Skip to content
TESA Open Knowledge

The edge_ai module: list the models, select one, read its answer

Module 1 — Getting started: run the real thing, then take it apart · Slides: slides.md · Module overview · Course page

Use the four core edge_ai calls — models, select, result and stop — read the model’s answer as a dict, interpret conf, CONF_FLOOR and latency correctly, and run the six-model menu for the first time on the emulator or the board.

Objectives

Section titled “Objectives”

By the end of this lesson, you will:

  1. Call edge_ai.models() and read each model’s index, name, sensor and labels keys correctly, and explain why you ask the firmware’s registry instead of hard-coding names or numbers.
  2. Explain that edge_ai.select(n) confirms the switch by observation, and wrap the call in try/except OSError.
  3. Read the dict from edge_ai.result(), say what label, top, conf, scores, latency_ms and seq mean, decide which results to trust against CONF_FLOOR (0.50), and convert latency to runs per second.
  4. Run the edge AI menu on the BENTO Emulator or the board, choose a model, press Load and see the winning class change with motion or sound.

Before you start

Section titled “Before you start”

Review from lesson 1.1 that MicroPython code lives on the Cortex-M33, while models live on the Cortex-M55 with the NPU. Keep BENTO IDE open, and have a USB cable ready if you have a board — the whole lesson also works on the BENTO Emulator if you don’t.

Concepts

Section titled “Concepts”

The edge_ai module is the one window we have into the inference engine on the M55. This lesson uses just four calls — models → select → result → stop — which is the whole story of running a model: read the registry, select a model, read its answer, then stop.

edge_ai.models() returns a list of one dict per model, with the keys index (used when selecting), name (shown on screen), sensor (0 = IMU, 1 = RADAR, 2 = MIC), and labels (the classes the model can answer). The good habit is: ask the hardware first — never guess. If the firmware gains or loses models, code that reads from models() adapts on its own, which matters a lot, since the board and the emulator don’t have the same number of models.

edge_ai.select(n) is a cross-core call. It sends the command, then keeps checking whether the engine has actually switched (confirm by observation). If it isn’t confirmed within a set time, it throws OSError, so we always wrap it in try/except OSError. start(n) behaves the same way as select(n).

edge_ai.result() returns the latest result as a dict, or None if there’s no result yet: label is the winning class, top is that class’s index, conf is confidence from 0 to 1, scores are the scores for every class, adding up to 1, latency_ms is how long the NPU took, and seq increases every time there’s a new result. In mathematical terms, the winning class is $\arg\max_k s_k$, and the confidence is $\max_k s_k$. The value CONF_FLOOR = 0.50 is the threshold the firmware recommends — below this, the model isn’t considered sure yet, because a model answers with a probability, not an absolute truth. Latency can also be converted to runs per second with $\text{fps} = 1000 / t_{\text{ms}}$ — for example, 4.1 ms ≈ 244 runs per second.

Worked example

Section titled “Worked example”

Open 12_edge_ai_menu.py and see first how the real Edge AI page uses these four calls. Predict before you run it: what happens if you select Motion and shake the board? Then run it and compare. The file we’ll fill in ourselves is in lesson 1.3.

File What this file teaches
examples/12_edge_ai_menu.py Edge AI Menu: select and run any AI model in one firmware image (no network needed)

This lesson’s slides also reference a file in another lesson, and one under shared/:

Check your understanding

Section titled “Check your understanding”

The same questions are in quiz.yaml for automated checking.

  1. Why does the code in this course read model names from edge_ai.models() instead of hard-coding a name or model number? (single choice · objective 1)

    • a) Because models() is faster than writing a list yourself
    • b) Because different boards’ firmware has different numbers of models — reading from the registry lets the code adapt automatically, with no changes needed
    • c) Because select() only accepts a model name
    • d) Because the emulator has no models at all
    Solution

    b — the habit of “ask the hardware first, never guess” means the same code works on both a six-model board and a five-model emulator. A different order doesn’t break anything either.

