[TOC]
In this repository we'll explore how to run a state-of-the-art object detection mode, Yolov5, on the Google Coral EdgeTPU.
This project was submitted to, and won, Ultralytic's competition for edge device deployment in the EdgeTPU category. The notes for the competition are at the bottom of this file, for reference.
Probably the most interesting aspect for people stumbling across this is that this project requires very few runtime dependencies (it doesn't even need PyTorch). It contains comprehensive benchmarking code, examples of how to compile and run a custom model on the EdgeTPU and a discussion of how to test on real edge hardware.
TL;DR (see the Dockerfile):
sudo apt-get update && sudo apt-get -y upgrade
sudo apt-get install -y git curl gnupg
# Install PyCoral (you don't need to do this on a Coral Board)
echo "deb https://packages.cloud.google.com/apt coral-edgetpu-stable main" | tee /etc/apt/sources.list.d/coral-edgetpu.list
curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | apt-key add -
sudo apt-get update
sudo apt-get install -y gasket-dkms libedgetpu1-std python3-pycoral
# Get Python dependencies
sudo apt-get install -y python3 python3-pip
pip3 install --upgrade pip setuptools wheel
python3 -m pip install numpy
python3 -m pip install opencv-python-headless
python3 -m pip install tqdm pyyaml
# Clone this repository
git clone https://github.com/jveitchmichaelis/edgetpu-yolo
cd edgetpu-yolo
# Run the test script
python3 detect.py -m yolov5s-int8-224_edgetpu.tflite --bench_image
Wasn't that easy? You can swap out different models and try other images if you like. You should see an inference speed of around 25 fps with a 224x224 px input model.
Note if you're using a PCIe accelerator, you will need to install an appropriate kernel driver. See the hardware notes for more information.
- Hardware setup (hardware.md)
- Briefly covers setup for the Coral Dev Board(s)
- Covers electrical and mechanical setup for the Jetson Nano, EdgeTPU driver installation, etc.
- On-device software setup (software.md)
- Setting up virtual environments and Docker
- Installing
pycoral
and related libraries - Notes on installing PyTorch, OpenCV etc from source [for development and testing work]
- Model generation and export (export.md)
- Exporting a TFLite model from PyTorch
- Notes on the
edgetpu_compiler
As the introduction says, all you need to do is install the dependencies and then run:
python3 detect.py -m yolov5s-int8-224_edgetpu.tflite --bench_speed
python3 detect.py -m yolov5s-int8-224_edgetpu.tflite --bench_image
This should give you first a speed benchmark (on 100 images - edit the file if you want to run more) and then on the Zidane test image (you should get two detections for the 224 model).
I've also included an (untested) option to run from a video stream.
The provided code is pretty much the minimal you need to get going with the TPU. It provides a simple class for loading the model and running inference. There are also a few utilities copied from Yolov5 for image annotation, but it's very basic at this stage.
You can also use the EdgeTPUModel
class in your own software quite easily:
from edgetpumodel EdgeTPUModel
from utils import get_image_tensor
model = EdgeTPUModel("model_name", "names.yaml")
input_shape = model.get_input_shape()
full_image, net_image, pad = get_image_tensor("/path/to/image", input_shape[0])
pred = model.predict(net_image)
It's not yet ready for production(!) but you should find it easy to adapt.
If you want, you can run everything inside a Docker container. I've set it up so that you should mount this repository as an external volume (easier for experimenting/modifying files on the fly).
cd docker
docker build -t edgetpu .
docker run -it --rm --privileged -v /path/to/repo:/yolo edgetpu bash
> cd /yolo
> python3 detect.py -m yolov5s-int8-224_edgetpu.tflite --bench_speed
Performance seems to be slightly faster in Docker, perhaps due to updated versions of some libraries?
Here is the result of running three different models. All benchmarks were performed using an M.2 accelerator on a Jetson Nano 4GB. Settings are conf_thresh
of 0.25, iou_thresh
of 0.45. If you fiddle these so you get more bounding boxes, speed will decrease as NMS takes more time.
- 96x96 input, runs fully on the TPU ~60-70fps
- 192x192 input, runs mostly on the TPU ~30-35fps
- 224x224 input, runs mostly on the TPU ~25-30 fps
- >= 256 px currently fails to compile due to large tensors. It's probable that the backbone alone would compile fine and then detection can run on CPU, but this is typically extremely slow - an order of magnitude slower. Better, I think, to explore options for Yolov5 models with smaller width/depth parameters.
(py36) josh@josh-jetson:~/code/edgetpu_yolo$ python3 detect.py -m yolov5s-int8-96_edgetpu.tflite --bench_speed
INFO:EdgeTPUModel:Loaded 80 classes
INFO:__main__:Performing test run
100%|¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦| 100/100 [00:01<00:00, 58.28it/s]
INFO:__main__:Inference time (EdgeTPU): 13.40 +- 1.68 ms
INFO:__main__:NMS time (CPU): 0.43 +- 0.39 ms
INFO:__main__:Mean FPS: 72.30
(py36) josh@josh-jetson:~/code/edgetpu_yolo$ python3 detect.py -m yolov5s-int8-192_edgetpu.tflite --bench_speed
INFO:EdgeTPUModel:Loaded 80 classes
INFO:__main__:Performing test run
100%|¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦| 100/100 [00:03<00:00, 30.85it/s]
INFO:__main__:Inference time (EdgeTPU): 26.43 +- 4.09 ms
INFO:__main__:NMS time (CPU): 0.77 +- 0.35 ms
INFO:__main__:Mean FPS: 36.77
(py36) josh@josh-jetson:~/code/edgetpu_yolo$ python3 detect.py -m yolov5s-int8-224_edgetpu.tflite --bench_speed
INFO:EdgeTPUModel:Loaded 80 classes
INFO:__main__:Performing test run
100%|¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦¦| 100/100 [00:03<00:00, 25.15it/s]
INFO:__main__:Inference time (EdgeTPU): 33.31 +- 3.69 ms
INFO:__main__:NMS time (CPU): 0.76 +- 0.12 ms
INFO:__main__:Mean FPS: 29.35
I would say that 96x96 is probably unusable unless it was a model that was properly quantisation-aware trained and was for a very limited task (see accuracy results below).
