crosstyan/OpenGait
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: add drf author checkpoint compatibility bundle

Browse Source
This commit is contained in:
crosstyan
2026-03-14 17:12:27 +08:00
parent d4e2a59ad2
commit 5f98844aff
18 changed files with 1144 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/drf_author_checkpoint_compat.md
+192
View File
@@ -0,0 +1,192 @@
# DRF Author Checkpoint Compatibility Note
This note records what happened when evaluating the author-provided DRF bundle in this repo:
- checkpoint: `artifact/scoliosis_drf_author_118_compat/DRF_118_unordered_iter2w_lr0.001_8830-08000.pt`
- config: `ckpt/drf_author/drf_scoliosis1k_20000.yaml`
The short version:
- the weight file is real and structurally usable
- the provided YAML is not a reliable source of truth
- the main problem was integration-contract mismatch, not a broken checkpoint
## What Was Wrong
The author bundle was internally inconsistent in several ways.
### 1. Split mismatch
The DRF paper says the main experiment uses `1:1:8`, i.e. the `118` split.
But the provided YAML pointed to:
- `./datasets/Scoliosis1K/Scoliosis1K_112.json`
while the checkpoint filename itself says:
- `DRF_118_...`
So the bundle already disagreed with itself.
### 2. Class-order mismatch
The biggest hidden bug was class ordering.
The current repo evaluator assumes:
- `negative = 0`
- `neutral = 1`
- `positive = 2`
But the author stub in `research/drf.py` uses:
- `negative = 0`
- `positive = 1`
- `neutral = 2`
That means an otherwise good checkpoint can look very bad if logits are interpreted in the wrong class order.
### 3. Legacy module-name mismatch
The author checkpoint stores PGA weights under:
- `attention_layer.*`
The current repo uses:
- `PGA.*`
This is a small compatibility issue, but it must be remapped before loading.
### 4. Preprocessing/runtime-contract mismatch
The author checkpoint does not line up with the stale YAML’s full runtime contract.
Most importantly, it did **not** work well with the more paper-literal local export:
- `Scoliosis1K-drf-pkl-118-paper`
It worked much better with the more OpenGait-like aligned export:
- `Scoliosis1K-drf-pkl-118-aligned`
That strongly suggests the checkpoint was trained against a preprocessing/runtime path closer to the aligned OpenGait integration than to the later local “paper-literal” summed-heatmap ablation.
## What Was Added In-Tree
The current repo now has a small compatibility layer in:
- `opengait/modeling/models/drf.py`
It does two things:
- remaps legacy keys `attention_layer.* -> PGA.*`
- supports configurable `model_cfg.label_order`
The model also canonicalizes inference logits back into the repo’s evaluator order, so author checkpoints can be evaluated without modifying the evaluator itself.
## Tested Compatibility Results
### Best usable author-checkpoint path
Config:
- `configs/drf/drf_author_eval_118_aligned_1gpu.yaml`
Dataset/runtime:
- dataset root: `Scoliosis1K-drf-pkl-118-aligned`
- partition: `Scoliosis1K_118.json`
- transform: `BaseSilCuttingTransform`
- label order:
- `negative`
- `positive`
- `neutral`
Result:
- `80.24 Acc / 76.73 Prec / 76.40 Rec / 76.56 F1`
This is the strongest recovered path so far.
