| all_vis_tokens: false | |
| as_images: false | |
| batch_size_test: 64 | |
| batch_size_train: 8 | |
| connector_per_text_layer: false | |
| dataset_name: msvd | |
| end_layer_idx: 31 | |
| image_res: 224 | |
| injected_hidden_states: 6 | |
| k_test: 64 | |
| lm_loss_weight: 0.1 | |
| modality: video | |
| num_frames: 8 | |
| num_tries: 1 | |
| num_workers: 4 | |
| optimizer: {lr: 2e-05, opt: adamW, prompt_lr: 1e-05, weight_decay: 0.02} | |
| pretrained_model: /gpfswork/rech/dyf/ugz83ue/.cache/torch/hub/checkpoints/TimeSformer_divST_8x32_224_K600.pyth | |
| prompt_len: 10 | |
| prompt_tuning: true | |
| replace_added_tokens: true | |
| sample_type: rand | |
| schedular: {cooldown_epochs: 0, decay_rate: 1, epochs: 26, lr: 2e-05, min_lr: 1e-06, | |
| sched: cosine, scheduler_groups: 0, warmup_epochs: 4, warmup_lr: 1e-05} | |
| special_answer_token: </a> | |
| start_layer_idx: 19 | |
| test_split: msvd_vqa_test | |
| text_step: 1 | |
| train_split: msvd_vqa_train | |
| unfreeze_text_layer_norm: false | |
| unfreeze_vision_layer_norm: false | |
| use_cache: false | |
| use_vis_prefix: true | |
| val_split: msvd_vqa_val | |
| vision_model_name: timesformer | |
| warm_up: true | |