  2. If edge_ai.select(n) sends the command but the engine doesn’t switch within the set time, what happens? (single choice · objective 2)

    • a) select() quietly returns None and the program continues
    • b) select() throws OSError, so we must wrap it in try/except OSError
    • c) The board reboots itself
    • d) The previous model is removed from the registry
    Solution

    b — select() confirms by observing that the active model genuinely changed. If it doesn’t see that, it throws OSError, so the program never lies about what’s running.

  3. The result is {‘label’: ‘circle’, ‘conf’: 0.41, …} and CONF_FLOOR = 0.50. What should you do with this answer? (single choice · objective 3)

    • a) Trust it right away, since circle won
    • b) Treat it as not sure yet, since conf is below CONF_FLOOR — don’t act on this answer yet
    • c) Call select() again every time conf is low
    • d) Multiply conf by 2 to pass the threshold
    Solution

    b — argmax always picks a winner, even when scores are close. CONF_FLOOR is the cutoff on the conf value: only trust the answer when conf ≥ 0.50.

  4. latency_ms = 5 means the model can run at most roughly how many times per second? (single choice · objective 3)

    • a) 5 times
    • b) 50 times
    • c) 200 times
    • d) 5000 times
    Solution

    c — fps = 1000 / t_ms = 1000 / 5 = 200 runs per second, which is plenty compared to a loop that reads results roughly every 180 ms.

  5. On the BENTO Emulator, you select Motion Detection, press Load, then drag to tilt the board or press Shake. What should you see? (single choice · objective 4)

    • a) The winning class switches between idle / circle / shaking, along with a confidence bar for every class
    • b) The model list disappears from the dropdown
    • c) The screen shows Push every time
    • d) Nothing changes until you connect WiFi
    Solution

    a — Motion has three classes: idle, circle, shaking. On the emulator, scores are simulated from a simulated sensor, so they change with the tilting or shaking we simulate, with no network needed at all.

Lab

Section titled “Lab”
  • In the REPL or in a file, type edge_ai.models() and note down how many models your board or emulator has, their names, and which sensor each uses.
  • Run the menu on the emulator (or the board), select Motion Detection, press Load, and watch the winning class change to idle / circle / shaking.
  • If you have a board, try switching to at least one microphone model, and make a sound to change the class.

Going further

Section titled “Going further”

In lesson 1.3, we’ll walk through the s01_first_inference.py file and fill in four gaps ourselves until the menu works end to end.

Next lesson: lesson 1.3 — Hands-on: our first model menu

Reflect

Section titled “Reflect”
  • If CONF_FLOOR is set too low or too high, what kind of mistake would your app make?
  • Why does select() have to “wait and see” that the switch genuinely happened, instead of trusting that the command succeeded?

Review questions

Answer on your own first, then open the answer.

  1. Why does this course read model names from edge_ai.models() instead of hard-coding names or numbers? (Objective 1)

    1. เพราะ models() เร็วกว่าการเขียน list เอง
    2. เพราะเฟิร์มแวร์แต่ละบอร์ดมีโมเดลไม่เท่ากัน ถ้าอ่านจากทะเบียน โค้ดจะปรับตามเองโดยไม่ต้องแก้
    3. เพราะ select() รับได้เฉพาะชื่อโมเดล
    4. เพราะ Emulator ไม่มีโมเดลเลย
    Show answer

    Answer: B. เพราะเฟิร์มแวร์แต่ละบอร์ดมีโมเดลไม่เท่ากัน ถ้าอ่านจากทะเบียน โค้ดจะปรับตามเองโดยไม่ต้องแก้

    นิสัย "ถามฮาร์ดแวร์ก่อน อย่าเดา" ทำให้โค้ดชุดเดียวใช้ได้ทั้งบอร์ดหกโมเดลและ Emulator ห้าโมเดล ลำดับที่ต่างกันก็ไม่ทำให้พัง

  2. If edge_ai.select(n) sends the command but the engine does not switch in time, what happens? (Objective 2)

    1. select() คืน None เงียบ ๆ แล้วโปรแกรมทำต่อ
    2. select() โยน OSError เราจึงต้องห่อด้วย try/except OSError
    3. บอร์ดรีบูตเอง
    4. โมเดลตัวเดิมถูกลบออกจากทะเบียน
    Show answer