### Other tested paths
`configs/drf/drf_author_eval_118_splitroot_1gpu.yaml`
- dataset root: `Scoliosis1K-drf-pkl-118`
- result:
- `77.17 Acc / 73.61 Prec / 72.59 Rec / 72.98 F1`
`configs/drf/drf_author_eval_112_1gpu.yaml`
- dataset root: `Scoliosis1K-drf-pkl`
- partition: `Scoliosis1K_112.json`
- result:
- `85.19 Acc / 57.98 Prec / 56.65 Rec / 57.30 F1`
`configs/drf/drf_author_eval_118_paper_1gpu.yaml`
- dataset root: `Scoliosis1K-drf-pkl-118-paper`
- transform: `BaseSilTransform`
- result:
- `27.24 Acc / 9.08 Prec / 33.33 Rec / 14.27 F1`
## Interpretation
What these results mean:
- the checkpoint is not garbage
- the original “very bad” local eval was mostly a compatibility failure
- the largest single hidden bug was the class-order mismatch
- the author checkpoint is also sensitive to which local DRF dataset root is used
What they do **not** mean:
- we have perfectly reconstructed the author’s original training path
- the provided YAML is trustworthy as-is
- the paper’s full DRF claim is fully reproduced here
The strongest recovered result:
- `80.24 / 76.73 / 76.40 / 76.56`
This is close to the paper’s reported `ScoNet-MT^ske` F1 and much better than our earlier broken compat evals, but it is still below the paper’s DRF headline result:
- paper DRF: `86.0 Acc / 84.1 Prec / 79.2 Rec / 80.8 F1`
## Practical Recommendation
If someone wants to use the author checkpoint in this repo today, the recommended path is:
1. use `configs/drf/drf_author_eval_118_aligned_1gpu.yaml`
2. keep the author label order:
- `negative, positive, neutral`
3. keep the legacy `attention_layer -> PGA` remap in the model
4. do **not** assume the stale `112` YAML is the correct training/eval contract
If someone wants to push this further, the highest-value next step is:
- finetune from the author checkpoint on the aligned `118` path instead of starting DRF from scratch
## How To Run
Recommended eval:
```bash
CUDA_VISIBLE_DEVICES=GPU-9cc7b26e-90d4-0c49-4d4c-060e528ffba6 \
uv run torchrun --nproc_per_node=1 --master_port=29693 \
opengait/main.py \
--cfgs ./configs/drf/drf_author_eval_118_aligned_1gpu.yaml \
--phase test
```
Other compatibility checks:
```bash
CUDA_VISIBLE_DEVICES=GPU-9cc7b26e-90d4-0c49-4d4c-060e528ffba6 \
uv run torchrun --nproc_per_node=1 --master_port=29695 \
opengait/main.py \
--cfgs ./configs/drf/drf_author_eval_112_1gpu.yaml \
--phase test
CUDA_VISIBLE_DEVICES=GPU-9cc7b26e-90d4-0c49-4d4c-060e528ffba6 \
uv run torchrun --nproc_per_node=1 --master_port=29696 \
opengait/main.py \
--cfgs ./configs/drf/drf_author_eval_118_splitroot_1gpu.yaml \
--phase test
CUDA_VISIBLE_DEVICES=GPU-9cc7b26e-90d4-0c49-4d4c-060e528ffba6 \
uv run torchrun --nproc_per_node=1 --master_port=29697 \
opengait/main.py \
--cfgs ./configs/drf/drf_author_eval_118_paper_1gpu.yaml \
--phase test
```
If someone wants to reproduce this on another machine, the usual paths to change are:
- `data_cfg.dataset_root`
- `data_cfg.dataset_partition`
- `evaluator_cfg.restore_hint`
The archived artifact bundle is:
- `artifact/scoliosis_drf_author_118_compat`
docs/scoliosis_next_experiments.md
+11 -1
View File
@@ -7,6 +7,7 @@ Related notes:
- [ScoNet and DRF: Status, Architecture, and Reproduction Notes](sconet-drf-status-and-training.md)
- [Scoliosis Training Change Log](scoliosis_training_change_log.md)
- [Scoliosis Reproducibility Audit](scoliosis_reproducibility_audit.md)
- [DRF Author Checkpoint Compatibility Note](drf_author_checkpoint_compat.md)
- [systemd-run Training](systemd-run-training.md)
## Current Best Known Result
@@ -72,6 +73,7 @@ These are the main branches already explored, so the next person does not rerun
