Amp

Auto mixed precision api.

class mindnlp.common.amp.DynamicLossScaler(scale_value, scale_factor, scale_window)[source]

Bases: LossScaler

Dynamic LossScaler

adjust(grads_finite)[source]: adjust scale value.

scale(inputs)[source]: scale inputs tensor.

unscale(inputs)[source]: unscale inputs tensor.

class mindnlp.common.amp.LossScaler(scale_value)[source]

Bases: object

Basic LossScaler.

adjust(grads_finite)[source]: adjust scale value.

scale(inputs)[source]: scale inputs tensor.

unscale(inputs)[source]: unscale inputs tensor.

class mindnlp.common.amp.NoLossScaler[source]

Bases: LossScaler

No LossScaler

adjust(grads_finite)[source]: adjust scale value.

scale(inputs)[source]: scale inputs tensor.

unscale(inputs)[source]: unscale inputs tensor.

class mindnlp.common.amp.StaticLossScaler(scale_value)[source]

Bases: LossScaler

Static LossScaler.

adjust(grads_finite)[source]: adjust scale value.

scale(inputs)[source]: scale inputs tensor.

unscale(inputs)[source]: unscale inputs tensor.

mindnlp.common.amp.all_finite(inputs)[source]: whether all inputs tensor are finite.

mindnlp.common.amp.auto_black_list(network, black_list=None)[source]: auto cast based on black list

mindnlp.common.amp.auto_mixed_precision(network, amp_level='O1')[source]: auto mixed precision cast.

mindnlp.common.amp.auto_white_list(network, white_list=None)[source]: auto cast based on white list