STMicroelectronics/st_mnistv1

ST MNIST v1

Use case : Image classification

Model description

This folder contains a custom model ST-MNIST for MNIST type datasets. ST-MNIST model is a depthwise separable convolutional based model architecture and can be used for different MNIST use-cases, e.g. alphabet recognition, digit recognition, or fashion MNIST etc.

ST-MNIST model accepts an input shape of 28 x 28, which is standard for MNIST type datasets. The pretrained model is also quantized in int8 using tensorflow lite converter.

Network information

Network Information	Value
Framework	TensorFlow Lite
Quantization	int8

Network inputs / outputs

For an image resolution of 28x28 and 36 classes : 10 integers (from 0-9) and 26 alphabets (upper-case A-Z)

Input Shape	Description
(1, 28, 28, 1)	Single 28x28 grey-scale image with UINT8 values between 0 and 255

Output Shape	Description
(1, 36)	Per-class confidence for 36 classes in FLOAT32

Recommended Platforms

Platform	Supported	Recommended
STM32L0	[]	[]
STM32L4	[x]	[x]
STM32U5	[x]	[x]
STM32H7	[x]	[x]
STM32MP1	[x]	[]
STM32MP2	[x]	[]
STM32N6	[x]	[]

Performances

Metrics

• Measures are done with default STM32Cube.AI configuration with enabled input / output allocated option.
• tfs stands for "training from scratch", meaning that the model weights were randomly initialized before training.

Reference MCU memory footprint based on EMNIST-Byclass dataset (see Accuracy for details on dataset)

Model	Format	Resolution	Series	Activation RAM	Runtime RAM	Weights Flash	Code Flash	Total RAM	Total Flash	STEdgeAI Core version
ST MNIST Byclass v1 tfs	Int8	28x28x1	STM32H7	17.21 KiB	0.3 KiB	10.08 KiB	27.99 KiB	17.51 KiB	38.07 KiB	3.0.0

Reference MCU inference time based on EMNIST-Byclass dataset (see Accuracy for details on dataset)

Model	Format	Resolution	Board	Frequency	Inference time (ms)	STEdgeAI Core version
ST MNIST Byclass v1 tfs	Int8	28x28x1	STM32H747I-DISCO	400 MHz	3.48 ms	3.0.0

Reference MPU inference time based on EMNIST-Byclass dataset (see Accuracy for details on dataset)

Model	Format	Resolution	Quantization	Board	Execution Engine	Frequency	Inference time (ms)	%NPU	%GPU	%CPU	X-LINUX-AI version	Framework
ST MNIST Byclass v1 tfs	Int8	28x28x1	per-channel**	STM32MP257F-DK2	2 CPU	1500 MHz	0.87	64.23	35.77	0	v6.1.0	TensorFlowLite 2.18.0
ST MNIST Byclass v1 tfs	Int8	28x28x1	per-channel	STM32MP157F-DK2	2 CPU	800 MHz	0.70	NA	NA	100	v6.1.0	TensorFlowLite 2.18.0
ST MNIST Byclass v1 tfs	Int8	28x28x1	per-channel	STM32MP135F-DK2	1 CPU	1000 MHz	1.02	NA	NA	100	v6.1.0	TensorFlowLite 2.18.0

** To get the most out of MP25 NPU hardware acceleration, please use per-tensor quantization

** Note: On STM32MP2 devices, per-channel quantized models are internally converted to per-tensor quantization by the compiler using an entropy-based method. This may introduce a slight loss in accuracy compared to the original per-channel models.

Accuracy with EMNIST-Byclass dataset

Dataset details: link , by_class, digits from [0-9] and capital letters [A-Z]. Number of classes: 36, Number of train images: 533,993, Number of test images: 89,264.

Model	Format	Resolution	Top 1 Accuracy
ST MNIST Byclass v1 tfs	Float	28x28x1	91.89 %
ST MNIST Byclass v1 tfs	Int8	28x28x1	91.47 %

Following we provide the confusion matrix for the model with Float32 weights.

Following we provide the confusion matrix for the quantized model with INT8 weights.

Retraining and Integration in a simple example:

Please refer to the stm32ai-modelzoo-services GitHub here

References

[1] "EMNIST : NIST Special Dataset," [Online]. Available: https://www.nist.gov/itl/products-and-services/emnist-dataset.

[2] "EMNIST: an extension of MNIST to handwritten letters". https://arxiv.org/abs/1702.05373