Skip to content
TESA Open Knowledge

Inside training: Keras, Conv1D, gradient descent, int8 and the confusion matrix

Module 5 — Training and deploying to several targets · Slides: slides.md · Module overview · Course page

Open train.py and take apart its four beats — build, fit, convert and eval — with the maths behind them: the Conv1D equation, softmax, cross-entropy, gradient descent, and int8 compression with scale and zero-point. Know why the representative dataset is needed, why normalization is part of the model, and learn to read a learning curve and a confusion matrix.

Objectives

Section titled “Objectives”

By the end of this lesson, you will:

  1. Compute one Conv1D output by hand with y[t] = Σ w[k]·x[t+k] + b, and count the parameters of a Conv1D layer and a Dense layer.
  2. Compute the softmax of three class scores, and interpret a learning curve as healthy, overfitting or underfitting.
  3. Explain full-integer int8 quantization with real ≈ scale × (q − zero_point), and state the job of representative_dataset and inference_input_type.
  4. Read a confusion matrix, compute the accuracy, and name the pair of classes the model confuses.

Before you start

Section titled “Before you start”

You’ve been through lesson 5.3, and have already run train.py and seen its log. Open train.py alongside the slides, and if you have time, download math_lab.html and open it in a browser (needs internet to load GeoGebra).

Concepts

Section titled “Concepts”

train.py reads as four beats: build → fit → convert → eval. The build beat lays out a 1-D CNN with tf.keras.Sequential: input (50, 6) → Conv1D(16, 5) → MaxPooling1D(2) → Conv1D(32, 3) → GlobalAveragePooling1D → Dense(32) → Dense(3, softmax). Every layer is an op the Ethos-U55 can accelerate, totalling about 3,200 parameters (the first Conv1D: 5·6·16 + 16 = 496). Conv1D is $y[t] = \sum_{k=0}^{K-1} w[k],x[t+k] + b$ — one filter sliding along time, producing a high output when that stretch resembles the filter’s shape. For example, kernel [.2 .5 .3] on .1 .4 .8 gives .46 (b = 0).

The fit beat loops through three equations every batch: softmax, $\hat{y}_i = e^{z_i}/\sum_j e^{z_j}$, turns raw scores into probabilities that add up to 1; cross-entropy, $L = -\sum_i y_i \log \hat{y}i$, measures how far a prediction is from the answer; and gradient descent, $\theta \leftarrow \theta - \eta,\nabla\theta L$, moves the weights downhill on the loss (adam is an optimizer in this family). epochs=25, batch_size=32 and validation_data let us read the learning curve: accuracy and val_accuracy rising together is healthy; train high but val low or falling is overfitting; both stuck low is underfitting.

The convert beat compresses float32 into int8 with $real \approx scale \times (q - zero_point)$, roughly four times smaller and accelerable by the NPU. The converter must calibrate the activation value range from a representative_dataset (200 real windows from the training set). If TFLITE_BUILTINS_INT8 is set but this set is never attached, convert() stops with a ValueError. inference_input_type = tf.int8 and inference_output_type = tf.int8 make the input and output pure int8. The normalize step’s mean/std are saved to .norm.npz, because normalization is part of the model — whoever uses it must use the same values. The eval beat quantizes the input with the file’s scale/zero-point, runs it, then dequantizes the output before argmax. The summary is a confusion matrix: rows are the true class, columns are the predicted class, the diagonal is correct predictions, and off-diagonal cells tell you which pair needs more data.

Worked example

Section titled “Worked example”

This lesson’s slides also reference files in another lesson and under shared/:

Check your understanding

Section titled “Check your understanding”

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

  1. kernel w = [.2, .5, .3] and b = 0, over the signal .4 .8 .6. What’s the output? (single choice · objective 1)

    • a) 0.46
    • b) 0.52
    • c) 0.66
    • d) 1.80
    Solution

    c — .2×.4 + .5×.8 + .3×.6 = .08 + .40 + .18 = .66. Multiply, then add, position by position.

  2. Raw scores z = [2.0, 3.1, 1.2] for idle, circle, shaking. After softmax, what’s circle’s probability, roughly? (single choice · objective 2)

    • a) 0.31
    • b) 0.49
    • c) 0.67
    • d) 1.00
    Solution

    c — e^3.1 ≈ 22.2, divided by e^2.0 + e^3.1 + e^1.2 ≈ 7.39 + 22.2 + 3.32 ≈ 32.9, giving about 0.67. idle is about 0.22, and shaking about 0.10.

  3. Training finishes with accuracy = 0.99, but val_accuracy = 0.60 and keeps dropping near the end. What does that mean? (single choice · objective 2)

    • a) The model is excellent
    • b) Overfitting — the model has memorized the training set but doesn’t generalize
    • c) Underfitting — the model is too small
    • d) The representative dataset is wrong
    Solution

    b — high train but val not following means it’s memorizing the exam. Fix it with more data, stopping earlier, or a smaller model. Underfitting is both staying stuck low instead.

  4. You set supported_ops = [TFLITE_BUILTINS_INT8] but forget to attach representative_dataset, then call convert(). What happens? (single choice · objective 3)

    • a) You get a file with the same accuracy
    • b) The converter stops with a ValueError, because full-integer needs samples to calibrate against
    • c) You get a float32 file
    • d) The file shrinks eight times smaller
    Solution

    b — with no samples, the converter can’t choose the activation’s scale and zero-point, so TensorFlow refuses right away. If given samples that don’t resemble reality, the file comes out but accuracy suffers.

