Rapid-INR: Storage Efficient CPU-free DNN Training Using Implicit Neural Representation
Hanqiu Chen, Hang Yang§, Stephen Fitzmeyer§ and Cong (Callie) Hao
(§Equal contribution)
What is Implicit Neural Representation (INR)?
Flexibility and generality
INR for image compression advantages:
Storage efficiency
Modeling with continuity and smoothness
What is special of Rapid-INR?
Accelerated CPU-free image decoding with pixel-level parallelism
Rapid-INR encoder-decoder architecture
Dynamic pruning and layer-wise quantization
Dynamic pruning ratio calculation
Overview of Rapid-INR
Hanqiu Chen: hanqiu.chen@gatech.edu
Sharc Lab | Georgia Tech | https://sharclab.ece.gatech.edu/
Experiment results
Neural architecture search and hyperparameter tuning
INR reconstructed image quality compared with JPEG
Rapid-INR achieves higher PSNR compared with JPEG
Quantize the hidden layer to 8 bits
Experiment results
Image classification backbone training accuracy and speedup
Up to 6x speedup compared with PyTroch
Up to 1.2x speedup compared with DALI