| # GLOBAL STUFF | |
| experiment_id: stage_c_3b_lora | |
| checkpoint_path: /path/to/checkpoint | |
| output_path: /path/to/output | |
| model_version: 3.6B | |
| # WandB | |
| wandb_project: StableCascade | |
| wandb_entity: wandb_username | |
| # TRAINING PARAMS | |
| lr: 1.0e-4 | |
| batch_size: 32 | |
| image_size: 768 | |
| multi_aspect_ratio: [1/1, 1/2, 1/3, 2/3, 3/4, 1/5, 2/5, 3/5, 4/5, 1/6, 5/6, 9/16] | |
| grad_accum_steps: 4 | |
| updates: 10000 | |
| backup_every: 1000 | |
| save_every: 100 | |
| warmup_updates: 1 | |
| # use_fsdp: True -> FSDP doesn't work at the moment for LoRA | |
| use_fsdp: False | |
| # GDF | |
| # adaptive_loss_weight: True | |
| # LoRA specific | |
| module_filters: ['.attn'] | |
| rank: 4 | |
| train_tokens: | |
| # - ['^snail', null] # token starts with "snail" -> "snail" & "snails", don't need to be reinitialized | |
| - ['[fernando]', '^dog</w>'] # custom token [snail], initialize as avg of snail & snails | |
| # ema_start_iters: 5000 | |
| # ema_iters: 100 | |
| # ema_beta: 0.9 | |
| webdataset_path: | |
| - s3://path/to/your/first/dataset/on/s3 | |
| - s3://path/to/your/second/dataset/on/s3 | |
| effnet_checkpoint_path: models/effnet_encoder.safetensors | |
| previewer_checkpoint_path: models/previewer.safetensors | |
| generator_checkpoint_path: models/stage_c_bf16.safetensors |