  5. Confusion matrix: idle [10 0 0], circle [0 8 2], shaking [0 1 9]. What’s the accuracy, and which pair is most confused? (single choice · objective 4)

    • a) 0.90 accuracy, and the model most often mistakes circle for shaking
    • b) 0.90 accuracy, and the model mistakes idle most often
    • c) 0.27 accuracy, and no pair is confused
    • d) 1.00 accuracy
    Solution

    a — the diagonal, 10 + 8 + 9 = 27 out of 30, is 0.90. circle’s row has 2 windows predicted as shaking — worth collecting more of these two motions.

Lab

Section titled “Lab”
  • Compute three more Conv1D outputs for kernel [.2 .5 .3] over .1 .4 .8 .6 .2 .3, and compare against an animation.
  • Count the parameters of every layer in the model by hand, and compare against model.count_params(), which the full version prints out.
  • Compute the softmax of [2.0, 3.1, 1.2] and the loss when the answer is circle, and note it in your learning log.

Going further

Section titled “Going further”

In lesson 5.5, we’ll fill four blanks in s12_train.py to complete all four beats, then run it in Docker until we have our own model.

Next lesson: lesson 5.5 — Hands-on: fill in a training script and run it in Docker

Reflect

Section titled “Reflect”
  • If a confusion matrix on real data often confuses circle with shaking, would you fix the data or the model first, and why?
  • Why is normalizing with the wrong mean/std set a bug that’s harder to find than a program error?

Review questions

Answer on your own first, then open the answer.

  1. With kernel w = [.2, .5, .3] and b = 0 over the signal .4 .8 .6, what is the output? (Objective 1)

    1. 0.46
    2. 0.52
    3. 0.66
    4. 1.80
    Show answer

    Answer: C. 0.66

    .2×.4 + .5×.8 + .3×.6 = .08 + .40 + .18 = .66 คือการคูณแล้วบวกทีละตำแหน่ง

  2. Raw scores z = [2.0, 3.1, 1.2] for idle, circle, shaking. After softmax, roughly what probability does circle get? (Objective 2)

    1. 0.31
    2. 0.49
    3. 0.67
    4. 1.00
    Show answer

    Answer: C. 0.67

    e^3.1 ≈ 22.2 หารด้วย e^2.0 + e^3.1 + e^1.2 ≈ 7.39 + 22.2 + 3.32 ≈ 32.9 ได้ราว 0.67 ส่วน idle ราว 0.22 และ shaking ราว 0.10

  3. At the end accuracy = 0.99 but val_accuracy = 0.60 and falling. What does it mean? (Objective 2)

    1. โมเดลดีมาก
    2. overfit โมเดลจำชุดฝึกแต่ไม่ generalize
    3. underfit โมเดลเล็กเกินไป
    4. representative dataset ผิด
    Show answer

    Answer: B. overfit โมเดลจำชุดฝึกแต่ไม่ generalize

    train สูงแต่ val ไม่ตามคือจำข้อสอบ แก้ด้วยข้อมูลเพิ่ม หยุดเร็วขึ้น หรือโมเดลเล็กลง ส่วน underfit คือทั้งคู่ต่ำค้าง

  4. You set supported_ops = [TFLITE_BUILTINS_INT8] but forget representative_dataset, then call convert(). What happens? (Objective 3)

    1. ได้ไฟล์ที่แม่นเท่าเดิม
    2. converter หยุดด้วย ValueError เพราะ full-integer ต้องมีตัวอย่างไว้ calibrate
    3. ได้ไฟล์ float32
    4. ไฟล์เล็กลงแปดเท่า
    Show answer

    Answer: B. converter หยุดด้วย ValueError เพราะ full-integer ต้องมีตัวอย่างไว้ calibrate

    ไม่มีตัวอย่าง converter ก็เลือก scale และ zero-point ของ activation ไม่ได้ TensorFlow จึงปฏิเสธตั้งแต่ต้น ถ้ามีแต่เป็นตัวอย่างที่ไม่เหมือนจริง ไฟล์จะออกมาแต่แม่นตก

  5. Confusion matrix: idle [10 0 0], circle [0 8 2], shaking [0 1 9]. What is the accuracy and the most confused pair? (Objective 4)

    1. ความแม่น 0.90 และโมเดลทาย circle ผิดเป็น shaking บ่อยที่สุด
    2. ความแม่น 0.90 และโมเดลทาย idle ผิดบ่อยที่สุด
    3. ความแม่น 0.27 และไม่มีคู่ใดสับสน
    4. ความแม่น 1.00
    Show answer

    Answer: A. ความแม่น 0.90 และโมเดลทาย circle ผิดเป็น shaking บ่อยที่สุด

    แนวทแยง 10 + 8 + 9 = 27 จาก 30 คือ 0.90 แถว circle มี 2 หน้าต่างที่ถูกทายเป็น shaking ควรเก็บสองท่านี้เพิ่ม

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.

"Inside training: Keras, Conv1D, gradient descent, int8 and the confusion matrix" 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: "ข้างในการฝึก: Keras, Conv1D, gradient descent, int8 และ confusion matrix" 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/m05-training/l04-inside-training/

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