| # Iquique Transfer and Test | |
| This bundle runs two-condition P/S ControlNet transfer learning and evaluates | |
| generated waveforms with the bundled EQTransformer checkpoint. | |
| ## Files | |
| - `data/processed/iquique/train_3000_two_conditions.h5`: 3000-waveform P/S training data | |
| - `checkpoints/base_checkpoint_epoch_00015_no_high_res_skip.pt`: SWaG base model | |
| - `eqt/best_val_state_dict.pt`: EQT checkpoint used for evaluation | |
| - `configs/config.yaml`: training and evaluation settings | |
| - `src/`: training, generation, diffusion, and EQT evaluation code | |
| Run from the `GenerationArticle` directory: | |
| ```bash | |
| conda activate DiT | |
| bash Train/Transfer_and_Test/run_transfer_and_test.sh --num-gpus 8 | |
| ``` | |
| `--num-gpus` is configurable and defaults to 8. The launcher uses paths inside | |
| this directory, trains with global batch size 256 (32 per GPU with 8 GPUs), | |
| saves checkpoints every 50 epochs, and distributes each 256-sample generation | |
| batch across the selected GPUs before merging the shards. At each milestone it | |
| generates 100 DDPM samples and evaluates them with | |
| the bundled EQT checkpoint using the selected GPUs and a 1.0-second tolerance. | |