File size: 6,304 Bytes
6dbcfef | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 |
1. OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 training/extract_video_training_latents.py > ./logs/extract_video_latents_16k.log 2>&1 &
2. huggingface cache dir: /root/.cache/huggingface/hub
2. torchhub cache dir: /root/.cache/torch/hub/checkpoints/
2. cache dir: /root/.cache/audioldm_eval/ckpt/Cnn14_16k_mAP=0.438.pth
training
3. OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 train.py exp_id=vgg_only_small_44k model=small_44k > ./logs/train_vgg_only_small_44k.log 2>&1 &
inference
4. OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 batch_eval.py duration_s=8 dataset=vggsound model=small_44k num_workers=8 > ./logs/inference_vgg_only_small_44k.log 2>&1 &
demo example
5. CUDA_VISIBLE_DEVICES=0 python demo.py --variant="small_16k" --duration=4 --video='' --prompt ""
CUDA_VISIBLE_DEVICES=0 python demo.py --variant="small_44k" --duration=10 --video='/inspire/hdd/ws-f4d69b29-e0a5-44e6-bd92-acf4de9990f0/gaopeng/zhoutao-240108120126/kwang/MMAudio/example_demo_videos/_jB-IM_77lI_000000_silent.mp4' --prompt ""
CUDA_VISIBLE_DEVICES=2 python demo.py --variant="small_44k" --duration=4 --video='/inspire/hdd/ws-f4d69b29-e0a5-44e6-bd92-acf4de9990f0/gaopeng/zhoutao-240108120126/kwang/MMAudio/example_demo_videos/demo2.mp4' --prompt ""
6. moviegen
CUDA_VISIBLE_DEVICES=1 python demo.py --variant="small_44k" --duration=11 --video='/inspire/hdd/ws-f4d69b29-e0a5-44e6-bd92-acf4de9990f0/gaopeng/zhoutao-240108120126/kwang/MMAudio/example_demo_videos/moviegen/video1.mp4' --prompt ""
7. audio waveforms for vggsound
/inspire/hdd/ws-f4d69b29-e0a5-44e6-bd92-acf4de9990f0/gaopeng/public/kwang/datasets/vggsound/audios_vggsound
8. demo new for paper
CUDA_VISIBLE_DEVICES=2 python demo.py --variant="small_44k" --duration=4 --video='/inspire/hdd/ws-f4d69b29-e0a5-44e6-bd92-acf4de9990f0/gaopeng/zhoutao-240108120126/kwang/MMAudio/example_demo_videos/demo_new/.mp4' --prompt ""
9 training for vggsound with text caption
OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 train.py exp_id=vgg_only_small_44k_caption_jan26 model=small_44k > ./logs/train_vgg_only_small_44k_caption_jan26.log 2>&1 &
10 generate data for dpo
(1) change eval_for_dpo_config.yaml
(2) change eval_data/base.yaml
OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 batch_eval_dpo.py duration_s=8 dataset=vggsound_dpo model=small_44k num_workers=8 > ./logs_dpo/inference_vgg_only_small_44k_new_model_lumina_v2a_two_stream_Feb28_depth16_caption_dpo.log 2>&1 &
11. nohup python dpo_training/create_dpo_file.py > ./logs/create_dop_file.log 2>&1 &
nohup python reward_models/cavp.py > ./logs/create_dop_cavp_file.log 2>&1 &
12. dpo training
OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 train_dpo.py exp_id=vgg_only_small_44k model=small_44k > ./logs_dpo/train_vgg_only_small_44k.log 2>&1 &
13. extraction features for dpo
OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 dpo_training/extract_video_training_latents.py > ./logs_dpo/extract_video_latents_44k_vggsound_dpo-new_model_lumina_v2a_two_stream_Mar1_depth16_caption_inference_ema_iter1_cavp.log 2>&1 &
14. after generating data for dpo, how to create dpo data file
(1) python ./dpo_training/generated_videos_file.py # get video json file
(2) python ./dpo_training/create_dpo_file.py # get av-align dpo file
(3) python ./reward_models/clap_multi_gpu.py (clap.py) # get clap dpo file
(4) python ./reward_models/cavp.py # get cavp dpo file, note the reencode_video.py should be runned first for 4fps video
# after that, extract video and audio features
(1) change dpo_training/extract_video_training_latents.py file
OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 dpo_training/extract_video_training_latents.py > ./logs_dpo/extract_video_latents_44k_vggsound_dpo-new_model_lumina_v2a_two_stream_Mar12_depth16_caption_10000samples_inference_ema_iter1_cavp.log 2>&1 &
# dpo training # change config files: train_dpo_config.yaml / dpo_base_config.yaml / ./dpo_data/base.yaml
OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 train_dpo.py exp_id=vgg_only_small_44k_lumina_v2a_two_stream_May7_depth16_caption_beta10000 model=small_44k > ./logs_dpo/train_vgg_only_small_44k_lumina_v2a_two_stream_May7_depth16_caption_beta10000.log 2>&1 &
# inference:
(1) change the 'model_path' in ./mmaudio/eval_utils.py file
(2) change the config files: eval_config.yaml, ./eval_data/base.yaml
(3) OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 batch_eval.py duration_s=8 dataset=vggsound model=small_44k num_workers=8 > ./logs_dpo/inference_vgg_only_small_44k_dpo_iter1_cavp_May7_lumina_v2a_two_stream_depth16_caption_2000samples_beta10000.log 2>&1 &
OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 batch_eval.py duration_s=8 dataset=moviegen model=small_44k num_workers=8 > ./logs_dpo/lumina_v2a_moviegen_Sep24_inference_ema.log 2>&1 &
# evaluation
(1) change ./av-benchmark/evaluate.sh
(2) bash evaluate.sh
OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 train_dpo.py exp_id=vgg_only_small_44k_lumina_v2a_two_stream_May17_depth16_caption_beta20000_full_reward_ib_desync_iter3_steps5k model=small_44k > ./logs_dpo/train_vgg_only_small_44k_lumina_v2a_two_stream_May17_depth16_caption_beta20000_full_reward_ib_desync_iter3_steps5k.log 2>&1 &
OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 batch_eval.py duration_s=8 dataset=vggsound model=small_44k num_workers=8 > ./logs_dpo/inference_vgg_only_small_44k_dpo_May17_lumina_v2a_two_stream_depth16_caption_2000samples_beta20000_full_reward_ib_desync_iter3_steps5k.log 2>&1 &
OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 batch_eval_dpo.py duration_s=8 dataset=vggsound_dpo model=small_44k num_workers=8 > ./logs_dpo/inference_vgg_only_small_44k_new_model_lumina_v2a_two_stream_May17_depth16_caption__beta20000_full_reward_ib_desync_iter3_steps5k_for_dpo.log 2>&1 &
OMP_NUM_THREADS=4 nohup torchrun --standalone --nproc_per_node=4 dpo_training/extract_video_training_latents.py > ./logs_dpo/extract_video_latents_44k_vggsound_dpo-new_model_lumina_v2a_two_stream_Mar17_depth16_caption_2000samples_steps5k_iter3_ib_desync.log 2>&1 & |