13.2. Model Zoo Reference

Complete reference of all available models in Tiny ML Tensorlab’s model zoo.

13.2.1. Classification Models

13.2.1.1. Standard Classification

For non-NPU devices:

Model Name	Parameters	Description
`CLS_100`	~100	Minimal, very simple tasks
`CLS_500`	~500	Small, simple classification
`CLS_1k`	~1,000	Baseline classification
`CLS_2k`	~2,000	Medium complexity
`CLS_4k`	~4,000	Good accuracy
`CLS_6k`	~6,000	Higher accuracy
`CLS_13k`	~13,000	Complex classification

13.2.1.2. NPU Classification

For NPU-enabled devices (F28P55, AM13E2, MSPM0G5187):

Model Name	Parameters	Description
`CLS_100_NPU`	~100	Minimal NPU model
`CLS_500_NPU`	~500	Small NPU model
`CLS_1k_NPU`	~1,000	Baseline NPU model
`CLS_2k_NPU`	~2,000	Medium NPU model
`CLS_4k_NPU`	~4,000	Recommended NPU model
`CLS_6k_NPU`	~6,000	Larger NPU model
`CLS_13k_NPU`	~13,000	Large NPU model
`CLS_20k_NPU`	~20,000	Very large NPU model
`CLS_55k_NPU`	~55,000	Maximum NPU model

13.2.1.3. Application-Specific Classification

Arc Fault Models:

Model Name	Parameters	Description
`ArcFault_model_200_t`	~200	Minimal arc fault detection
`ArcFault_model_400_t`	~400	Balanced arc fault
`ArcFault_model_800_t`	~800	Higher accuracy arc fault
`ArcFault_model_1400_t`	~1,400	Maximum arc fault accuracy

Motor Fault Models:

Model Name	Parameters	Description
`MotorFault_model_1_t`	~1,000	Baseline motor fault
`MotorFault_model_2_t`	~2,000	Improved motor fault
`MotorFault_model_3_t`	~4,000	Best motor fault accuracy

13.2.2. Regression Models

13.2.2.1. Standard Regression

Model Name	Parameters	Description
`REGR_500`	~500	Small regression
`REGR_1k`	~1,000	Baseline regression
`REGR_2k`	~2,000	Medium regression
`REGR_4k`	~4,000	Good accuracy
`REGR_8k`	~8,000	Higher accuracy

13.2.2.2. NPU Regression

Model Name	Parameters	Description
`REGR_500_NPU`	~500	Small NPU regression
`REGR_1k_NPU`	~1,000	Baseline NPU regression
`REGR_2k_NPU`	~2,000	Medium NPU regression
`REGR_4k_NPU`	~4,000	Recommended NPU regression
`REGR_8k_NPU`	~8,000	Large NPU regression
`REGR_20k_NPU`	~20,000	Maximum NPU regression

13.2.3. Anomaly Detection Models

Autoencoder-based models for anomaly detection.

13.2.3.1. Standard AD

Model Name	Parameters	Description
`AD_500`	~500	Small autoencoder
`AD_1k`	~1,000	Baseline autoencoder
`AD_2k`	~2,000	Medium autoencoder
`AD_4k`	~4,000	Good complexity
`AD_8k`	~8,000	Higher complexity

13.2.3.2. NPU AD

Model Name	Parameters	Description
`AD_500_NPU`	~500	Small NPU autoencoder
`AD_1k_NPU`	~1,000	Baseline NPU autoencoder
`AD_2k_NPU`	~2,000	Medium NPU autoencoder
`AD_4k_NPU`	~4,000	Recommended NPU autoencoder
`AD_8k_NPU`	~8,000	Large NPU autoencoder
`AD_20k_NPU`	~20,000	Maximum NPU autoencoder

13.2.4. Forecasting Models

13.2.4.1. Standard Forecasting

Model Name	Parameters	Description
`FCST_500`	~500	Small forecasting
`FCST_1k`	~1,000	Baseline forecasting
`FCST_2k`	~2,000	Medium forecasting
`FCST_4k`	~4,000	Good accuracy
`FCST_8k`	~8,000	Higher accuracy

13.2.4.2. NPU Forecasting

Model Name	Parameters	Description
`FCST_500_NPU`	~500	Small NPU forecasting
`FCST_1k_NPU`	~1,000	Baseline NPU forecasting
`FCST_2k_NPU`	~2,000	Medium NPU forecasting
`FCST_4k_NPU`	~4,000	Recommended NPU forecasting
`FCST_8k_NPU`	~8,000	Large NPU forecasting
`FCST_20k_NPU`	~20,000	Maximum NPU forecasting

13.2.5. Image Classification Models

Model Name	Parameters	Description
`MobileNetV2_Tiny`	~10,000	Minimal image model
`MobileNetV2_Small`	~50,000	Small image model
`MobileNetV2_Medium`	~100,000	Medium image model
`CustomCNN_Small`	~20,000	Simple custom CNN

13.2.6. Model Selection Guide

By Device Type:

Device Class	Recommended Models	Max Size
Entry-level M0+	`_100`, `_500`	~1k params
Mid-range	`_1k`, `_2k`	~4k params
High-performance	`_4k`, `_6k`	~13k params
NPU devices	`*_NPU` variants	~55k params

By Task Complexity:

Task Complexity	Classes/Output	Recommended Size
Simple (2 classes)	Binary	100-500 params
Moderate (3-5 classes)	Few classes	1k-2k params
Complex (6+ classes)	Many classes	4k+ params
Very complex	High accuracy needed	6k-13k params

13.2.7. Model Architecture Details

Classification Architecture (CLS_*k_NPU):

Input (1, 512, 1)
├── Conv1: 1→4 ch, kernel=5
├── BN + ReLU
├── MaxPool: 2x1
├── Conv2: 4→8 ch, kernel=5
├── BN + ReLU
├── MaxPool: 2x1
├── Conv3: 8→16 ch, kernel=3
├── BN + ReLU
├── Flatten
└── FC: → num_classes

Anomaly Detection Architecture (AD_*k_NPU):

Encoder:
├── Conv1: 1→4 ch
├── Conv2: 4→8 ch
└── Conv3: 8→bottleneck

Decoder:
├── ConvT1: bottleneck→8 ch
├── ConvT2: 8→4 ch
└── ConvT3: 4→1 ch

13.2.8. Using Models

In Configuration:

training:
  model_name: 'CLS_4k_NPU'

Listing Available Models:

from tinyml_tinyverse.common.models import MODEL_REGISTRY
print(list(MODEL_REGISTRY.keys()))

Model Information:

from tinyml_tinyverse.common.models import MODEL_REGISTRY

model_class = MODEL_REGISTRY['CLS_4k_NPU']
model = model_class(config={}, input_features=512, variables=1, num_classes=2)
params = sum(p.numel() for p in model.parameters())
print(f"Parameters: {params}")