"shape_dhw": [
|
| 111 |
+
64,
|
| 112 |
+
256,
|
| 113 |
+
256
|
| 114 |
+
],
|
| 115 |
+
"voxel_spacing_mm": [
|
| 116 |
+
3.0,
|
| 117 |
+
2.0,
|
| 118 |
+
2.0
|
| 119 |
+
],
|
| 120 |
+
"coverage_mm": [
|
| 121 |
+
192.0,
|
| 122 |
+
512.0,
|
| 123 |
+
512.0
|
| 124 |
+
],
|
| 125 |
+
"latent_shape_dhw": [
|
| 126 |
+
17,
|
| 127 |
+
16,
|
| 128 |
+
16
|
| 129 |
+
],
|
| 130 |
+
"tokens_after_patch": 1088,
|
| 131 |
+
"stage": "B"
|
| 132 |
+
},
|
| 133 |
+
{
|
| 134 |
+
"name": "B-pan",
|
| 135 |
+
"fov": "pan_crop",
|
| 136 |
+
"shape_dhw": [
|
| 137 |
+
64,
|
| 138 |
+
192,
|
| 139 |
+
256
|
| 140 |
+
],
|
| 141 |
+
"voxel_spacing_mm": [
|
| 142 |
+
2.0,
|
| 143 |
+
1.0,
|
| 144 |
+
1.0
|
| 145 |
+
],
|
| 146 |
+
"coverage_mm": [
|
| 147 |
+
128.0,
|
| 148 |
+
192.0,
|
| 149 |
+
256.0
|
| 150 |
+
],
|
| 151 |
+
"latent_shape_dhw": [
|
| 152 |
+
17,
|
| 153 |
+
12,
|
| 154 |
+
16
|
| 155 |
+
],
|
| 156 |
+
"tokens_after_patch": 816,
|
| 157 |
+
"stage": "C"
|
| 158 |
+
}
|
| 159 |
+
],
|
| 160 |
+
"token_cap": 8192,
|
| 161 |
+
"vae": {
|
| 162 |
+
"compression": [
|
| 163 |
+
4,
|
| 164 |
+
16,
|
| 165 |
+
16
|
| 166 |
+
],
|
| 167 |
+
"channels": 48,
|
| 168 |
+
"family": "wan22"
|
| 169 |
+
},
|
| 170 |
+
"dit_patch": [
|
| 171 |
+
1,
|
| 172 |
+
2,
|
| 173 |
+
2
|
| 174 |
+
]
|
| 175 |
+
}
|
raw_pants_train_test/metadata/pants-captions-ldm/splits/splits.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
raw_pants_train_test/metadata/raw_summary.txt
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
PanTS raw root: /scratch/user/yuhwang/dataset/PanTS
|
| 2 |
+
Created: Sat May 23 10:24:22 PDT 2026
|
| 3 |
+
|
| 4 |
+
Top-level sizes:
|
| 5 |
+
371G /scratch/user/yuhwang/dataset/PanTS
|
| 6 |
+
|
| 7 |
+
Train/test/label directory counts:
|
| 8 |
+
ImageTr dirs=9000
|
| 9 |
+
ImageTr files=9000
|
| 10 |
+
296G /scratch/user/yuhwang/dataset/PanTS/data/ImageTr
|
| 11 |
+
ImageTe dirs=901
|
| 12 |
+
ImageTe files=901
|
| 13 |
+
27G /scratch/user/yuhwang/dataset/PanTS/data/ImageTe
|
| 14 |
+
LabelTr dirs=9000
|
| 15 |
+
LabelTr files=260999
|
| 16 |
+
45G /scratch/user/yuhwang/dataset/PanTS/data/LabelTr
|
| 17 |
+
LabelTe dirs=901
|
| 18 |
+
LabelTe files=26129
|
| 19 |
+
4.0G /scratch/user/yuhwang/dataset/PanTS/data/LabelTe
|
| 20 |
+
|
| 21 |
+
Metadata files:
|
| 22 |
+
1334813 /scratch/user/yuhwang/dataset/PanTS/data/metadata.xlsx
|
| 23 |
+
15379 /scratch/user/yuhwang/dataset/PanTS/repo/LICENSE
|
| 24 |
+
8689 /scratch/user/yuhwang/dataset/PanTS/repo/README.md
|
| 25 |
+
2115 /scratch/user/yuhwang/dataset/PanTS/repo/data/README.md
|
| 26 |
+
1339 /scratch/user/yuhwang/dataset/PanTS/repo/download_PanTS_label.sh
|
| 27 |
+
1131 /scratch/user/yuhwang/dataset/PanTS/repo/download_PanTS_data.sh
|
raw_pants_train_test/raw/RESTORE.txt
ADDED
|
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
|
|
|
| 1 |
+
cat PanTS.tar.zst.part-* > PanTS.tar.zst
|
| 2 |
+
tar --use-compress-program zstd -xf PanTS.tar.zst
|