Cfg
Namespace: Models
Parent Module: Train
Training configuration.
Record Fields
| Record Field | Description |
BatchSize
Signature: int64
|
batch size |
BestOn
Signature: Partition
|
partition to use for determination of best loss |
CheckpointFile
Signature: string option
|
Path to a checkpoint file (HDF5 format). Used to save the training state if training is interrupted and/or periodically. If the file exists, the training state is loaded from it and training resumes. The string %ITER% in the filename is replaced with the iteration number. |
CheckpointInterval
Signature: int option
|
number of iterations between automatic writing of checkpoints |
DiscardCheckpoint
Signature: bool
|
If true, checkpoint is not loaded from disk. |
DumpPrefix
Signature: string option
|
If set, during each iteration the dump prefix will be set to the given string concatenated with the iteration number. |
LearningRates
Signature: float list
|
Learning rates that will be used. After training terminates with one learning rate, it continues using the next learning rate from this list. |
LoadCheckpointIter
Signature: int option
|
If specified, loads the checkpoint corresponding to the specified iteration. Otherwise, the latest checkpoint is loaded. |
LossRecordFunc
Signature: Entry -> unit
|
function that is called after loss has been evaluated |
LossRecordInterval
Signature: int
|
number of iterations between evaluation of the loss |
MaxIters
Signature: int option
|
maximum training iterations |
MinImprovement
Signature: float
|
minimum loss decrease to count as improvement |
MinIters
Signature: int option
|
minimum training iterations |
PerformTraining
Signature: bool
|
If false, no training is performed after loading the checkpoint. |
Seed
Signature: int
|
seed for parameter initialization |
SlotSize
Signature: int64 option
|
time slot length for sequence training |
TargetLoss
Signature: float option
|
target loss that should lead to termination of training |
Termination
Signature: TerminationCriterium
|
termination criterium |
UserQualityFunc
Signature: int -> UserQualities
|
Function that takes the current iteration number and calculates one or more user-defined quality metrics using the current model state. |
