Convert

You can automatically convert the AI model's framework to the target framework.

Only single-input models are supported.
The supported channel ordering format is 'Channel First'. The three-dimensional array structure of images should be organized in the order of Batch, Number of Channels, Height, and Width.

Batch size: The number of combined input datasets that the model processes simultaneously.

Channel: enter 3 for RGB channel and 1 for gray channel.

Input size: In computer vision tasks, input size refers to the size of the input images.

ONNX to TensorRT

Target Device	SW version	Input datatype	Batch size	Channel	Input size	Output datatype
NVIDIA Jetson Nano	JetPack 4.4.1, JetPack 4.6	FP32	1~4 (Static, Dynamic)	1~4	height, width	FP16
NVIDIA Jetson Xavier NX	JetPack 5.0.2, JetPack 4.6	FP32	1~4 (Static, Dynamic)	1~4	height, width	FP16
NVIDIA Jetson TX2	JetPack 4.6	FP32	1~4 (Static, Dynamic)	1~4	height, width	FP16
NVIDIA Jetson AGX Xavier	JetPack 4.6	FP32	1~4 (Static, Dynamic)	1~4	height, width	FP16
NVIDIA Jetson AGX Orin	JetPack 5.0.1	FP32	1~4 (Static, Dynamic)	1~4	height, width	FP16
NVIDIA T4	-	FP32	1~4 (Static, Dynamic)	1~4	height, width	FP16

ONNX to OpenVino

Input datatype	Batch size	Channel	Input size	Output datatype
FP32	1~4 (Static, Dynamic)	1~4	height, width	FP16

ONNX to TFlite

Input datatype	Batch size	Channel	Input size	Output datatype
FP32	1~4 (Static, Dynamic)	1~4	height, width	FP16 INT8

TensorFlow to TensorFlowLite

Input datatype	Batch size	Channel	Input size	Output datatype
FP32	1~4 (Static, Dynamic)	1~4	height, width	FP16 INT8

Benchmark

Measure the performance of the AI model on the target device. You can measure the actual performance on the devices listed below without the need to purchase or install the device.

Below is a table for benchmarking compatible model frameworks.

Arm MCU

Target Device	.tflite
Renesas RA8D1 (Arm Cortex-M85)	O (only INT8)
Renesas RA8D1 (Arm Cortex-M85) with helium	O (only INT8)
Alif Ensemble DevKit Gen2 (Arm Cortex-M85 + Ethos-U55)	O (only INT8)
Alif Ensemble DevKit Gen2 (Arm Cortex-M85 + Ethos-U55) with helium	O (only INT8)

NVIDIA

When benchmarking on Jetson, it is essential for the model file and target device to match the Jetpack version.

Target Device	.tflite	.engine	.tflite	.onnx
Jetson Nano JetPack 4.4.1	O	O	O	O
Jetson Nano JetPack 4.6	O	O	O	O
Jetson Xavier NX JetPack 4.6	O	O	O	O
Jetson Xavier NX JetPack 5.0.2	O	O	O	O
Jetson TX2 JetPack 4.6	O	O	O	O
Jetson AGX Xavier JetPack 4.6	O	O	O	O
Jetson AGX Orin JetPack 5.0.1	O	O	O	O
AWS-T4	O	O	O	O

Raspberry Pi

Target Device	.tflite	.onnx
Raspberry Pi ZeroW	O	O
Raspberry Pi Zero2W	O	O
Raspberry Pi 2B	O	O
Raspberry Pi 3B	O	O
Raspberry Pi 3B+	O	O
Raspberry Pi 4B	O	O

Intel

Target Device	.zip(bin+xml)
Xeon W-2223	O