    Answer: B. select() โยน OSError เราจึงต้องห่อด้วย try/except OSError

    select() ยืนยันด้วยการสังเกตว่าโมเดลที่ active เปลี่ยนจริง ถ้าไม่เห็นก็โยน OSError เพื่อไม่ให้โปรแกรมโกหกว่ากำลังรัน

  3. The result is {'label': 'circle', 'conf': 0.41, ...} and CONF_FLOOR = 0.50. What should you do with this answer? (Objective 3)

    1. เชื่อเลยเพราะ circle ชนะ
    2. ถือว่ายังไม่ชัวร์ เพราะ conf ต่ำกว่า CONF_FLOOR ยังไม่ควรสั่งการตามคำตอบนี้
    3. เรียก select() ใหม่ทุกครั้งที่ conf ต่ำ
    4. คูณ conf ด้วย 2 ให้ผ่านเกณฑ์
    Show answer

    Answer: B. ถือว่ายังไม่ชัวร์ เพราะ conf ต่ำกว่า CONF_FLOOR ยังไม่ควรสั่งการตามคำตอบนี้

    argmax เลือกผู้ชนะเสมอแม้คะแนนสูสี CONF_FLOOR คือเกณฑ์ตัดบนค่า conf เชื่อคำตอบก็ต่อเมื่อ conf ≥ 0.50

  4. latency_ms = 5 means the model can run at most about how many times per second? (Objective 3)

    1. 5 ครั้ง
    2. 50 ครั้ง
    3. 200 ครั้ง
    4. 5000 ครั้ง
    Show answer

    Answer: C. 200 ครั้ง

    fps = 1000 / t_ms = 1000 / 5 = 200 ครั้งต่อวินาที เหลือเฟือเมื่อเทียบกับลูปที่อ่านผลทุกราว 180 ms

  5. On the BENTO Emulator you choose Motion Detection, press Load and use the tilt pad or Shake. What should you see? (Objective 4)

    1. คลาสที่ชนะเปลี่ยนระหว่าง idle / circle / shaking พร้อมแถบความมั่นใจของทุกคลาส
    2. รายชื่อโมเดลหายไปจาก dropdown
    3. หน้าจอขึ้น Push ทุกครั้ง
    4. ไม่มีอะไรเปลี่ยนจนกว่าจะต่อ WiFi
    Show answer

    Answer: A. คลาสที่ชนะเปลี่ยนระหว่าง idle / circle / shaking พร้อมแถบความมั่นใจของทุกคลาส

    Motion มีสามคลาส idle circle shaking บน Emulator คะแนนเลียนแบบจากเซนเซอร์จำลอง จึงเปลี่ยนตามการเอียงหรือเขย่าที่เราจำลอง ไม่ต้องใช้เน็ตเลย

Cite this lesson

If you teach from this lesson or reuse it in slides or documents, credit it with the text below. If you changed it, add (adapted) after the title.

"The edge_ai module: list the models, select one, read its answer" from TESA Open Knowledge by the Thai Embedded Systems Association (TESA), https://github.com/tesaiot/tesa-qualification-program, licensed under CC BY-NC 4.0

Thai attribution: "โมดูล edge_ai: ถามทะเบียนโมเดล เลือก แล้วอ่านคำตอบ" จาก TESA Open Knowledge โดยสมาคมสมองกลฝังตัวไทย (Thai Embedded Systems Association: TESA) https://github.com/tesaiot/tesa-qualification-program สัญญาอนุญาต CC BY-NC 4.0

Lesson link: https://tesaiot.github.io/tesa-qualification-program/en/courses/edge-ai-developer/m01-onboarding/l02-edge-ai-module/

Full guide: how to cite TESA

TESA Open Knowledge · © 2026 สมาคมสมองกลฝังตัวไทย (TESA) · CC BY-NC 4.0

Content is licensed CC BY-NC 4.0. Reuse it non-commercially and credit the Thai Embedded Systems Association (TESA) every time. · How to cite TESA