How a Config File is Structured

Indicates items that can be changed in the configuration. See Catalog for the modules available in each section. If you want to try out & test, copy and use Putting It Together at the bottom. Modules that can be changed in the config include models, optimizers, loss functions, and data augmentation, etc. Other modules that can be changed include the device to be used, how training and test data are split, etc.

Basic Usage

Config can be specified and used during training. In this case, the value of the config is reflected.

python tools/train.py --cfg-name "" ...

General Settings

# Which device to use
DEVICE: "cuda:0"
# Random seed
SEED: 42

IO:
  ...

Input & Output

SEED: ...

IO:
  # Number of input time steps
  LOOKBACK: 48
  # Number of output time steps
  HORIZON: 16

Training

IO:
  ...

TRAINING:
  # How to split train dataset {"col" or "row"}
  TRAIN_DATA_SPLIT: "col"
  # Ratio of train dataset size over val dataset size
  TRAIN_DATA_RATIO: 0.75
  # Number of epochs
  NUM_EPOCHS: 100
  # If True, datasets are split randomly (valid only if TRAIN_DATA_SPLIT = "col")
  RANDOM_SPLIT: False
  # Try to load pretrained model & local scaler in this directory
  PRETRAIN: None

Optimizer

See a full list of optimizers.

TRAINING:
  ...

OPTIMIZER:
  # Optimizer name
  NAME: "Adam"
  # Learning rate
  LR: 0.001

Learning Rate Scheduler

See a full list of learning rate schedulers.

OPTIMIZER:
  ...

SCHEDULER:
  # Scheduler name
  NAME: "CosineAnnealing"

Trainer

SCHEDULER:
  ...

TRAINER:
  # Trainer name
  NAME: "SupervisedTrainer"
  # Maximum gradient norm
  MAX_GRAD_NORM: 1.0
  # Denormalize before computing metric values
  DENORM: False

Model

See a full list of models.

TRAINER:
  ...

MODEL:
  # Model name
  NAME: "Seq2Seq"

Local Scaler

MODEL:
  ...

LOCAL_SCALER:
  # Local scaler name
  NAME: "NOOP"

Loss Function

See a full list of loss functions. Multiple loss functions can be passed.

LOCAL_SCALER:
  ...

LOSSES:
  # Loss function names
  NAMES: ["MSE"]
  # Loss function arguments
  ARGS: [{}]
  # Loss function weights
  WEIGHT_PER_LOSS: [1.0]

Metric

LOSSES:
  ...

METRICS:
  # Metric names
  NAMES: ["RMSE"]
  # Metric arguments
  ARGS: [{}]

Dataset

If i-th sequence is smaller than IO.LOOKBACK, it is zero padded to match IO.LOOKBACK.

Note

Zero padding may reduce the accuracy. To avoid zero padding, set DATASET.BASE_START_INDEX to the same value to 2 * IO.LOOKBACK.

METRICS:
  ...

DATASET:
  # Dataset index starts with this value
  BASE_START_INDEX: 0
  # Last BASE_END_INDEX samples are not used for training
  BASE_END_INDEX: -1

Pipeline

See a full list of transforms.

DATASET:
  ...

PIPELINE:
  # List of transforms
  # Each dictionary must contain `name` and `args` pairs
  # Ex: [{"name": "GaussianNoise", "args": {"mean": 0.0, "std": 0.001}}]
  TRANSFORMS_TRAIN: []
  TRANSFORMS_VALID: []

Scaler

PIPELINE:
  ...

X_SCALER:
  # Scaler for input time series
  NAME: "StandardScaler"

Y_SCALER:
  # Scaler for output time series
  NAME: "StandardScaler"

Dataloader

X_SCALER:
  ...

Y_SCALE:
  ...

DATALOADER:
  # Train dataloader name
  NAME_TRAIN: "DataLoader"
  # Validation dataloader name
  NAME_VALID: "DataLoader"
  # Batch size of train dataset
  BATCH_SIZE_TRAIN: 100
  # Batch size of validation dataset
  BATCH_SIZE_VALID: 100

Logger

DATALOADER:
  ...

LOGGER:
  # Log directory name (if "auto", it is randomly generated)
  LOG_DIR: "auto"

Putting It Together

# Which device to use
DEVICE: "cuda:0"
# Random seed
SEED: 42

IO:
  # Number of input time steps
  LOOKBACK: 48
  # Number of output time steps
  HORIZON: 16

TRAINING:
  # How to split train dataset {"col" or "row"}
  TRAIN_DATA_SPLIT: "col"
  # Ratio of train dataset size over val dataset size
  TRAIN_DATA_RATIO: 0.75
  # Number of epochs
  NUM_EPOCHS: 100
  # If True, datasets are split randomly (valid only if TRAIN_DATA_SPLIT = "col")
  RANDOM_SPLIT: False
  # Try to load pretrained model & local scaler in this directory
  PRETRAIN: None

OPTIMIZER:
  # Optimizer name
  NAME: "Adam"
  # Learning rate
  LR: 0.001

SCHEDULER:
  # Scheduler name
  NAME: "CosineAnnealing"

TRAINER:
  # Trainer name
  NAME: "SupervisedTrainer"
  # Maximum gradient norm
  MAX_GRAD_NORM: 1.0
  # Denormalize before computing metric values
  DENORM: False

MODEL:
  # Model name
  NAME: "Seq2Seq"

LOCAL_SCALER:
  # Local scaler name
  NAME: "NOOP"

LOSSES:
  # Loss function names
  NAMES: ["MSE"]
  # Loss function arguments
  ARGS: [{}]
  # Loss function weights
  WEIGHT_PER_LOSS: [1.0]

METRICS:
  # Metric names
  NAMES: ["RMSE"]
  # Metric arguments
  ARGS: [{}]

DATASET:
  # Dataset index starts with this value
  BASE_START_INDEX: 0
  # Last BASE_END_INDEX samples are not used for training
  BASE_END_INDEX: -1

PIPELINE:
  # List of transforms
  # Each dictionary must contain `name` and `args` pairs
  # Ex: [{"name": "GaussianNoise", "args": {"mean": 0.0, "std": 0.001}}]
  TRANSFORMS_TRAIN: []
  TRANSFORMS_VALID: []

X_SCALER:
  # Scaler for input time series
  NAME: "StandardScaler"

Y_SCALER:
  # Scaler for output time series
  NAME: "StandardScaler"

DATALOADER:
  # Train dataloader name
  NAME_TRAIN: "DataLoader"
  # Validation dataloader name
  NAME_VALID: "DataLoader"
  # Batch size of train dataset
  BATCH_SIZE_TRAIN: 100
  # Batch size of validation dataset
  BATCH_SIZE_VALID: 100

LOGGER:
  # Log directory name (if "auto", it is randomly generated)
  LOG_DIR: "auto"