In this paper, we describe a framework that permits a neural network model trained in Keras, to be modified by inserting custom layers or operation nodes before or after any existing node or layer. Pre-defined custom layer templates are made available for a user to customize as per their needs. Typical applications for such a framework include introspection of a network at runtime for instance to measure statistical parameters like max/min/variance of the activation tensors which can be used to design effective quantization schemes. Other applications include studying the impact of noise on the accuracy of the network, which can be useful for modeling non-idealities and determining end-to-end network performance on systems with resistive memory technologies like Phase Change Memory (PCM) where the weights stored in memory are susceptible to non-idealities from various sources like conductance drift.
Paper
Full text
nn-morph: Simplifying deep neural network manipulation, transformation and analysis
Semantic Scholar · Computer Science · 2023
Abstract
In this paper, we describe a framework that permits a neural network model trained in Keras, to be modified by inserting custom layers or operation nodes before or after any existing node or layer. Pre-defined custom layer templates are made available for a user to customize as per their needs. Typical applications for such a framework include introspection of a network at runtime for instance to measure statistical parameters like max/min/variance of the activation tensors which can be used to design effective quantization schemes. Other applications include studying the impact of noise on the accuracy of the network, which can be useful for modeling non-idealities and determining end-to-end network performance on systems with resistive memory technologies like Phase Change Memory (PCM) where the weights stored in memory are susceptible to non-idealities from various sources like conductance drift.