tinyml-modelzoo-agent

TinyML ModelZoo — Add Custom Model

Adds a new neural network architecture to tinyml-modelzoo so it works end-to-end with training, quantization, and compilation.

No changes needed in tinyml-modelmaker or tinyml-tinyverse — modelzoo only.

---

Session Prerequisites

Load TINYML_MODELZOO_PATH from ~/.tinyml-agent-skills/.env and activate the venv:

source ~/.tinyml-agent-skills/.env 2>/dev/null || true
source "$TINYML_MODELZOO_PATH/.venv/bin/activate" 2>/dev/null || true

If TINYML_MODELZOO_PATH is unset: stop and tell user to run /tinyml-agent-skills:setup first.

---

Files Modified (Overview)

File	Purpose
tinyml_modelzoo/models/<task>.py	Model class definition
tinyml_modelzoo/device_info/run_info.py	Device performance info
tinyml_modelzoo/model_descriptions/<task>.py	Pipeline + GUI integration

---

Step 1: Gather Requirements

Ask the user:

1. Task type — classification / regression / anomaly detection / forecasting / image classification?
2. Model name — choose a class name following SCREAMING_SNAKE_CASE with param count suffix:
   • Classification: CNN_TS_MY_MODEL_2K or CNN_TS_MY_MODEL_2K_NPU (if NPU-targeted)
   • Regression: REG_TS_MY_MODEL_2K
   • Anomaly detection: AE_MY_MODEL_4K
   • Forecasting: FC_MY_MODEL_13K
3. Architecture approach — spec-based (recommended, uses standard layer blocks) or custom PyTorch?
4. NPU target? — if yes, enforce NPU constraints (see Step 2C)
5. Architecture details — how many layers, what types, approximate param count?

Set session variables:

TASK_TYPE=<classification|regression|anomalydetection|forecasting|image>
MODEL_CLASS_NAME=<e.g., CNN_TS_MY_MODEL_2K>
MODEL_GUI_NAME=<e.g., My_Model_2k_t>   # snake_case_t suffix for GUI name

Task → file mapping:

Task	Model file	Description file
classification	models/classification.py	model_descriptions/classification.py
regression	models/regression.py	model_descriptions/regression.py
anomalydetection	models/anomalydetection.py	model_descriptions/anomalydetection.py
forecasting	models/forecasting.py	model_descriptions/forecasting.py
image	models/image.py	model_descriptions/classification.py