gbahlnxp/mobilenetv1

MobileNet V1

Introduction

MobileNet V1 [1] is a family of lightweight convolutional neural networks for image classification. We use the weights from the implementation found in TensorFlow Model Garden.

Our base recipe script offers the smallest model (width multiplier 0.25, input resolution 128x128) by default.

Model Information

Information	Value
Input shape	Color image (128, 128, 3)
Input example	(Public domain image)
Output shape	Vector of probabilities shape (1, 1001). Labels can be found in labels.py.
Output example	Output tensor: [[0., 0., 0., 58, ... 0., 0., 0. ]]
FLOPS	28 MOPS
Number of parameters	468k
Source framework	Tensorflow/Keras
Target platform	MPUs, MCUs

Version and changelog

Initial release of quantized int8 model.

Tested configurations

The quantized int8 models have been tested on i.MX 8M Plus and i.MX 93 using benchmark-model (see i.MX Machine Learning User Guide), and on i.MX RT1170 and i.MX RT1050 using TensorFlow Lite Micro.

Training and evaluation

The model has been trained and evaluated on the ImageNet dataset [2]. It achieved a score of 39.5% Top-1 accuracy on the test set according to the source of the model.

The original training procedure is detailed here.

Conversion/Quantization

The model is downloaded in an archive which contains the quantized model and original (*.pb) TensorFlow model. See the source of the model for information on the quantization procedure that was used.

Use case and limitations

This model can be used for simple object recognition applications. Since its accuracy is relatively low, it should not be trusted for critical scenarios. If necessary, a more accurate model can be obtained by modifying the recipe.sh script and selecting a larger model scale and/or input resolution. A larger model will be slower, however.

Performance

Here are performance figures evaluated on i.MX 8MP and i.MX 93 using BSP LF6.1.1_1.0.0:

Model	Average latency	Platform	Accelerator	Command
Int8	6.53ms	i.MX 8M Plus	CPU (1 thread)	/usr/bin/tensorflow-lite-2.10.0/examples/benchmark_model --graph=mobilenet_v1_0.25_128_quant.tflite
Int8	2.44ms	i.MX 8M Plus	CPU (4 threads)	/usr/bin/tensorflow-lite-2.10.0/examples/benchmark_model --graph=mobilenet_v1_0.25_128_quant.tflite --num_threads=4
Int8	0.84ms	i.MX 8M Plus	NPU	/usr/bin/tensorflow-lite-2.10.0/examples/benchmark_model --graph=mobilenet_v1_0.25_128_quant.tflite --external_delegate_path=/usr/lib/libvx_delegate.so
Int8	3.63ms	i.MX 93	CPU (1 thread)	/usr/bin/tensorflow-lite-2.10.0/examples/benchmark_model --graph=mobilenet_v1_0.25_128_quant.tflite
Int8	2.41ms	i.MX 93	CPU (2 threads)	/usr/bin/tensorflow-lite-2.10.0/examples/benchmark_model --graph=mobilenet_v1_0.25_128_quant.tflite --num_threads=2
Int8	0.37ms	i.MX 93	NPU	/usr/bin/tensorflow-lite-2.10.0/examples/benchmark_model --graph=mobilenet_v1_0.25_128_quant_vela.tflite --external_delegate_path=/usr/lib/libethosu_delegate.so
Int8	47ms	i.MX RT1170	CPU	Tested with TensorFlow Lite Micro
Int8	70ms	i.MX RT1050	CPU	Tested with TensorFlow Lite Micro

Download and run

To create the TFLite model fully quantized in int8 with int8 input and float32 output, run bash recipe.sh.

The TFLite model file for i.MX 8M Plus, i.MX RT 1170 and i.MX RT1050 is mobilenet_v1_0.25_128_quant.tflite. The file for i.MX 93 is output in the model_imx93 directory.

An example of how to use the model is in example.py. The output labels are listed in labels.py.

Origin

Model implementation: https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md

[1] Howard, Andrew G., et al. "Mobilenets: Efficient convolutional neural networks for mobile vision applications." arXiv preprint arXiv:1704.04861 (2017). [2] Deng, Jia, et al. "Imagenet: A large-scale hierarchical image database." 2009 IEEE conference on computer vision and pattern recognition. Ieee, 2009.