- `DRF`
- implemented and tested
- underperformed the plain skeleton baseline in the first serious practical run
- the author-provided checkpoint is now partially recoverable on the aligned `118` path
- still worth one final warm-start retry, but not the default winner
## What Not To Do First
@@ -120,7 +122,9 @@ Goal:
Recommended setup:
- split: `1:1:2`
- representation: start from the same `body-only` skeleton maps
- initialization: warm-start from the strong skeleton model, not from random init
- initialization: warm-start from either:
- the strong practical skeleton model, or
- the recovered author DRF checkpoint if the goal is paper-side DRF recovery
- optimizer: small-LR `AdamW`
- scheduler: cosine
@@ -130,9 +134,15 @@ Recommended strategy:
3. use a smaller LR than the plain baseline finetune
4. if needed, freeze or partially freeze the backbone for a short warmup so the PAV/PGA branch learns without immediately disturbing the already-good visual branch
If the goal is specifically DRF recovery rather than practical best-model development, use this as the author-checkpoint starting point:
- config: `configs/drf/drf_author_eval_118_aligned_1gpu.yaml`
- recovered author checkpoint result:
- `80.24 Acc / 76.73 Prec / 76.40 Rec / 76.56 F1`
Why:
- the earlier DRF bridge peaked early and then degraded
- local evidence suggests the prior branch is currently weak or noisy
- the scratch DRF path was weak, but the recovered author checkpoint is already much better than that path
- DRF deserves one fair test from a strong starting point, not another scratch run
Success criterion:
docs/scoliosis_reproducibility_audit.md
+37
View File
@@ -11,6 +11,7 @@ Related notes:
- [ScoNet and DRF: Status, Architecture, and Reproduction Notes](sconet-drf-status-and-training.md)
- [Scoliosis Training Change Log](scoliosis_training_change_log.md)
- [Scoliosis: Next Experiments](scoliosis_next_experiments.md)
- [DRF Author Checkpoint Compatibility Note](drf_author_checkpoint_compat.md)
Primary references:
- ScoNet paper: [arXiv-2407.05726v3-main.tex](papers/arXiv-2407.05726v3-main.tex)
@@ -155,6 +156,41 @@ Conclusion:
- DRF currently loses to the stronger plain skeleton baseline in this repo
- the published DRF advantage is not established locally
### 9a. The author-provided DRF checkpoint is partially recoverable
This changed one important part of the audit:
- the author checkpoint itself is not unusable
- the earlier very poor local eval was mostly a compatibility failure
The recovered best author-checkpoint path is:
- config: `configs/drf/drf_author_eval_118_aligned_1gpu.yaml`
- result:
- `80.24 Acc / 76.73 Prec / 76.40 Rec / 76.56 F1`
This is still below the paper's DRF headline result:
- `86.0 / 84.1 / 79.2 / 80.8`
But it is far better than the earlier broken compat evals, which means:
- the weight file is real
- the stale author YAML is not a reliable runtime contract
The main causes were:
- split mismatch:
- checkpoint name says `118`
- provided YAML points to `112`
- class-order mismatch:
- author stub uses `negative=0, positive=1, neutral=2`
- repo evaluator assumes `negative=0, neutral=1, positive=2`
- legacy module naming mismatch:
- `attention_layer.*` vs `PGA.*`
- preprocessing/runtime mismatch:
- the checkpoint aligns much better with `Scoliosis1K-drf-pkl-118-aligned`
- it performs very badly on the local `118-paper` export
Conclusion:
- the author checkpoint can be made meaningfully usable in this repo
- but the provided bundle still does not fully specify the original training/eval contract
