FelTris/pjepa_features

P-JEPA feature archives

Precomputed input features and aligned annotations for the P-JEPA encoders. These are backbone/pooler outputs, before P-JEPA temporal encoding. No raw videos or P-JEPA output caches are included. All feature tensors are float32.

The files use the directory layout expected by the P-JEPA code. That GitHub repository is currently private pending the code release. The plain PyTorch example below works without it; the full training/evaluation recipes require access to the code.

Contents and encoder inputs

Dataset / feature family	Feature extractor	Width	Stored sampling	Videos / splits	Size
LEMON	PL-Stitch ViT-B/16, final CLS	768	4 fps; released P-JEPA recipe takes every fourth token (1 fps)	4,194 pretrain	40.24 GB
Cholec80	Same PL-Stitch	768	1 fps	80; train 01–40, test 41–80	0.57 GB
M2CAI16	Same PL-Stitch	768	1 fps	27 train / 14 test	0.29 GB
Assembly101	Released TSM features, 8-frame input	2048	Stride 8 on a 30-Hz feature timeline: 3.75 fps	204 train / 61 val; camera C10119	3.37 GB
EgoProceL / FACT features	I3D features distributed with FACT	2048	10-Hz source resampled to nominal 4 fps	337 train / 86 test	8.33 GB
EgoProceL / pooled V-JEPA	V-JEPA 2.1 ViT-g + supervised framewise pooler	1408	Nominal 4 fps	337 train / 86 test	5.73 GB

Sizes are decimal GB; total features are approximately 58.54 GB, plus three small CSV segment manifests. Coverage is the archived experimental subset, not a claim to contain every video/view of each original dataset. Both EgoProceL families contain the same 423 video IDs and train/test assignments. Use each family's own manifest for its matching annotation timeline.

EgoProceL camera coverage

Both feature families contain 337 training and 86 test videos. The FACT archive is restricted to the video IDs in the pooled archive. This excludes 491 CMU-Kitchens static-camera recordings (394 train / 97 test) from the original FACT distribution: cameras 7151020 (top), 7151062 (right side), and 8421130 (left/top), according to the upstream camera mapping. All non-CMU video IDs are retained. Both archives include CMU IDs 7150991 and 7150996, following the original pooled experiment's video selection. Counts refer to video files/camera recordings, not independent sessions.

The retained FACT vectors and timestamps are unchanged, and all 116 label IDs are preserved. The released P-JEPA FACT-input checkpoint was trained on the original 731-train / 183-test mixed-view split. The current archive and recipes use the smaller matched subset; training on them does not reproduce that checkpoint's original training data. The pooled checkpoint's data is unchanged. The archives have matching video coverage, but the pooled representation still has supervised feature-extractor provenance.

Feature provenance and timing

• PL-Stitch: pl_lemon.pth, SHA-256 8c4166f24c02a7cb5d10a89ebe179dfb50de25319031f4128af22d57a992d8a3; upstream code commit f4a3b19de71c041d77e091817dbe93a7da168498. Final CLS features precede the projection head. Images are RGB, resized directly to 224×224 using bilinear antialiased interpolation, with ImageNet channel mean/std. Feature computation and storage use float32. This image preprocessing is already applied; pass the stored vectors directly to P-JEPA.
• Assembly101: the official TSM feature release is a supervised feature representation. This archive samples every eighth feature on its 30-Hz annotation timeline; times contains seconds. Do not substitute the video's 60-fps container rate when interpreting annotation frame numbers. The exact upstream extractor checkpoint hash is not recorded in the archive.
• FACT features: I3D vectors repacked from the FACT EgoProceL .npy distribution. FACT names the provider here, not an additional temporal model to run. Features are trimmed to available ground-truth length before resampling. Indices follow floor(arange(0, duration, 0.25) * 10): stored timestamps can alternate between 0.2- and 0.3-second gaps. Use times, not arange(T)/4.
• Pooled V-JEPA: extraction used vjepa2_1_vit_giant_384_framewise, a frozen pretrained V-JEPA 2.1 ViT-g backbone with a supervised framewise pooler trained using FACT step labels (model-epoch=21.ckpt). The extraction configuration uses a 256-pixel test crop, 64-frame non-overlapping windows sampled at 4 fps, and 64 learned frame queries (pooler depth 4, 16 heads). Saved features are frame-pooler outputs before classifier pooling, not raw V-JEPA patch averages. Timestamps follow sampled video frames and can deviate from an exact 0.25-s grid. The final window is padded during extraction; padded outputs are omitted. This representation has supervised provenance even when subsequent P-JEPA fitting uses an SSL loss. Its extractor weights are not part of this dataset.

