| batch_size: 8 | |
| data_parallel: false | |
| dataset_name: MP-DocVQA | |
| device: cuda | |
| eval_start: false | |
| freeze_encoder: false | |
| images_dir: /data/users/rperez/MP-DocVQA/images/ | |
| imdb_dir: /data/users/rperez/imdb/collection | |
| lr: 2e-4 | |
| max_pages: 2 | |
| model_name: Hi-VT5 | |
| model_weights: t5-base | |
| page_retrieval: Custom | |
| page_tokens: 10 | |
| retrieval_module: | |
| loss: CrossEntropy | |
| loss_weight: 0.25 | |
| save_dir: save/ | |
| seed: 42 | |
| train_epochs: 10 | |
| visual_module: | |
| model: dit | |
| model_weights: microsoft/dit-base-finetuned-rvlcdip | |
| warmup_iterations: 1000 | |