LiME-TM: A Lightning-fast and Memory-Efficient ML Framework
for On-Device Training on MCUs
Han Wu, Shengyu Duan and Tousif
EMG and Why TM?
Vanilla TM: Memory Consumption
- 100 clauses
- 10 classes
- 28 x 28 pixels
- 8 bits
100 clauses x 10 classes x 28 x 28 pixels x 8 bits
= 12,544,000 Bytes (~ 12MB)
- On-Device Training: 20 KB SRAM
- On-Device Inference: 2 KB SRAM
Vanilla TM (12 MB) --> LiME-TM Pruning (ESP32 520 KB)
LiME-TM: Pruning (ESP32)
Vanilla TM (12 MB) --> LiME-TM Pruning (ESP32 520 KB) --> LiME-TM Compiler (STM32F4 192 KB)
LiME-TM: Compiler (STM32F4)
Vanilla TM (12 MB) --> LiME-TM Pruning (ESP32 520 KB) --> LiME-TM Compiler (STM32F4 192 KB) --> LiME-TM Compression (STM32L4 128 KB)
LiME-TM: Compression (STM32L4)
Vanilla TM (12 MB) --> LiME-TM Pruning (ESP32 520 KB) --> LiME-TM Compiler (STM32F4 192 KB) --> LiME-TM Compression (STM32L4 128 KB) --> LiME-TM Bare-metal (STM32F103 20 KB)
LiME-TM: Bare-metal (STM32F103)
Vanilla TM (12 MB) --> LiME-TM Pruning (ESP32 520 KB) --> LiME-TM Compiler (STM32F4 192 KB) --> LiME-TM Compression (STM32L4 128 KB) --> LiME-TM Bare-metal (STM32F103 20 KB) --> LiME-TM Inference-Only (Arduino 2 KB)
LiME-TM: Inference Only (Arduino)
Thanks
