Skip to content

training_input.md

Latest commit

 

History

History
103 lines (88 loc) · 5.45 KB

training_input.md

File metadata and controls

103 lines (88 loc) · 5.45 KB
Raw

INPUT PARAMETERS

global configs

  • START: Literal[0, 1, 2], whether to load the model checkpoint file.

    (

    1. 0: Default, from scratch.
    2. 1: load checkpoint of model parameters, optimizer state, current minimum validation error, and lr_scheduler state (if set) from LOAD_CHK_FILE_PATH.
    3. 2: only load model parameters from LOAD_CHK_FILE_PATH.

    )

  • DEVICE: str, the device that model and data run on. It is the same as device of torch.Tensor. Default: 'cpu'.

  • VERBOSE: int, to control the output frequency and details. 0 for silence mode, 1 for brief mode. Larger number means more detailed output information.

  • EPOCH: int, number of total epochs that model will train. Default: 0.

  • BATCH_SIZE: int, batch size of training samples. Default: 1.

  • VAL_BATCH_SIZE: int, batch size of validation samples. Default: the same as BATCH_SIZE.

  • VAL_PER_STEP: int, the validation frequency. Model will validate per VAL_PER_STEP training step. Default: 10.

  • VAL_IF_TRN_LOSS_BELOW: float, not validate until training loss is less than VAL_IF_TRN_LOSS_BELOW. Default: inf

  • ACCUMULATE_STEP: int, the step number of gradient accumulation. Default: 1.

  • DEBUG_MODE: bool, whether to turn on the debug mode, which will output the range of parameters, parameter gradients and gradient differences in each layer, and turn on the NaN check. Default: False.

  • CHECK_NAN: bool, whether to check NaN after inputting each batch during training. Default: False

I/O configs

  • REDIRECT: bool, whether to output training logs to OUTPUT_PATH or directly print it to the screen. Default: True.
  • SAVE_CHK: bool, whether to save the checkpoint file of the training model. Default: False.
  • LOAD_CHK_FILE_PATH: str, only work when START == 1. The path of checkpoint file to load.
  • STRICT_LOAD: bool, to control the parameter strict in torch.Module.load_state_dict.
  • CHK_SAVE_PATH: str, only work when SAVE_CHECK == True. The directory of checkpoint file to save. Default: "./".
  • CHK_SAVE_POSTFIX: str, the postfix of checkpoint file. Default: "". The checkpoint file name will be "best_checkpoint_CHK_SAVE_POSTFIX.pt".
  • OUTPUT_PATH: str, only work when REDIRECT == True. The output directory of training log. Default: "./".
  • OUTPUT_POSTFIX: str, the postfix of training log file. Default: "Untitled". The log file name will be f"time.strftime("%Y%m%d_%H_%M_%S")_OUTPUT_POSTFIX.out".

loss configs

  • LOSS: Literal["MSE", "MAE", "Huber", "CrossEntropy", "Energy_Force_Loss", "Energy_Loss"]

    (

    1. "MSE": nn.MSELoss
    2. "MAE": nn.L1Loss
    3. "Hubber": nn.HuberLoss
    4. "CrossEntropy": nn.CrossEntropyLoss
    5. "Energy_Force_Loss": $loss = coeff_E * loss_E(x, y) + coeff_F * loss_F(x, y)$
    6. "Energy_Loss": $loss = coeff_E * loss_E(x, y)$
    7. "custom": Any, one need to set custom loss function manually by Trainer.set_loss_fn(loss_fn, loss_config: Optional[Dict] = None)

    )

  • LOSS_CONFIG: Dict, the kwargs of LOSS.

    1. for LOSS == "Energy_Force_Loss":
      • loss_E: Literal["MAE", "MSE"], the loss function of energies.
      • loss_F: Literal["MAE", "MSE"], the loss function of forces.
      • coeff_E: float, coefficient of energy loss.
      • coeff_F: float, coefficient of force loss.
    2. for LOSS == "Energy_Loss":
      • loss_E: Literal["MAE", "MSE", "SmoothMAE", "Hubber"], the loss function of energies.

  • METRICS: Literal["E_MAE", "F_MAE", "F_MaxE", "E_R2", "MSE", "MAE", "R2", "RMSE"], the metrics function of training and validation results.

model configs

  • MODEL_NAME: str, the model name.
  • MODEL_CONFIG: Dict, the hyperparameters of model.

optimizer configs

  • OPTIM: Literal["Adam", "SGD", "AdamW", "Adadelta", "Adagrad", "ASGD", "Adamax", "custom"], the model optimizer.

(

  1. 'Adam': th.optim.Adam,
  2. 'SGD': th.optim.SGD,
  3. 'AdamW': th.optim.AdamW,
  4. 'Adadelta': th.optim.Adadelta,
  5. 'Adagrad': th.optim.Adagrad,
  6. 'ASGD': th.optim.ASGD,
  7. 'Adamax': th.optim.Adamax,
  8. 'custom': Any, need to set custom optimizer manually by Trainer.set_optimizer(self, optimizer, optim_config: Optional[Dict])

)

  • OPTIM_CONFIG: Dict, the kwargs of model optimizer.
  • GRAD_CLIP: bool, whether to use gradient clip. Default: false.
  • GRAD_CLIP_MAX_NORM: float, only for GRAD_CLIP == True. The max norm of gradient to clip. Default: 100.
  • LR_SCHEDULER: Literal[{'StepLR', 'ExponentialLR', 'ChainedScheduler', 'ConstantLR', 'LambdaLR', 'LinearLR', 'custom'}], the learning rate scheduler.

(

  1. 'StepLR': torch.optim.lr_scheduler.StepLR,
  2. 'ExponentialLR': torch.optim.lr_scheduler.ExponentialLR,
  3. 'ChainedScheduler': torch.optim.lr_scheduler.ChainedScheduler,
  4. 'ConstantLR': torch.optim.lr_scheduler.ConstantLR,
  5. 'LambdaLR': torch.optim.lr_scheduler.LambdaLR,
  6. 'LinearLR': torch.optim.lr_scheduler.LinearLR,
  7. 'custom': Any, need to set custom lr_scheduler manually by Trainer.set_lr_scheduler(self, lr_scheduler, lr_scheduler_config)

)

  • LR_SCHEDULER_CONFIG: Dict, the kwargs of lr scheduler. Default: dict().
  • EMA: bool, whether to apply the exponential moving average strategy for training. Note that the best_checkpoint will save with EMA parameters, while checkpoint and stop_checkpoint will not. Default: False.
  • EMA_DECAY: float, the decay coefficient of EMA. Default: 0.999.