kunli-cs/MiGA26_Track1_RGB_FT

2026 MIGA Track 1 Baseline: Linear Probe

This package is a challenge distribution version of the imigue3k linear probe baseline. It is reduced to the files needed for Track 1 classification.

What This Baseline Does

• loads a pretrained VideoMAE-style backbone
• freezes the backbone
• trains only the classification head
• optionally trains fc_norm
• evaluates with multi-segment and multi-crop testing

The main entrypoint is run_class_linearprobe.py.

Included Files

• run_class_linearprobe.py: linear probe training entrypoint
• datasets.py: reduced dataset router for imigue3k
• imigue3k_mp4.py: iMiGUE dataset loader
• engine_for_linearprobe.py: train / val / test loop
• modeling_finetune.py: VideoMAE finetuning model definition
• optim_factory.py: optimizer helpers
• utils_mae.py: distributed training, checkpointing, schedulers, logging
• mixup.py, random_erasing.py, rand_augment.py
• video_transforms.py, volume_transforms.py, functional.py
• linearprobe_imigue3k.sh: simple launch script

Expected Dataset Layout

<data_path>/
  train/<video_id>/<video_id>_<clip_idx>.mp4
  train/<video_id>/<video_id>_label.csv
  valid/<video_id>/<video_id>_<clip_idx>.mp4
  valid/<video_id>/<video_id>_label.csv
  test/<video_id>/<video_id>_<clip_idx>.mp4
  test/<video_id>/<video_id>_label.csv

label.csv is expected to contain clip-level labels. The loader converts labels to 0-based class ids and treats class 99 as class 32.

Quick Start

cd 2026miga_track1_baseline_linearprobe
DATA_PATH=/path/to/rgb_reorg \
MODEL_PATH=/path/to/pretrained_checkpoint.pth \
bash linearprobe_imigue3k.sh
MASTER_PORT=29502 sh linearprobe_imigue3k.sh

Recommended Default Setting

The provided launch script uses the following default recipe:

• backbone: vit_base_patch16_224
• frames: 16
• sampling rate: 2
• optimizer: AdamW
• learning rate: 3e-3
• weight decay: 1e-4
• epochs: 50
• batch size: 32
• --lp_train_fc_norm
• --lp_reset_head

Notes For Participants

• This baseline is intentionally minimal and only supports imigue3k.
• The backbone is frozen in linear probe mode.
• Final test metrics are merged from per-rank prediction files.
• You are free to modify augmentations, optimizer settings, head design, or evaluation settings as long as they remain within the challenge rules.

Environment

This package still depends on the original Python stack, especially:

• torch
• torchvision
• timm
• decord
• scipy
• tensorboardX