RobertRussell/Puppeteer-human-skeleton-selected

Selected human skeleton checkpoints

This repository preserves four selected historical Puppeteer-style human skeleton checkpoints, their evaluation summaries and visualization samples. The primary checkpoint is hard4_full_valid_1836_step3300.pth: it achieved 1,836 valid predictions out of 1,836 evaluated samples in its recorded test run. These are original training checkpoint files, not newly trained models or converted Hugging Face Transformers models.

Selection and recorded results

The two explicitly preferred visual checkpoints are user_favorite_1836_step400.pth and user_favorite_1836_step500.pth, both from the March 13 run. The hard4 checkpoint is retained separately for its full-validity record. Numerical validity is not treated as a substitute for visual quality. A further manually selected low-error alternative is included. Distances are historical averages over valid predictions; lower is better. The evaluation summaries are archived under evaluation/, and example images under previews/.

Checkpoint	Role	Valid predictions	J2J	J2B	B2B
hard4_full_valid_1836_step3300.pth	Primary; hard4 fine-tuning	1836/1836	0.0187	0.0138	0.0113
user_favorite_1836_step400.pth	Explicit user visual favorite	1833/1836	0.0185	0.0136	0.0113
user_favorite_1836_step500.pth	Explicit user visual favorite	1828/1836	0.0185	0.0138	0.0113
manual_favorite_1412_step1200.pth	Manually selected low-error alternative	1405/1412	0.0104	0.0081	0.0066

The 1,836-sample and 1,412-sample cohorts are different historical evaluation settings. Their distances must not be pooled into one ranking or interpreted as directly comparable. The manually selected alternative has seven failed samples in its recorded 1,412-sample evaluation, so it is not the full-validity default. The older 1,412-sample membership is not reconstructed by simply taking the first 1,412 entries of the currently distributed test file. Matching sample counts also do not by themselves establish identical historical preprocessing or evaluation settings across separate runs.

The legacy step1700 checkpoint and its preview were withdrawn from the current repository version following visual inspection. Its numerical summary remains under evaluation/ as a historical record, not a model recommendation.

On the primary checkpoint's own 1,836-sample report, the original puppeteer_skeleton_w_diverse_pose baseline records 0.0213 / 0.0163 / 0.0135 for J2J / J2B / B2B, also with 1836/1836 valid. These values come from that same archived report, not a fresh inference run.

Training provenance

Published name	Original run	Original filename
hard4_full_valid_1836_step3300.pth	finetune_human_refined_1e-6_1e-7_03_15_02-24-25	checkpoint_3300.pth
user_favorite_1836_step400.pth	finetune_human_refined_1e-6_1e-7_03_13_01-17-17	checkpoint_400.pth
user_favorite_1836_step500.pth	finetune_human_refined_1e-6_1e-7_03_13_01-17-17	checkpoint_500.pth
manual_favorite_1412_step1200.pth	finetune_human_refined_20_09-44-15	checkpoint_1200.pth

The primary training command records six GPU processes, batch size 80 per GPU, 10 maximum epochs, base/final learning rates 1e-6 / 1e-7, maximum 101 joints, and initialization from puppeteer_skeleton_w_diverse_pose.pth. Its training input was train_human_final_hard4_aug.npz; the associated validation and test files were val_human.npz and test_human.npz. Those datasets and the augmentation implementation are archived in the data repository.

File format and loading

Each .pth is an original PyTorch training checkpoint containing model, optimizer, epoch, args, and best_val_metrics, rather than just a bare state dictionary. Optimizer state is retained for archival fidelity. The model requires the matching custom SkeletonGPT architecture and point-encoder dependencies; these files are not loadable with a generic AutoModel.from_pretrained call.

from huggingface_hub import hf_hub_download
import torch

path = hf_hub_download(
    repo_id="RobertRussell/Puppeteer-human-skeleton-selected",
    filename="hard4_full_valid_1836_step3300.pth",
)
# Only load checkpoints you trust: args contains a pickled argparse Namespace.
checkpoint = torch.load(path, map_location="cpu", weights_only=False)
state_dict = checkpoint["model"]
# Instantiate the matching SkeletonGPT implementation, remove a DDP 'module.'
# prefix if present, and check state-dict compatibility before inference.

Use SHA256SUMS to verify downloads. selection_metadata.json maps the public filenames to their original runs, metrics and content hashes. No new inference, weight conversion or model fusion was performed during this archival upload. Evaluation records demonstrate successful historical loading, not a newly validated environment setup.

Attribution

This is an unofficial archive of fine-tuned checkpoints based on Puppeteer and data derived from Articulation XL2. Retain the applicable upstream model, code, component and data notices; the dataset's CC BY 4.0 declaration is not itself a license declaration for every model component.