| # Configuration: base_config.yaml | |
| # | |
| # Global training parameters. Consumed by training/train_semantic.py, | |
| # training/train_frequency.py, and training/data/{dataset,datamodule}.py. | |
| seed: 42 | |
| image_size: 512 | |
| batch_size: 16 | |
| num_workers: 0 | |
| lr: 0.0001 | |
| # Stream-specific settings. | |
| semantic: | |
| pretrained: true | |
| freeze_backbone: true | |
| unfreeze_last_n_blocks: 2 | |
| output_dim: 1024 | |
| frequency: | |
| pretrained: true | |
| freeze_backbone: false | |
| highpass_ratio: 0.08 | |
| output_dim: 768 | |
| data: | |
| # Path to a manifest CSV with an `image_path,label` header (0=Real, 1=AI-Generated). | |
| # Falls back to an in-memory synthetic dataset for pipeline testing if this | |
| # path doesn't exist yet. Produced by (see training/data/import_hf.py and | |
| # training/data/shuffle_manifest.py): | |
| # python -m training.data.import_hf --split train --output data/sid_subset --limit <N> | |
| # python -m training.data.shuffle_manifest --input data/sid_subset/manifest.csv --output data/sid_subset/manifest_shuffled.csv | |
| manifest_path: data/sid_binary_10000/manifest.csv | |
| num_synthetic_samples: 32 | |
| val_split: 0.2 | |
| log_level: INFO | |