In edge AI applications, designing the neural network properly is very important. Customers usually use example models from ai8x or create their own models and train them on their data, the training results are usually unsatisfactory. This happens because the MAX78002 uses low-bit models (like INT8, INT4, or INT2), where each value stores very little information, making training more difficult and less stable compared to normal FP32 models. Because of this, the model design must be very careful. To solve this problem, adopting AutoML to search for an optimal model architecture can be an effective approach.
The ai8x tool provides an AutoML method called Once-For-All (OFA), The goal of OFA is to train a model once and automatically obtain multiple small subnets with compatible performance. It consists of three steps:
Supernet Training:
A supernet is a huge FP32-based CNN with deepaer layer,more channel and larger kernel sizes, which cannot be directly deployed on an MCU. In OFA, a sequential supernet is trained using knowledge distillation (KD) and shared weights to ensure robustness across multiple subnets. Its goal is not to iterate a subnet with the highest accuracy because searching for an optimal solution is time-consuming. During KD, subnets sampled from the supernet reuse its weights, enabling efficient training and maintaining performance consistency.
Network Architecture Searching (NAS):
Once the supernet is trained, NAS is applied to identify an optimal smaller subnet (TinyDurianNet). OFA adopts an evolutionary search strategy, which is more efficient than traditional NAS methods in terms of time and computational resources. The NAS procedure outputs multiple candidate subnets that offer similar accuracy but differ in architecture. Depending on the target hardware’s performance characteristics and power-consumption constraints, a suitable subnet can be selected as the final TinyDurianNet.
Quantization Aware Training (QAT)
Quantization-aware training (QAT) enables neural networks to operate efficiently on the MAX78002’s CNN accelerator, which supports INT1, INT2, INT4, and INT8 weights.
The core idea of QAT is to simulate low bit quantization errors during training (fake quantization) so the model learns to tolerate quantization noise in both forward and backward passes, preserving accuracy when deployed on low-precision hardware.
For ai8x’s QAT, the straight-through estimator (STE) rule is applied in back propagation. Figure 6 demonstrates the process of QAT. After QAT, INT8-based TinyDurianNet can be obtained and deployed on the MAX78002 directly.
Applicable Part Numbers
| DigiKey Part Numbers | Manufacturer Part Number |
|---|---|
| MAX9867ETJ+TTR-ND,MAX9867ETJ+TCT-ND,MAX9867ETJ+TDKR-ND | MAX9867ETJ+T |
| MAX9867EWV+TTR-ND,MAX9867EWV+TCT-ND | MAX9867EWV+T |
| MAX9867ETJ±ND | MAX9867ETJ+ |
| 505-MAX9867ETJ+G3U-ND | MAX9867ETJ+G3U |
| MAX9867ETJ+TG3U-ND | MAX9867ETJ+TG3U |
| 505-MAX9867EWV+G47-ND | MAX9867EWV+G47 |
| 505-MAX78002GXE±ND | MAX78002GXE+ |
| 3318-PG1302XA-NA0-ND | PG1302XA-NA0 |
