TI Deep Learning Product User Guide
|
This structure define the parameters of ReLU activation layer in TIDL.
Data Fields | |
int32_t | activationRangeMethod |
int32_t | weightRangeMethod |
float32_tidl | percentileActRangeShrink |
float32_tidl | percentileWtRangeShrink |
float32_tidl | biasCalibrationFactor |
int32_t | biasCalibrationIterations |
int32_t sTIDL_CalibParams_t::activationRangeMethod |
This parameter is only applicable when calibrationOption is TIDL_CalibOptionActivationRange. This option tells the method to be used for activation range collection. Refer eTIDL_ActivationRangeMethod for various supported methods
int32_t sTIDL_CalibParams_t::weightRangeMethod |
This parameter is only applicable when calibrationOption is TIDL_CalibOptionWeightRange. This option tells the method to be used for weights range collection. Refer eTIDL_WeightRangeMethod for various supported methods. It is highly recommended that this option is used with bias calibration otherwise it may result in accuracy degradation
float32_tidl sTIDL_CalibParams_t::percentileActRangeShrink |
This parameter is only applicable when activationRangeMethod is TIDL_ActivationRangeMethodHistogram. This is percentile of the total number of elements in a activation tensor which needs to be discarded from both side of activation distribution. If input is unsigned then this is applied to only one side of activation distribution. For example percentileActRangeShrink = 0.01, means to discard 1/10000 elements from both or one side of activation distribution.
float32_tidl sTIDL_CalibParams_t::percentileWtRangeShrink |
This parameter is only applicable when weightRangeMethod is TIDL_weightRangeMethodHistogram. This is percentile of the total number of elements in a weight filter which needs to be discarded from both side of weight distribution. For example percentileWtRangeShrink = 0.01, means to discard 1/10000 elements from both or one side of weight distribution.
float32_tidl sTIDL_CalibParams_t::biasCalibrationFactor |
This is contribution which is used to update the bias in each iteration based on the difference of actual mean with respect to the mean after quantization
int32_t sTIDL_CalibParams_t::biasCalibrationIterations |
This parameter is only applicable when calibrationOption is TIDL_CalibOptionBiasCalibration. This is the number of iterations for which bias calibration will be iteratively run.