Effects of Enabling or Disabling Custom Deep Learning Processor Layer Modules

Analyze how deep learning processor layer modules affect custom bitstream generation. Identify when to enable or disable modules in order to speed up custom bitstream generation.

This table lists the deep learning processor configuration modules and when you should enable or disable these modules to optimize custom bitstream generation.

Deep Learning Processor Parameter	Deep Learning Processor Module	Reason to Enable Module	Reason to Disable Module
ModuleGeneration	conv	When network contains convolution or pooling layers.	Reduce FPGA resource utilization when the network does not contain any convolution or pooling layers.
LRNBlockGeneration	conv	The network uses one or more cross-channel normalization layers.	The network does not use cross-channel normalization layers.
SegmentationBlockGeneration	conv	The network contains one or more max unpooling layers.	Reduce bitstream resource utilization or the network does not contain any fully connected layers.
ModuleGeneration	fc	The network has a fully connected layer.	The network has no fully connected layer.
SoftmaxBlockGeneration	fc	Implement the softmax layer in hardware.	Implement the softmax layer in software.
ModuleGeneration	custom	The network has a custom layer or any layers supported by the custom processor module.	The network does not contain any custom layers or layers supported by the custom processor module.
Addition	custom	The network has addition operations.	The network has no addition operations.
MishLayer	custom	The network uses a mish activation layer.	The network does not use a mish activation layer.
Multiplication	custom	The network has layers that perform multiplication operations.	The network has no layers that perform multiplication operations.
Resize2D	custom	The network has a resize layer.	The network does not have a resize layer.
Sigmoid	custom	The network has a sigmoid layer.	The network does not have sigmoid layers.
SwishLayer	custom	The network has a swish activation layer.	The network does not have a swish activation layer.
TanhLayer	custom	The network has a hyperbolic tangent activation layer.	The network does not have a hyperbolic tangent activation layer.