### 10. The paper-level `1:1:8` skeleton story is not reproduced here
What happened locally:
@@ -254,6 +290,7 @@ Practical bottom line:
- keep the `body-only` skeleton baseline as the mainline path
- keep the retained best checkpoint family as the working artifact
- treat DRF as an optional follow-up branch, not the current winner
- if DRF work continues, the author checkpoint on the aligned `118` path is now a much better starting point than scratch DRF
Current strongest practical checkpoint:
- `92.38 Acc / 90.30 Prec / 87.39 Rec / 88.70 F1`
docs/scoliosis_training_change_log.md
+5
View File
@@ -38,6 +38,10 @@ Use it for:
| 2026-03-10 | `ScoNet_skeleton_112_sigma15_joint8_bodyonly_plaince_bridge_1gpu_10k` | ScoNet-MT-ske bridge | `Scoliosis1K-drf-pkl-118-sigma15-joint8-bodyonly` + `Scoliosis1K_112.json` | Same plain-CE `1:1:2` bridge, relaunched on the `5070 Ti` only per user request | complete | best proxy subset at `7000`: `88.28/69.12/74.15/68.80`; full test at `7000`: `83.16/68.24/80.02/68.47`; final proxy at `10000`: `75.00/65.00/63.41/54.55` (Acc/Prec/Rec/F1) |
| 2026-03-10 | `ScoNet_skeleton_112_sigma15_joint8_headlite_plaince_bridge_1gpu_10k` | ScoNet-MT-ske bridge | `Scoliosis1K-drf-pkl-112-sigma15-joint8-headlite` + `Scoliosis1K_112.json` | Added `head-lite` structure (nose plus shoulder links, no eyes/ears) on top of the plain-CE `1:1:2` bridge; first `3090` launch OOMed due unrelated occupancy, then relaunched on the UUID-pinned `5070 Ti` | complete | best proxy subset at `7000`: `86.72/70.15/89.00/70.44`; full test at `7000`: `78.07/65.42/80.50/62.08` (Acc/Prec/Rec/F1) |
| 2026-03-10 | `DRF_skeleton_112_sigma15_joint8_bodyonly_plaince_bridge_1gpu_10k` | DRF bridge | `Scoliosis1K-drf-pkl-118-sigma15-joint8-bodyonly` + `Scoliosis1K_112.json` | First practical DRF run on the winning `1:1:2` skeleton recipe: `body-only`, plain CE, SGD, `10k` bridge schedule, fixed proxy subset seed `112` | complete | best proxy subset at `2000`: `88.28/61.79/60.31/60.93`; full test at `2000`: `80.21/58.92/59.23/57.84` (Acc/Prec/Rec/F1) |
| 2026-03-14 | `DRF_author_eval_112_1gpu` | DRF author checkpoint compat | `Scoliosis1K-drf-pkl` + `Scoliosis1K_112.json` | Re-evaluated the author-provided checkpoint after adding legacy module-name remap and correcting the author class order; kept the stale `112` path to test whether the provided YAML was trustworthy | complete | `85.19/57.98/56.65/57.30` (Acc/Prec/Rec/F1) |
| 2026-03-14 | `DRF_author_eval_118_splitroot_1gpu` | DRF author checkpoint compat | `Scoliosis1K-drf-pkl-118` + `Scoliosis1K_118.json` | Same author-checkpoint compat path, but switched to the `118` split-specific local DRF dataset root while keeping `BaseSilCuttingTransform` and author class order | complete | `77.17/73.61/72.59/72.98` (Acc/Prec/Rec/F1) |
| 2026-03-14 | `DRF_author_eval_118_aligned_1gpu` | DRF author checkpoint compat | `Scoliosis1K-drf-pkl-118-aligned` + `Scoliosis1K_118.json` | Same author-checkpoint compat path, but evaluated on the aligned `118` DRF export; this is currently the best recovered author-checkpoint runtime contract | complete | `80.24/76.73/76.40/76.56` (Acc/Prec/Rec/F1) |
| 2026-03-14 | `DRF_author_eval_118_paper_1gpu` | DRF author checkpoint compat | `Scoliosis1K-drf-pkl-118-paper` + `Scoliosis1K_118.json` | Tested the author checkpoint against the local paper-literal summed-heatmap export with `BaseSilTransform` to see whether it matched the later paper-style preprocessing branch | complete | `27.24/9.08/33.33/14.27` (Acc/Prec/Rec/F1) |
