mmcli

name	mmcli
description	Use when the user wants to train, compile, analyze, or deploy a TI tinyml model via the mmcli command-line interface. Covers dataset inspection, model recommendation, training, compilation, and hardware deployment to TI MCUs. Always verify MMCLI_PYTHON is set before running any command.

mmcli — TI TinyML CLI Skill

Environment setup

Before running any mmcli command, verify the Python interpreter is configured:

# Set MMCLI_PYTHON to the venv that has tinyml-modelmaker installed
export MMCLI_PYTHON=~/.venv-tinyml/bin/python

# Verify:
$MMCLI_PYTHON -c "import tinyml_modelmaker; print('OK')"

For C2000-target compilation (F28P55, F28P65, etc.) on any platform, set:

# Point to the TI C2000 CGT installation root (version 25.11.1.LTS or compatible)
export C2000_CG_ROOT=~/ti/ti-cgt-c2000_25.11.1.LTS

modelmaker reads C2000_CG_ROOT at import time; it must be set before invoking mmcli.

macOS ARM64 note: on macOS, mmcli run / mmcli compile may exit with code 245 after the pipeline completes. This is a crash in onnxsim's C extension during Python interpreter shutdown — all artifacts (mod.a, ONNX files) are written before it fires.

If tinyml-modelzoo recommendations are needed, also set:

export MMCLI_MODELZOO_PATH=~/tinyml-tensorlab/tinyml-modelzoo

---

Core workflow

1. mmcli analyze    — inspect dataset (size, classes, sample count)
2. mmcli recommend  — pick model + FE preset from modelzoo examples
3. mmcli info       — browse supported devices, tasks, models, FE presets
4. mmcli init       — create a project from a bundled example dataset
5. mmcli train      — train (ONNX export only)
6. mmcli compile    — compile an existing ONNX → TVM artifacts
7. mmcli run        — train + compile in one pass
8. mmcli deploy     — create CCS project, build, and flash to device

---

1. Dataset analysis

mmcli analyze -i <project_dir>
# <project_dir>/dataset/ must contain classes/ or files/ layout

# Output: row counts, class balance bar chart, dataset size bucket
# (tiny / small / medium / large) — use bucket in recommend step

---

2. Model recommendation

mmcli recommend -t <task_type> -d <device> [options]

# Examples:
mmcli recommend -t motor_fault -d F28P55 --variables 3 --dataset-size-bucket small
mmcli recommend -t arc_fault   -d F28P55 --top 10
mmcli recommend -t generic_timeseries_forecasting -d MSPM0G5187

# Key options:
#   --variables N          sensor channel count (+2 score bonus for exact match)
#   --dataset-size-bucket  tiny | small | medium | large
#   --top N                how many ranked matches to show (default 5)
#   --modelzoo-path DIR    explicit path to modelzoo repo root

Scoring: task(+1) device(+1) module(+1) variables-exact(+2) size(+1) = max 6. Use the recommended model name with mmcli train -n <model>.

---

3. Device / task / preset browser

mmcli info -m timeseries -t motor_fault -d F28P55
mmcli info -m audio
mmcli info -m timeseries -t generic_timeseries_classification

---

4. Create project from bundled dataset

mmcli init --list                                  # show available datasets
mmcli init --dataset motor_fault -t motor_fault -p /path/to/project
mmcli init --dataset generic_audio_classification -t audio_classification -p /path/to/project

Bundled datasets: generic_timeseries_classification, generic_timeseries_regression, generic_timeseries_anomalydetection, generic_timeseries_forecasting, arc_fault_classification, ecg_classification, fan_blade_fault, pir_detection, mnist_image_classification, generic_audio_classification.

