TI Deep Learning Product User Guide
|
This structure define the parameters Convolution Layer in TIDL.
Data Fields | |
int32_t | weights |
int32_t | bias |
int32_t | perChannelWeightScaleOffset |
int32_t | convolutionType |
int32_t | numInChannels |
int32_t | numOutChannels |
int32_t | numGroups |
int32_t | kernelW |
int32_t | kernelH |
int32_t | strideW |
int32_t | strideH |
int32_t | dilationW |
int32_t | dilationH |
int32_t | padW |
int32_t | padH |
float32_tidl | weightScale |
float32_tidl | biasScale |
int32_t | weightsQ |
int32_t | zeroWeightValue |
int32_t | biasB |
int32_t | biasQ |
int32_t | inDataQ |
int32_t | outDataQ |
int32_t | interDataQ |
int32_t | enableBias |
int32_t | enablePooling |
int32_t | enableEltWise |
int32_t | enableEWRelU |
int32_t | kernelType |
int32_t | enableDepthToSpace |
int32_t | upscaleFactor |
sTIDL_PoolingParams_t | poolParams |
int32_t sTIDL_ConvParams_t::weights |
Offset to the kernel parameters
int32_t sTIDL_ConvParams_t::bias |
Offset to the bias parameters
int32_t sTIDL_ConvParams_t::perChannelWeightScaleOffset |
Offset where per channel weight scales are stored for depthwise convolution layers. A value of zero indicates that this infomation is not valid
int32_t sTIDL_ConvParams_t::convolutionType |
Type of the convolution, Reserved for future use
int32_t sTIDL_ConvParams_t::numInChannels |
Number of input channels channels to be processed
int32_t sTIDL_ConvParams_t::numOutChannels |
Number of output channels to be processed
int32_t sTIDL_ConvParams_t::numGroups |
Number of groups in the convolutions
int32_t sTIDL_ConvParams_t::kernelW |
Kernel width
int32_t sTIDL_ConvParams_t::kernelH |
Kernel height
int32_t sTIDL_ConvParams_t::strideW |
Stride in horizontal direction
int32_t sTIDL_ConvParams_t::strideH |
Stride in vertical direction
int32_t sTIDL_ConvParams_t::dilationW |
Dilation in horizontal direction
int32_t sTIDL_ConvParams_t::dilationH |
Dilation in vertical direction
int32_t sTIDL_ConvParams_t::padW |
Horizontal Padding requirement in number of elements
int32_t sTIDL_ConvParams_t::padH |
Vertical Padding requirement in number of elements
float32_tidl sTIDL_ConvParams_t::weightScale |
Floating point scale on Kernel weights
float32_tidl sTIDL_ConvParams_t::biasScale |
Floating point scale on bias
int32_t sTIDL_ConvParams_t::weightsQ |
Q value of Kernel weights
int32_t sTIDL_ConvParams_t::zeroWeightValue |
value of weights added for dynamic quantSytle
int32_t sTIDL_ConvParams_t::biasB |
Not used
int32_t sTIDL_ConvParams_t::biasQ |
Q value kernel Bias
int32_t sTIDL_ConvParams_t::inDataQ |
Q value expected for in data
int32_t sTIDL_ConvParams_t::outDataQ |
Q value expected for out data
int32_t sTIDL_ConvParams_t::interDataQ |
Q value intermediate output data
int32_t sTIDL_ConvParams_t::enableBias |
Enable/Disable output bias
int32_t sTIDL_ConvParams_t::enablePooling |
Enable/Disable 2x2 Spatial pooling
int32_t sTIDL_ConvParams_t::enableEltWise |
Enable/Disable EltWise
int32_t sTIDL_ConvParams_t::enableEWRelU |
Enable/Disable Relu for EltWise
int32_t sTIDL_ConvParams_t::kernelType |
Defines the different types of optimizations in kernel types supported by TIDL
int32_t sTIDL_ConvParams_t::enableDepthToSpace |
Enable/Disable depth to Space layer
int32_t sTIDL_ConvParams_t::upscaleFactor |
Used only when depth to Space layer is enabled - This is derived from numOutChannels and data Params Out num channels
sTIDL_PoolingParams_t sTIDL_ConvParams_t::poolParams |
Used only if enablePooling is true