| 3318-PG2405CA-NA0-ND | PG2405CA-NA0 |
| 757-NHD-2.4-240320AF-CSXP-ND | NHD-2.4-240320AF-CSXP |
| NHD-2.4-240320CF-CSXN#-F-ND | NHD-2.4-240320CF-CSXN#-F |
| 1548-NHD-2.4-240320CF-BSXV-F-ND | NHD-2.4-240320CF-BSXV-F |
| 757-NHD-2.4-240320CF-BSXV-FT-ND | NHD-2.4-240320CF-BSXV-FT |
| NHD-2.8-240320AF-CSXP-F-ND | NHD-2.8-240320AF-CSXP-F |
| 1597-104990802-ND | 1.05E+08 |
| 757-NHD-2.4-240320AF-CTXP-ND | NHD-2.4-240320AF-CTXP |
| 5491-AFY240320B0-2.4INTH-ND | AFY240320B0-2.4INTH |
| 3387-ATM0200B3-ND | ATM0200B3 |
| NHD-2.4-240320CF-CSXN#-FT-ND | NHD-2.4-240320CF-CSXN#-FT |
| 5491-AFY240320A1-2.8INTH-ND | AFY240320A1-2.8INTH |
| 757-NHD-2.4-240320AF-CTXP-T-ND | NHD-2.4-240320AF-CTXP-T |
| 757-NHD-2.4-240320AF-CSXP-T-ND | NHD-2.4-240320AF-CSXP-T |
| 4863-WF0200BSYAJDNN0-ND | WF0200BSYAJDNN0 |
| 3390-GLT035240320WH1-ND | GLT035240320WH1 |
| 3086-MDT0280A1IH-MULTI-ND | MDT0280A1IH-MULTI |
| 1528-4421-ND | 4421 |
| 1597-104990803-ND | 1.05E+08 |
| 5491-AFL240320A0-2.0INTM-ANO-ND | AFL240320A0-2.0INTM-ANO |
| 5491-AFL240320A0-2.4N12NTM-ANO-ND | AFL240320A0-2.4N12NTM-ANO |
| 5491-AFL240320A0-2.8N12NTM-ANO-ND | AFL240320A0-2.8N12NTM-ANO |
| 757-NHD-2.4-240320AF-CSXP-CTP-ND | NHD-2.4-240320AF-CSXP-CTP |
| 757-NHD-2.8-CSXP-BREAKOUT-ND | NHD-2.8-CSXP-BREAKOUT |
| 5491-AFR240240A0-1.5INTN-ND | AFR240240A0-1.5INTN |
| 5491-AFL240240A0-1.3INTM-ANO-ND | AFL240240A0-1.3INTM-ANO |
| 5491-AFR240320A0-2.0INTM-ND | AFR240320A0-2.0INTM |
| 5491-AFR240240A0R-1.3INTM-ND | AFR240240A0R-1.3INTM |
| 757-NHD-2.4-240320AF-CTXP-CTP-ND | NHD-2.4-240320AF-CTXP-CTP |
| 5491-AFY240320A1-2.4INTH-C1-ND | AFY240320A1-2.4INTH-C1 |
| 5491-AFY240320A1-2.8INTH-C1-ND | AFY240320A1-2.8INTH-C1 |
| NHD-2.8-240320AF-CSXP-FT-ND | NHD-2.8-240320AF-CSXP-FT |
| 277-1132710-ND | 1132710 |
| 3387-ATM0240B32-ND | ATM0240B32 |
| 5491-AFY240320A1-2.8INTH-R-ND | AFY240320A1-2.8INTH-R |
| 757-NHD-1.5-240240AF-CSXP-ND | NHD-1.5-240240AF-CSXP |
| 757-NHD-1.5-240240AF-CSXP-T-ND | NHD-1.5-240240AF-CSXP-T |
| 5491-AFY240320A2-2.4INTH-R-ND | AFY240320A2-2.4INTH-R |
| 757-NHD-2.4-CTXP-BREAKOUT-ND | NHD-2.4-CTXP-BREAKOUT |
| 757-NHD-2.4-CSXP-BREAKOUT-ND | NHD-2.4-CSXP-BREAKOUT |
| 757-NHD-2.4R-CTXP-BREAKOUT-ND | NHD-2.4R-CTXP-BREAKOUT |
| 757-NHD-2.4R-CSXP-BREAKOUT-ND | NHD-2.4R-CSXP-BREAKOUT |
| 757-NHD-2.8R-CSXP-BREAKOUT-ND | NHD-2.8R-CSXP-BREAKOUT |
| 757-NHD-2.4C-CTXP-BREAKOUT-ND | NHD-2.4C-CTXP-BREAKOUT |
| 757-NHD-2.4C-CSXP-BREAKOUT-ND | NHD-2.4C-CSXP-BREAKOUT |
| 757-NHD-2.8C-CSXP-BREAKOUT-ND | NHD-2.8C-CSXP-BREAKOUT |
| 2544-AFL240135A0-1.14INTM-ANO-ND | AFL240135A0-1.14INTM-ANO |
| 4863-WF0200BTYAJDNN0-ND | WF0200BTYAJDNN0 |
| 3390-GLT024240320TR2-ND | GLT024240320TR2 |
| 3190-DLC-T2.0ASC#1-ND | DLC-T2.0ASC#1 |
| 4863-WF24MTLAJDNN0-ND | WF24MTLAJDNN0 |
| 3749-ST0154A1W-RSLW-C-ND | ST0154A1W-RSLW-C |
| 2544-AFY240320A0-2.4INTH-R-ND | AFY240320A0-2.4INTH-R |
| 3086-MDT0200FIH-MULTI-ND | MDT0200FIH-MULTI |
| 2632-E15RA-FW255-N-ND | E15RA-FW255-N |
| 2632-E24GA-I-CW850-RTB-ND | E24GA-I-CW850-R |
| 2632-E24GA-I-CW1000-N-ND | E24GA-I-CW1000-N |
