Downloads · 30 days
0
salaheiyo/MAC2026-track2-salaheiyo1
MAC2026-track2-salaheiyo1 is a machine learning model from salaheiyo. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as mit.
Model checkpoints for reproducing our MMAD Track2 (Multi-label Micro-Action Detection, MMA-52) best submission (detection avg-mAP 0.26762). The code, README and data live in the accompanying reproduction package; this…
Downloads · 30 days
0
Access
Public
Updated Jun 10, 2026
Repo size
4.6 GB
Likes
0
Public
Click a slice to open those files.
.zip4.6 GB · 100%
From the Hugging Face model README
Model checkpoints for reproducing our MMAD Track2 (Multi-label Micro-Action Detection, MMA-52) best submission (detection avg-mAP 0.26762). The code, README and data live in the accompanying reproduction package; this repo only hosts the 16 large checkpoints (4.8 GB), which are impractical to email.
This repo stores everything as a single archive: HF_checkpoints_0.26762.zip.
backbones/ vmae_base_TRACK1_epoch30.pth VideoMAEv2-B finetuned on MA-52 (init for the track2 finetune; regular features)
vmae_base_TRACK2FT.pth VideoMAEv2-B further finetuned on track2 GT segments (t2ft features)
vswin_small.pth VideoSwin-small (v1 sliding-window scores)
members_t2ft_trainval/ lr1e4/i2s4/v3/tmaxer/tridet/pose TAD heads trained on train+val (t2ft) -> produce TEST proposals
members_regular_val/ lr1e4/i2s4/v3/tmaxer/tridet/pose TAD heads trained on train only -> produce VAL proposals (GBDT training)
rescore/ stgcnpp_epoch18.pth ST-GCN++ re-scorer (track2 skeletons)
CHECKPOINTS.md5 md5 manifest for all 16 .pth
pip install -U "huggingface_hub[cli]"
# 1) download the archive
huggingface-cli download salaheiyo/MAC2026-track2-salaheiyo1 HF_checkpoints_0.26762.zip \
--repo-type model --local-dir /tmp/hfdl
# 2) unzip
unzip /tmp/hfdl/HF_checkpoints_0.26762.zip -d /tmp/hfdl
# 3) move the checkpoint dirs into the reproduction package's checkpoints/ , then verify:
SRC=$(dirname "$(find /tmp/hfdl -type d -name backbones | head -1)")
cp -r "$SRC"/backbones "$SRC"/members_* "$SRC"/rescore /path/to/package/checkpoints/
cd /path/to/package/checkpoints && md5sum -c ../CHECKPOINTS.md5 # expect all 16 OK
These checkpoints are needed only for the FULL from-scratch pipeline (Path B in the package
README). The FAST reproduction of 0.26762 does NOT need them — it runs from the shipped
intermediates/ in the reproduction package.