Download

Download everything into your feature-data root (allow at least 60 GB):

python -m pip install huggingface_hub
export PJEPA_DATA_ROOT=/path/to/pjepa-data
hf download FelTris/pjepa_features --repo-type dataset --local-dir "$PJEPA_DATA_ROOT"
cd "$PJEPA_DATA_ROOT"
sha256sum -c SHA256SUMS

For only Assembly101, for example:

hf download FelTris/pjepa_features --repo-type dataset \
  --include 'assembly101_small/*' 'assembly101_annotations/*' \
  --local-dir "$PJEPA_DATA_ROOT"

The corresponding pairs for EgoProceL are egoprocel_small/fact_npy_features_4fps.pt with egoprocel_annotations/egoprocel_fact_features_segments.csv, and egoprocel_small/pooled_features_fact.pt with egoprocel_annotations/egoprocel_segments.csv. Download every LEMON shard along with its index. Partial downloads will not pass the full-repository checksum check.

Format and loading

Use PyTorch 2.6+; these are packed torch archives rather than a Hugging Face datasets.load_dataset table. manifest.json records file hashes, byte sizes, dimensions, token/video counts, split counts, and extraction metadata.

from pathlib import Path
import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "FelTris/pjepa_features",
    "Cholec80/cholec80_pl_stitch_vitb16_1fps.pt",
    repo_type="dataset",
)
archive = torch.load(path, map_location="cpu", mmap=True, weights_only=True)
video_index = 0
start, end = map(int, archive["offsets"][video_index:video_index + 2])
features = archive["tokens"][start:end]  # [time, 768], float32
times = archive["times"][start:end]      # seconds
labels = archive["phase_labels"][start:end]
video_id = archive["video_ids"][video_index]

For a packed Assembly101/EgoProceL file, use the same tokens, times, and offsets slicing; paths[video_index] identifies the sequence. Labels and intervals come from its matching CSV, joined by video_path (replace a video extension with .pt to match archive paths). CSV official_split supplies the split. Preserve class IDs and background handling from the supplied recipes; EgoProceL uses fact_step_id with background 0, Assembly101 uses action_id. The release loaders handle interval alignment and take grouping.

LEMON's index maps each video through shard_ids and local_indices to a shard's offsets. Resolve shards[shard_id] relative to the downloaded index directory. For 1-fps P-JEPA inputs, slice each video's tokens/times with [::4]. Historical absolute paths in archive metadata document extraction provenance; they are not required files on the user's machine. No raw-video decoder, PL-Stitch installation, FACT installation, or V-JEPA installation is needed.

The released LEMON loader memory-maps shards with a bounded shard cache. The Assembly101/EgoProceL training loader loads its packed archive into host memory; budget RAM for the selected archive plus training buffers. The plain loading example uses memory mapping to avoid eagerly reading the entire tensor.

Protocols and matching P-JEPA checkpoints

Input	Encoder in FelTris/pjepa
PL-Stitch at 1 fps	lemon_plstitch_block32.pt
Assembly101 TSM at 3.75 fps	assembly101_tsm_block64.pt or oracle assembly101_tsm_clip.pt
EgoProceL FACT features	Oracle egoprocel_fact_features_clip.pt
EgoProceL pooled V-JEPA	Oracle egoprocel_pooled_clip.pt

Clip-causal models require ground-truth segment boundaries and are oracle comparisons. Use block-causal models going forward. Block-causal P-JEPA does not require true boundaries for attention; historical training sequence construction and upstream feature learning can still use annotations.

LEMON SSL uses only its pretrain split. Cholec80 development selection fits probes on videos 01–32 and selects on 33–40; 41–80 remain held out. Full Cholec80 probing uses 01–40 / 41–80. M2CAI16 uses all 27 / 14 official train/test videos; inline source_splits records membership. Surgical phase names are included in phase_class_names (7 Cholec80 classes; 8 M2CAI16 classes).

Historical EgoProceL configurations use the split named test for model/probe selection. It must not be described as an untouched final test set. The pooled feature extractor's configuration also uses test for validation. Feature-extractor supervision and the original training data of each saved P-JEPA checkpoint must be accounted for when interpreting comparisons.

Sources and terms

Credit the original datasets and feature methods when using these archives: LEMON, PL-Stitch, Cholec80 and M2CAI16, Assembly101 and its TSM features, FACT's EgoProceL distribution, and V-JEPA 2.1. See LICENSE.md for source-specific terms. The MIT license on P-JEPA code/encoder weights does not relicense the source datasets or annotations.