---

5–7. Train / compile / run

# Train (exports ONNX, no compilation)
mmcli train -m timeseries -t motor_fault -d F28P55 -n CLS_1k_NPU -i /path/to/project

# Compile only (from existing ONNX)
mmcli compile -m timeseries -t motor_fault -d F28P55 -n CLS_1k_NPU \
  --onnx /path/to/model.onnx

# Full pipeline
mmcli run -m timeseries -t motor_fault -d F28P55 -n CLS_1k_NPU -i /path/to/project

# Key options (all three commands):
#   --quantization QUANTIZATION_TINPU | NO_QUANTIZATION
#   --auto-quantization              binary search for optimal bitwidth
#   --epochs N  --batch-size N  --lr FLOAT
#   --preset auto | default_preset | forced_soft_npu_preset | compress_npu_layer_data
#   --report                         generate live HTML training report
#   --dry-run                        print YAML config without executing

Preset auto-resolution

--preset auto picks the right compilation path:

• NPU device + standard task + QUANTIZATION_TINPU → default_preset (hardware NPU)
• NPU device + anomaly/forecasting task → forced_soft_npu_preset (soft-NPU)
• Non-NPU device or non-TI quantization → default_preset

NPU-capable devices: F28P55, F28P65, MSPM0G5187, MSPM33C34, CC2755, CC35X1, AM13E2.

---

8. Device deployment

8A. Verify SDK

mmcli deploy check-sdk -d F28P55
mmcli deploy check-sdk -d CC1312 --sdk-path ~/ti/simplelink_cc13xx_sdk_7.10.00

8B. Locate compiled artifacts

mmcli deploy artifacts -t motor_fault --run-id 20240115_143022 --model-id CLS_1k_NPU
# Add --no-quantization for float model artifacts
# Add --tinyml-base <PATH> if checkout is not at ~/tinyml-tensorlab

8C. Create CCS project

mmcli deploy create \
  -d F28P55 -t motor_fault \
  --run-id 20240115_143022 --model-id CLS_1k_NPU \
  --project-name my_motor_fault_project
# --device-type f28p55x  (auto-resolved; override if needed)
# --sdk-path / --ccs-templates-path  (override SDK auto-detection)

CCS device_type strings (auto-resolved from --device):

Device	CCS type
F28P55	f28p55x
F28P65	f28p65x
F28004	f28004x
MSPM0G3507	mspm0g3507
MSPM0G5187	mspm0g5187
AM263	am263
CC2755	cc2755
CC1312	cc1312
CC1314	cc1314
CC1352	cc1352

8D. Build

mmcli deploy build \
  --project-path /path/to/project \
  --ccs-path /opt/ti/ccs1260
# --build-type incremental  (default: full)

8E. Flash

mmcli deploy flash \
  --project-path /path/to/project \
  --ccs-path /opt/ti/ccs1260
# --ccxml TMS320F28P550SJ9_LaunchPad.ccxml  (auto-searched if omitted)

---

Supported devices

Family	Devices	SDK
c2000	F280013, F280015, F28003, F28004, F2837, F28P55, F28P65, F29H85, F29P58, F29P32	C2000Ware
mspm0	MSPM0G3507, MSPM0G3519, MSPM0G5187	MSPM0 SDK
mspm33	MSPM33C32, MSPM33C34	MSPM33 SDK
am13	AM13E2	MCU+ SDK
am26x	AM263, AM263P, AM261	MCU+ SDK
simplelink	CC1312, CC1314, CC1352, CC1354, CC2755, CC35X1	SimpleLink LP

---

Supported task types

Timeseries: generic_timeseries_classification, generic_timeseries_regression, generic_timeseries_anomalydetection, generic_timeseries_forecasting, arc_fault, ecg_classification, motor_fault, blower_imbalance, pir_detection

Vision: image_classification

Audio: audio_classification

---

Common flag reference

Flag	Commands	Notes
-m / --module	most	timeseries / vision / audio
-t / --task	most	task type string
-d / --device	most	canonical device ID
-n / --model	train/run	model name from modelzoo
-i / --project	train/run	project dir (contains dataset/)
--quantization	train/run/compile	QUANTIZATION_TINPU / NO_QUANTIZATION
--auto-quantization	train/run	binary search for best bitwidth
--preset	compile/run	auto / default_preset / forced_soft_npu_preset
--feature-extraction	train/run	FE preset name from schema
--report	train/run	generate HTML training report
--dry-run	train/compile/run	print YAML, do not execute
--verbose / -v	all	debug logging