nuscenes.yaml

output_dir: null  # The output directory where the model predictions and checkpoints will be written.
train_batch_size: 64   # Batch size (per device) for the training dataloader.
num_epochs: 1000   
gradient_accumulation_steps: 1   
use_ema: True   # Whether to use Exponential Moving Average for the final model weights.
learning_rate: 1e-4   
lr_warmup_steps: 500   # Number of steps for the warmup in the lr scheduler.
save_images_epochs: 1 # How often to save images during training.
ddim: True
ddpm_num_inference_steps: 50 
with_vae: True
vae_config: ../vae/configs/nuscenes.yaml
vae_checkpoint: path_to_vae_checkpoint_ckpt_file
block_out_channels: [128, 128, 256, 256]  # unet block out channels
pos_encoding: True
nuscenes: True
range_mean: 50.
range_std: 50.
all_circonv: True
model_config: 
  sample_size: [256, 8]
  in_channels: 5
  out_channels: 4
  layers_per_block: 2
  block_out_channels: [128, 128, 256, 256] 
  down_block_types: [ "DownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D",]
  up_block_types: [ "AttnUpBlock2D", "AttnUpBlock2D", "AttnUpBlock2D", "UpBlock2D",]
resolution: [1024, 32]  # The resolution for input images, all the images in the train/validation dataset will be resized to this resolution

# default parameters
model_config_name_or_path: null  # The config of the UNet model to train, leave as None to use standard DDPM configuration.
eval_batch_size: 16  # The number of images to generate for evaluation.
dataloader_num_workers: 8  # The number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process.
save_model_epochs: 10 # How often to save the model during training.
lr_scheduler: cosine  # The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial", "constant", "constant_with_warmup"]
adam_beta1: 0.95 # The beta1 parameter for the Adam optimizer.
adam_beta2: 0.999 # The beta2 parameter for the Adam optimizer.
adam_weight_decay: 1e-6 # Weight decay magnitude for the Adam optimizer.
adam_epsilon: 1e-08 # Epsilon value for the Adam optimizer.
ema_inv_gamma: 1.0 # The inverse gamma value for the EMA decay.
ema_power: 0.75 # The power value for the EMA decay.
ema_max_decay: 0.9999  # The maximum decay magnitude for EMA.
push_to_hub: False # Whether or not to push the model to the Hub.
hub_token: null # The token to use to push to the Model Hub.
hub_model_id: null # The name of the repository to keep in sync with the local `output_dir`.
hub_private_repo: False # Whether or not to create a private repository.
logger: tensorboard  # Whether to use [tensorboard](https://www.tensorflow.org/tensorboard) or [wandb](https://www.wandb.ai) for experiment tracking and logging of model metrics and model checkpoints"
logging_dir: logs  # [TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
local_rank: -1 # For distributed training: local_rank
mixed_precision: 'no' # "no", "fp16", "bf16" Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10. and an Nvidia Ampere GPU.
prediction_type: epsilon # "epsilon", "sample" Whether the model should predict the 'epsilon'/noise error or directly the reconstructed image 'x0'.
ddpm_num_steps: 1000
ddpm_beta_schedule: linear
checkpointing_steps: 500   # Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming  training using `--resume_from_checkpoint`.
checkpoints_total_limit: 10  # Max number of checkpoints to store.
resume_from_checkpoint: null # Whether training should be resumed from a previous checkpoint. Use a path saved by  `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.
enable_xformers_memory_efficient_attention: False  # Whether or not to use xformers.
snr_gamma: null  # SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. More details here: https://arxiv.org/abs/2303.09556.