image_finetune.yaml 1.2 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748
  1. image_finetune: true
  2. output_dir: "outputs"
  3. pretrained_model_path: "models/StableDiffusion/stable-diffusion-v1-5"
  4. noise_scheduler_kwargs:
  5. num_train_timesteps: 1000
  6. beta_start: 0.00085
  7. beta_end: 0.012
  8. beta_schedule: "scaled_linear"
  9. steps_offset: 1
  10. clip_sample: false
  11. train_data:
  12. csv_path: "/mnt/petrelfs/guoyuwei/projects/datasets/webvid/results_2M_val.csv"
  13. video_folder: "/mnt/petrelfs/guoyuwei/projects/datasets/webvid/2M_val"
  14. sample_size: 256
  15. validation_data:
  16. prompts:
  17. - "Snow rocky mountains peaks canyon. Snow blanketed rocky mountains surround and shadow deep canyons."
  18. - "A drone view of celebration with Christma tree and fireworks, starry sky - background."
  19. - "Robot dancing in times square."
  20. - "Pacific coast, carmel by the sea ocean and waves."
  21. num_inference_steps: 25
  22. guidance_scale: 8.
  23. trainable_modules:
  24. - "."
  25. unet_checkpoint_path: ""
  26. learning_rate: 1.e-5
  27. train_batch_size: 50
  28. max_train_epoch: -1
  29. max_train_steps: 100
  30. checkpointing_epochs: -1
  31. checkpointing_steps: 60
  32. validation_steps: 5000
  33. validation_steps_tuple: [2, 50]
  34. global_seed: 42
  35. mixed_precision_training: true
  36. enable_xformers_memory_efficient_attention: True
  37. is_debug: False