hhieupt/mer2026-features

git clone https://github.com/zeroQiaoba/MERTools.git
cd MERTools/MER2026/MER2026_Track1

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install numpy scikit-learn omegaconf pandas tqdm matplotlib openai fire huggingface_hub
pip install opencv-python-headless pytorchvideo ftfy timm einops decord regex iopath

HF_DIR=/path/to/download/mer2026-hf   # thư mục tải về từ HuggingFace

python3 - <<EOF
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="hhieupt/mer2026-features",
    repo_type="dataset",
    local_dir="$HF_DIR",
    # token="YOUR_HF_TOKEN",  # bỏ comment nếu repo private
)
print("Download complete!")
EOF

python3 - <<EOF
from huggingface_hub import hf_hub_download

files = [
    "track1_label_6way.npz",
    "track1_track2_candidate.csv",
    "features.zip",                          # UTT features (~1.12 GB)
    # "wavlm-large-FRA.zip",                 # FRA audio (~18.9 GB)
    # "chinese-roberta-wwm-ext-large-FRA.zip", # FRA text (~5.96 GB)
    # "clip-vit-large-patch14-FRA.zip",      # FRA video (~5 GB)
]

for f in files:
    path = hf_hub_download(
        repo_id="hhieupt/mer2026-features",
        filename=f,
        repo_type="dataset",
        local_dir="$HF_DIR",
    )
    print(f"Downloaded: {path}")
EOF

DATA_DIR=/path/to/your/mer2026   # thư mục chứa dữ liệu đã xử lý
mkdir -p $DATA_DIR/embeddings

cp $HF_DIR/track1_label_6way.npz        $DATA_DIR/
cp $HF_DIR/track1_track2_candidate.csv  $DATA_DIR/

# UTT features (nhỏ, nhanh, phù hợp để bắt đầu)
unzip $HF_DIR/features.zip -d $DATA_DIR/embeddings/

# FRA features (lớn hơn, nhiều thông tin hơn)
unzip $HF_DIR/wavlm-large-FRA.zip                    -d $DATA_DIR/embeddings/
unzip $HF_DIR/chinese-roberta-wwm-ext-large-FRA.zip  -d $DATA_DIR/embeddings/
unzip $HF_DIR/clip-vit-large-patch14-FRA.zip         -d $DATA_DIR/embeddings/

$DATA_DIR/embeddings/
├── features/
│   ├── chinese-hubert-large-UTT/           # audio UTT (từ features.zip)
│   ├── chinese-macbert-large-UTT/          # text UTT (từ features.zip)
│   ├── clip-vit-large-patch14-UTT/         # video UTT (từ features.zip)
├── wavlm-large-FRA/                    # audio FRA
├── chinese-roberta-wwm-ext-large-FRA/  # text FRA
├── clip-vit-large-patch14-FRA/         # video FRA
├── track1_label_6way.npz
└── track1_track2_candidate.csv

chinese-hubert-large-UTT/
├── sample_00001.npy   # shape: (1024,) hoặc (1, 1024)
├── sample_00002.npy
└── ...

import os

DATA_DIR = {
    'MER2026Raw': '/path/to/your/mer2026',   # submission.py đọc track_all_candidates.csv từ key này (set cả 2 bằng đường dẫn DATA_DIR bên trên luôn)
    'MER2026':    '/path/to/your/mer2026',
}

PATH_TO_LABEL    = {'MER2026': os.path.join(DATA_DIR['MER2026'], 'track1_label_6way.npz')}
PATH_TO_FEATURES = {'MER2026': os.path.join(DATA_DIR['MER2026'], 'embeddings/features')}

python main-release.py \
  --dataset=MER2026 \
  --model=attention \
  --feat_type=utt \
  --audio_feature=chinese-hubert-large-UTT \
  --text_feature=chinese-macbert-large-UTT \
  --video_feature=clip-vit-large-patch14-UTT \
  --epochs=100 \
  --gpu=0 \
  --save_root=./saved-utt \
  > train_utt.log 2>&1 &

echo "Training started, PID: $!"
tail -f train_utt.log

python main-release.py \
  --dataset=MER2026 \
  --model=attention \
  --feat_type=frm_unalign \
  --audio_feature=wavlm-large-FRA \
  --text_feature=chinese-roberta-wwm-ext-large-FRA \
  --video_feature=clip-vit-large-patch14-FRA \
  --epochs=100 \
  --gpu=0 \
  --save_root=./saved-fra \
  > train_fra.log 2>&1 &

echo "Training started, PID: $!"
tail -f train_fra.log

ModuleNotFoundError: No module named 'torchvision.transforms.functional_tensor'

python3 -c "
import torchvision, os
path = os.path.join(os.path.dirname(torchvision.__file__), 'transforms', 'functional_tensor.py')
with open(path, 'w') as f:
    f.write('from torchvision.transforms.functional import *\n')
print('Created:', path)
"

epoch:1; metric:emo; train results:0.5628; eval results:0.7122
epoch:2; metric:emo; train results:0.6691; eval results:0.7369
...

./saved-utt-trimodal/result/
├── cv_..._f1:0.XXXX_acc:0.XXXX_....npz     # CV results (tham khảo)
└── test1_..._f1:0.XXXX_acc:0.XXXX_....npz  # Test predictions (dùng để submit)

cp /path/to/your/mer2026/track1_track2_candidate.csv /path/to/your/mer2026/track_all_candidates.csv

ls ./saved-utt-trimodal/result/test1_*.npz

NPZ=$(ls ./saved-utt-trimodal/result/test1_*.npz | head -1)
echo "Using: $NPZ"

python submission.py generate_submission \
  --result_npz="$NPZ" \
  --save_csv=answer.csv

head -5 answer.csv
# name,discrete
# sample_00001,neutral
# sample_00002,happy
# ...

wc -l answer.csv   # phải là 20001 (20000 mẫu + 1 dòng header)

zip answer.zip answer.csv