| 2026-03-10 | `ScoNet_skeleton_112_sigma15_joint8_bodyonly_plaince_main_1gpu_20k` | ScoNet-MT-ske mainline | `Scoliosis1K-drf-pkl-118-sigma15-joint8-bodyonly` + `Scoliosis1K_112.json` | Promoted the winning practical skeleton recipe to a longer `20k` run with full `TEST_SET` eval and checkpoint save every `1000`; no proxy subset, same plain CE + SGD setup | interrupted | superseded by the true-resume continuation below |
| 2026-03-10 | `ScoNet_skeleton_112_sigma15_joint8_bodyonly_plaince_resume_1gpu_20k` | ScoNet-MT-ske mainline | `Scoliosis1K-drf-pkl-118-sigma15-joint8-bodyonly` + `Scoliosis1K_112.json` | True continuation of the earlier plain-CE `1:1:2` `10k` bridge from its `latest.pt`, extended to `20k` with full `TEST_SET` eval and checkpoint save every `1000` | interrupted | superseded by the AdamW finetune branch below |
| 2026-03-10 | `ScoNet_skeleton_112_sigma15_joint8_bodyonly_plaince_adamw_finetune_1gpu_20k` | ScoNet-MT-ske finetune | `Scoliosis1K-drf-pkl-118-sigma15-joint8-bodyonly` + `Scoliosis1K_112.json` | AdamW finetune from the earlier plain-CE `1:1:2` `10k` checkpoint; restores model weights only, resets optimizer/scheduler state, keeps full `TEST_SET` eval and checkpoint save every `1000` | interrupted | superseded by the longer overnight 40k finetune below |
@@ -69,4 +73,5 @@ Current best `ScoNet-MT-ske`-style result:
- A later full-test rerun of the retained `body-only + plain CE` `7000` checkpoint reproduced the same `83.16 / 68.24 / 80.02 / 68.47` result exactly, so that number is now explicitly reconfirmed rather than just carried forward from the original run log.
- `Head-lite` looked stronger than `body-only` on the fixed 128-sequence proxy subset at `7000`, but it did not transfer to the full test set: `78.07 Acc / 65.42 Prec / 80.50 Rec / 62.08 F1`, which is clearly below the `body-only + plain CE` full-test result.
- The first practical DRF bridge on the winning `1:1:2` recipe did not beat the plain skeleton baseline. Its best retained checkpoint (`2000`) reached only `80.21 Acc / 58.92 Prec / 59.23 Rec / 57.84 F1` on the full test set, versus `83.16 / 68.24 / 80.02 / 68.47` for `body-only + plain CE` at `7000`. The working local interpretation is that the added PAV/PGA path is currently injecting a weak or noisy prior rather than a useful complementary signal.
- The author-provided DRF checkpoint turned out to be partially recoverable once the local integration contract was corrected. The main hidden bug was class-order mismatch: the author stub uses `negative=0, positive=1, neutral=2`, while the repo evaluator assumes `negative=0, neutral=1, positive=2`. After adding label-order compatibility and the legacy `attention_layer -> PGA` remap, the checkpoint became meaningfully usable. The best recovered path so far is the aligned `118` export, which reached `80.24 Acc / 76.73 Prec / 76.40 Rec / 76.56 F1`. The stale author YAML is therefore not a trustworthy source of the original runtime contract; the checkpoint aligns much better with an `118` aligned/OpenGait-like path than with the later local `118-paper` export.
- Extending the practical `1:1:2` body-only plain-CE baseline with an `AdamW` cosine finetune changed the picture substantially. The final `80000` checkpoint still evaluated well (`90.64 / 72.87 / 93.19 / 75.74`), but the real win was an earlier retained checkpoint: `27000` reproduced exactly at `92.38 Acc / 90.30 Prec / 87.39 Rec / 88.70 F1` on a standalone full-test rerun. So for this practical path, the best result is no longer the original SGD bridge checkpoint but the retained `AdamW` cosine finetune checkpoint.