| OUTPUT_DIR='/ccn2/u/honglinc/cwm_checkpoints/ablation_3frame_clumping_debug/' | |
| DATA_PATH_Ktr="/ccn2/dataset/Kinetics700/kinetics_700_train_list.txt" | |
| DATA_PATH_M="/ccn2/dataset/Moments/multi_moment_train_list.txt" | |
| DATA_PATH_E="${HOME}/BBNet/bbnet/models/VideoMAE-main/video_file_lists/ego4d_train_list_320p_chunked_imu.txt" | |
| DATA_PATH_H="/ccn2/dataset/how_to_100m/how_to_100m_train_list.txt" | |
| CUDA_VISIBLE_DEVICES=0 OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=1 \ | |
| --master_addr=10.102.2.137 --master_port=32240 \ | |
| --nnodes=1 --node_rank=0 \ | |
| run_cwm_pretraining.py \ | |
| --data_path_list ${DATA_PATH_H} ${DATA_PATH_Ktr} ${DATA_PATH_M} ${DATA_PATH_E} \ | |
| --mask_type rotated_table \ | |
| --mask_ratio 0.99 \ | |
| --mask_kwargs '{"tube_length": 1}' \ | |
| --model vitbase_8x8patch_3frames_1tube \ | |
| --context_frames 2 \ | |
| --target_frames 1 \ | |
| --temporal_units 'ms' \ | |
| --sampling_rate 150 \ | |
| --context_target_gap 150 150 \ | |
| --batch_size 3 \ | |
| --opt adamw \ | |
| --opt_betas 0.9 0.95 \ | |
| --warmup_epochs 1 \ | |
| --save_ckpt_freq 1 \ | |
| --epochs 19 \ | |
| --no_normlize_target \ | |
| --rescale_size 224 \ | |
| --augmentation_type 'multiscale' \ | |
| --augmentation_scales 1.0 0.875 0.75 0.66 \ | |
| --log_dir ${OUTPUT_DIR} \ | |
| --output_dir ${OUTPUT_DIR} \ | |
| --print_freq 1 \ | |
| --num_workers 0 | |