quaterion.train.xbm.xbm_config module

class XbmConfig(weight: Optional[float] = 1.0, buffer_size: Optional[int] = 10000, start_iteration: Optional[int] = 1000, device: Optional[XbmDevice] = XbmDevice.AUTO)

Bases: object

Determine XBM settings.

This class should be returned from configure_xbm()

buffer_size: Optional[int] = 10000

Size of the memory buffer that holds embeddings from previous batches

device: Optional[XbmDevice] = 'auto'

Placement of the buffer

start_iteration: Optional[int] = 1000

Iteration step to start considering the buffer loss

weight: Optional[float] = 1.0

Value to scale the buffer loss before adding it to the final loss

class XbmDevice(value)

Bases: str, Enum

Device selection for placement of the buffer

AUTO = 'auto'

Buffer created in GPU if available. In CPU, otherwise.

CPU = 'cpu'

Buffer created in CPU

CUDA = 'cuda'

Buffer created in GPU