| 2632-E28RA-I-CW480-N-ND | E28RA-I-CW480-N |
| 2632-E24RA1-FW420-N-ND | E24RA1-FW420-N |
| 3190-DLC0283BCM04CB-1-ND | DLC0283BCM04CB-1 |
| 3749-ST0280A3WY-RSLW-F-ND | ST0280A3WY-RSLW-F |
| 4863-WF0200BTYAJDNG0-ND | WF0200BTYAJDNG0 |
| 2544-AFY240320A0-2.4INTH-C-ND | AFY240320A0-2.4INTH-C |
| 2632-E24RA-I-CW500-N-ND | E24RA-I-CW500-N |
| 2632-E32RA-CW380-N-ND | E32RA-CW380-N |
| 2632-E32RA-CW350-R-ND | E32RA-CW350-R |
| 3190-DLC0140AZP03SF-1-ND | DLC0140AZP03SF-1 |
| 2632-E28RA-FW480-N-ND | E28RA-FW480-N |
| 2632-E28RA-I-CW600-C-ND | E28RA-I-CW600-C |
| 3749-ST0430E2WCY-RSLW-F-ND | ST0430E2WCY-RSLW-F |
| 2632-E32RA-CW370-C-ND | E32RA-CW370-C |
| 2632-E28RA-FW580-C-ND | E28RA-FW580-C |
| 3749-ST0144S4W-RSLW-F-ND | ST0144S4W-RSLW-F |
| 3318-PG28006A-ND | PG28006A |
| 2632-E24RB-I-CW1150-N-ND | E24RB-I-CW1150-N |
| 3749-ST0280A3W-RSLW-F-ND | ST0280A3W-RSLW-F |
| 2632-E24RE-I-CW400-C-ND | E24RE-I-CW400-C |
| 5491-AFL240240A0-1.54INTM-ANO-ND | AFL240240A0-1.54INTM-ANO |
| 5491-AFR240320A0-1.77INTM-ND | AFR240320A0-1.77INTM |
| NHD-2.8-240320AF-CSXP-FCTP-ND | NHD-2.8-240320AF-CSXP-FCTP |
| 1597-104990542-ND | 1.05E+08 |
| 2470-ATM0280B3-ND | ATM0280B3 |
| 1471-MIKROE-6587-ND | MIKROE-6587 |
| 1471-MIKROE-6588-ND | MIKROE-6588 |
| 2114-4DLCD-24-240-320-RT-NM-165-M16/S-ND | 4DLCD-24-240-320-RT-NM-165-M16/S |
| 2114-4DLCD-24-240-320-NT-CLB-170-M16/S-ND | 4DLCD-24-240-320-NT-CLB-170-M16/S |
| 2114-4DLCD-28-240-320-NT-NM-175-M8/M16-ND | 4DLCD-28-240-320-NT-NM-175-M8/M16 |
| 2114-4DLCD-28-240-320-RT-NM-140-M8/M16-ND | 4DLCD-28-240-320-RT-NM-140-M8/M16 |
| 2114-4DLCD-28-240-320-RT-NM-170-M16/S-ND | 4DLCD-28-240-320-RT-NM-170-M16/S |
| 2114-4DLCD-24-320-240-CT-NM-170-M16/S-ND | 4DLCD-24-320-240-CT-NM-170-M16/S |
| 2114-4DLCD-24-240-320-CT-CLB-165-M16/S-ND | 4DLCD-24-240-320-CT-CLB-165-M16/S |
| 2114-4DLCD-28-320-240-CT-NM-170-M8/M16-ND | 4DLCD-28-320-240-CT-NM-170-M8/M16 |
| 2114-4DLCD-28-240-320-CT-CLB-157-M8/M16-ND | 4DLCD-28-240-320-CT-CLB-157-M8/M16 |
| 5491-AFY240320A1-2.8INTH-C-ND | AFY240320A1-2.8INTH-C |
| 2544-AGU240320A00-1.77INSM-ND | AGU240320A00-1.77INSM |
| 5491-AFY240320A1-2.4INTH-C-ND | AFY240320A1-2.4INTH-C |
| 5491-AFR240240A0-1.3INTM-ND | AFR240240A0-1.3INTM |
| 3749-ST0103A2W-WSNLW-F-ND | ST0103A2W-WSNLW-F |
| 2544-AFY240320A0-2.8INTH-C-ND | AFY240320A0-2.8INTH-C |
| 2114-4DLCD-25-240-240-NT-NM-300-M16/S-ND | 4DLCD-25-240-240-NT-NM-300-M16/S |
| 2114-4DLCD-25-240-240-CT-NM-255-M16/S-ND | 4DLCD-25-240-240-CT-NM-255-M16/S |
| 2114-4DLCD-25-240-240-CT-CLB-300-M16/S-ND | 4DLCD-25-240-240-CT-CLB-300-M16/S |
| 2544-AFY240320A0-2.8INTH-ND | AFY240320A0-2.8INTH |
| 2544-AFY240320A0-2.8INTH-R-ND | AFY240320A0-2.8INTH-R |
| 2632-E28RA-I-CW384-R-ND | E28RA-I-CW384-R |
| 5491-AFY240320A1-2.4INTH-R-ND | AFY240320A1-2.4INTH-R |
| 5491-AFY240320A1-2.4INTH-ND | AFY240320A1-2.4INTH |
| 3749-ST0280H4W-RSMLW-F-ND | ST0280H4W-RSMLW-F |
| 3749-ST0200H2W-RSLW-F-ND | ST0200H2W-RSLW-F |
| 4863-WF0200BSYAJDNG0-ND | WF0200BSYAJDNG0 |
| 6233-GF-M36XGF20383-01-ND | GF-M36XGF20383-01 |
| 2632-E24BA-CW1000-N-ND | E24BA-CW1000-N |