REACT24 / submission /README.md
HJHGJGHHG's picture
Upload 14 files
c7c9ef3 verified

Results

Results on the validation set

Task with latent Top-K Top-P Temp. FRCorr↑ FRDist↓ FRDiv↑ FRVar↑ FRDvs↑ FRSyn↓
Random - - - - 0.0696 200.74 0.3346 0.1673 0.3346 46.36
Mimic - - - - 0.3747 84.93 0.0000 0.0636 0.2117 43.51
online 50 1.0 0.33 0.196 146.5 0.2755 0.1415 0.2871 46.55
8 1.0 0.33 0.1984 145.9 0.2797 0.1421 0.1912 46.39
10 0.8 0.33 0.2042 144.8 0.2782 0.1416 0.2929 46.38
10 0.5 0.33 0.2117 143.6 0.2754 0.1407 0.2933 46.03
5 0.5 0.2 0.21 143.4 0.2757 0.14 0.294 46.65
3 0.33 0.2 0.2083 144.8 0.2754 0.149 0.2951 46.33
2 0.2 0.2 0.2094 145.1 0.2769 0.1411 0.2956 46.27
50 1.0 1.0 0.1633 156.7 0.2826 0.1508 0.2850 46.65
50 1.0 0.33 0.2077 91.21 0.1143 0.0414 0.117 47.29
50 1.0 0.67 0.1996 95.26 0.1371 0.05576 0.1407 46.68
50 1.0 1.0 0.1819 101.9 0.176 0.078 0.1795 46.28
8 0.5 1.0 0.2036 91.23 0.1141 0.0413 0.1171 47.57
10 1.0 1.0 0.1969 96.24 0.1466 0.0611 0.15 46.6
10 1.0 1.33 0.1804 99.4 0.1667 0.0733 0.1697 46.67
50 1.0 1.33 0.1487 111.5 0.2233 0.1045 0.226 46.03
8 0.5 1.33 0.202 92.1 0.1186 0.0437 0.1215 47.23
10 0.8 1.33 0.1965 95.3 0.1377 0.0568 0.1398 46.37
10 0.5 1.5 0.2029 92.89 0.1189 0.0466 0.1223 47.31
offline 50 1.0 0.33 0.1957 138.7 0.2551 0.1306 0.267 46.56
10 1.0 1.0 0.2038 93.85 0.1334 0.0545 0.1374 46.31
50 1.0 1.0 0.1801 102.1 0.1785 0.0786 0.1816 46.18

†: self-producted baseline results

Results on the test set (Online task)

Model with latent Top-K Top-P Temp. FRCorr↑ FRDist↓ FRDiv↑ FRVar↑ FRDvs↑ FRSyn↓
Random - - - - 0.05 237.21 0.1667 0.0833 0.1667 43.84
Mimic - - - - 0.38 92.94 0.0000 0.0724 0.2483 38.54
Trans-VAE - - - - 0.03 92.81 0.0008 0.0002 0.0006 43.75
BeLFusion (k=10) - - - - 0.12 91.60 0.0105 0.0082 0.0116 44.87
REGNN - - - - 0.19 84.54 0.0007 0.0061 0.0342 41.35
× 50 0.8 0.33 0.128 98.18 0.1343 0.0647 0.1368 43.95
× 50 1.0 0.33
× 10 1.0 1.0 0.1543 95.2 0.131 0.0534 0.1346 43.84
× 10 0.8 1.33 0.1553 94.45 0.1239 0.0504 0.1273 43.89
SUBMISSION1 × 8 0.5 1.0 0.1587 92.05 0.1029 0.0387 0.1065 44.52
× 10 1.0 1.33 0.1633 96.37 0.0825 0.0404 0.0839 44.02
SUBMISSION2 10 0.5 0.33 0.1436 135.5 0.303 0.139 0.2878 44.05
SUBMISSION3 - - - - 0.1609 104.3 0.1706 0.055 0.1506 44.71

Results on the test set (Offline task)

Model with latent Top-K Top-P Temp. FRCorr↑ FRDist↓ FRDiv↑ FRVar↑ FRDvs↑ FRSyn↓
Random - - - - 0.05 237.21 0.1667 0.0833 0.1667 43.84
Mimic - - - - 0.38 92.94 0.0000 0.0724 0.2483 38.54
Trans-VAE - - - - 0.03 92.81 0.0008 0.0002 0.0006 43.75
BeLFusion (k=10) - - - - 0.12 91.60 0.0105 0.0082 0.0116 44.87
REGNN - - - - 0.19 84.54 0.0007 0.0061 0.0342 41.35
× 10 0.8 1.33 0.1638 99.8 0.1503 0.0482 0.1552 44.3
× 10 1.0 0.67 0.168 98.08 0.131 0.0415 0.1379 44.61
SUBMISSION1 × 10 0.5 0.67 0.1664 93.97 0.1018 0.0325 0.1 44.71
× 8 0.5 1.0 0.168 96.82 0.1188 0.03560 0.1191 44.71
SUBMISSION2 5 0.5 0.2 0.139 140.5 0.3059 0.1409 0.2881 44.19
SUBMISSION3 - - - - 0.1692 104 0.1675 0.0535 0.1385 44.71

NOTE: results of the submission3 is a combination of submission1 (k=8) and submission2 (k=2).

The listener reaction videos are generated using the same setting of online submission1 on the val dataset.

The generation config .yaml file and the logging files are provided in the raw_results folder. To compute the metrics, run

python compute_metrics.py --n_threads 24  --phase test --decoded_labels_dir submission\{task}\submission{1/2/3}\raw_results

To re-generate the results, pute the generation_config.yaml in the recipes folder and run

CUDA_VISIBLE_DEVICES=4,5,6,7 accelerate launch --config_file ../recipes/accelerate_config.yaml --num_processes 4 generate_results.py