pre_trained_models

Micron Deep Learning Accelerator (MDLA)

Introduction

This section gives a deeper insight with detailed scripts to run neural networks (NNs) on MDLA.

Few possible combinations as earlier discussed and covered in these examples are:

       Mode 0: single image, single model, single FPGA
       Mode 1: multiple images, single model, single FPGA
       Mode 2: multiple images, single model, multiple FPGAs
       Mode 3: multiple images, multiple models, single FPGA
       Mode 4: multiple images, multiple models, multiple FPGAs

Steps before running the example

1. Install the SDK by using the installation instruction
2. Download and extract the trained onnx models (email us a request for the download link)
3. Place the binary file (libmicrondla.so) in the parent folder (SDK/)

Running the example scripts

Before looking into all the variations, try out the simplest mode (mode 0).

$ python3 main.py --image sample_driving_image.png --model linknet --model-path micron_model_zoo/linknet.onnx --bitfile ac511.tgz

Provide bitfile path only during the first run after every system reboot or if a new bitfile is to be tested. Ignore this option for subsequent runs. Use --help option any time to see the details about the available options and to get list of supported models.

To run model with batched input use main_batch.py.

For example to run in mode 1 with multiple images and clusters:

python3 main_batch.py --model superresolution --model-path micron_model_zoo/super-resolution-10.onnx --numclus 4

For example to run in mode 1 with multiple (16) images and multiple clusters (2):

python3 main_batch.py --model inception --model-path micron_model_zoo/inception_v3.onnx --numclus 2

For example to run in mode 2 with multiple images (one per fpga):

python3 main_batch.py --model yolov3 --model-path micron_model_zoo/yolov3.onnx --numfpga N

For example to run in mode 3 with two images and two models on a single FPGA:

python3 main_batch.py --model resnet34_18 --model-path micron_model_zoo/resnet34.onnx,micron_model_zoo/resnet18.onnx

For example to run in mode 4 with two images, two models (the model-path contains the path to the two models separated by comma) each on a separate FPGA:

python3 main_batch.py --model resnet34_50 --model-path micron_model_zoo/resnet34.onnx,micron_model_zoo/resnet50.onnx --numfpga 2

List of models in the example set

ID	Category	Model	Mode	Notes
1	Categorization	inception	1, 2	The Inception models applied to multiple images on multiple clusters and 1 FPGA. The most number of clusters is 4. The model is from the ONNX model zoo.
2	Categorization	resnet34_18	3	Two models - resnet34 and resnet18 applied to two images on a single FPGA. The two models are from the ONNX model zoo.
3	Categorization	resnet34_50	4	Two models - resnet34 and resnet50 applied to two images on 2 FPGAs. The two models are from the ONNX model zoo.
4	Localization	RetinaNet	0	RetinaNet object detection model from ONNX model zoo
5	Localization	YOLO v3	0, 2	RetinaNet object detection model from ONNX model zoo
6	Segmentation	Linknet	0	Trained on street scenes
7	Super resolution	Super resolution	0, 1, 2	Originally from ONNX model zoo