224px gives good results on standard images, e.g. zidane
, but it might not always find the tie. This is quite normal for edge-based models with small inputs.
You could attempt to tile the model on larger images which may give reasonable results.
Note that benchmarks use the same parameters as Ultralytics/yolov5; conf=0.001, iou=0.65. These settings significantly slow down performance due to the large number of bounding boxes created (and NMS'd). You will find that inference speed drops up to 50%. There are sample prediction files in the repo for the default conf=0.25/iou=0.45 - these result in a slightly lower mAP but are much faster.
-
96x96: mAP 6.3 , mAP50 11.0
-
192x192: mAP 16.1, mAP50 26.7
-
224x224: mAP 18.4, mAP50 30.5
Performance is considerably worse than the benchmarks on yolov5s.pt, however this is a post-training quantised model on images 3x smaller.
There are prediction.json
files for each model in the coco_eval
folder. You can re-run with:
python3 detect.py -m yolov5s-int8-224_edgetpu.tflite --bench_coco --coco_path /home/josh/data/coco/images/val2017/ -q
The q
option silences logging to stdout. You may wish to turn this off to see that stuff is being detected.
Once you've run this, you can run the coco_eval.py
script to process the results. Run with something like:
python3 eval_coco.py --coco_path /home/josh/data/coco/images/val2017/ --pred_pat ./coco_eval/yolov5s-int8-192_edgetpu.tflite_predictions.json --gt_path /home/josh/data/coco/annotations/instances_val2017.json
and you should get out something like:
(py36) josh@josh-jetson:~/code/edgetpu_yolo$ python3 eval_coco.py --coco_path /home/josh/data/coco/images/val2017/ --pred_pat ./coco_eval/yolov5s-int8-224_edgetpu.tflite_predictions.json --gt_path /home/josh/data/coco/annotations/instances_val2017.json
INFO:COCOEval:Looking for: /home/josh/data/coco/images/val2017/*.jpg
loading annotations into memory...
Done (t=1.92s)
creating index...
index created!
Loading and preparing results...
DONE (t=0.45s)
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=52.38s).
Accumulating evaluation results...
DONE (t=8.63s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.158
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.251
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.168
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.012
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.136
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.329
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.150
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.185
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.185
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.012
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.158
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.397
INFO:COCOEval:mAP: 0.15768057519574114
INFO:COCOEval:mAP50: 0.25142469970806514
This repository is an entry into the Ultralytics export challenge for the EdgeTPU. It provides the following solution:
- A minimal repository which has extremely few dependencies:
pycoral
,opencv
for image handling (you could drop this using e.g Pillow) andnumpy
- Other "light" dependencies include
tqdm
for progress reporting, andyaml
for parsing names files.json
is also used for output logs (e.g. benchmarks) - No dependency on Torch, which means no building Torch - from clone to inference is extremely fast.
- Code has been selectively taken from the original Ultralytics repository and converted to use Numpy where necessary, for example non-max suppression. There is essentially no speed penalty for this on a CPU-only device.
- I chose not to fork ultralytics/yolov5 because the competition scoring was weighted by deployment simplicity. Installing Torch and various dependencies on non-desktop hardware can be a significant challenge - and there is no need for it when using the tflite-runtime.
- Accuracy benchmark code is provided for running on COCO 2017. It's a slimmed down version of
val.py
and there is also a script for checking the output. mAP results are provided in this readme.- For the 224x224 model: mAP 18.4, mAP50 30.5
- Packages are easily installable on embedded platforms such as the Google Coral Dev board and the Jetson Nano. It should also work on any platform that an EdgeTPU can be connected to, e.g. Desktop.
- This repository uses the Jetson Nano as an example, but the code should be transferrable given the few dependencies required
- Setup instructions are given for the Coral, but these are largely based on Google's guidelines and are not tested as I didn't have access to a dev board at time of writing.
- tflite export is taken from https://github.com/ultralytics/yolov5/blob/master/models/tf.py
- These models have the detection step built-in as a custom Keras layer. This provides a significant speed boost, but does mean that larger models are unable to compile.
- Speed benchmarks are good: you can expect 24 fps using the EdgeTPU on a Jetson Nano for a 224 px input.
- You can easily swap in a different model/input size, but larger/smaller models are going to vary in runtime and accuracy.
- The workaround for exporting a 416 px model is to use an older runtime version where the transpose operation is not supported. This significantly slows model performance because then the
Detect
stage must be run as a CPU operation. See bogdannedelcu's solution for an example of this.- Note this approach doesn't work any more because the compiler supports the Transpose option. I tried exporting with different model runtimes in an attempt to force the compiler to switch to CPU execution before these layers, but it didn't seem to help.
- Extensive documentation is provided for hardware setup and library testing. This is more for the Jetson than anything else, as library setup on the Coral Dev Board should be minimal.
- A Dockerfile is provided for a repeatable setup and test environment