edgetpu | Coral issue tracker (and legacy Edge TPU API source) | Computer Vision library

Since the Yolov5s model has a different input file than the EfficientDet, the output tensor will be different. The trick here is understanding how to process this output tensor.

Fortunately, Ultralytics/Yolov5 held an export competition where the goal was to execute Yolov5 models on EdgeTPU devices.

This guy Josh won the coral devboard section. He wrote python library to process these wonky tensor outputs from Yolov5s models. Here is the repo. The real processing of the output tensor is done in his non-max-suppression code.

I've forked his repo and added the ability to execute/process these Yolov5s models on desktops.

Thanks so much Josh!

It seems to be a bug in the PyCoral API. To solve the issue, I replaced the last line from the "detect.py"-file (in my case located in "/usr/lib/python3/dist-packages/pycoral/adapters/detect.py") with this updated line:

return [make(i) for i in range(len(scores)) if scores[i] >= score_threshold]

Basically what you want to do is make a dict where the key is the class and the value is a color in the same format that is here.

cv2.rectangle(image, (xmin,ymin), (xmax,ymax), (10, 255, 0), 2)

Replace (10, 255, 0) with something like color_dict[classes[i]] and then you will be able to get a different color for each class.

I recently came into this for a thesis supervision. We tested face detection in a raspberry pi 4 with Coral USB an without (inference on rpi CPU). Are you using the same model file for both? If this is the case, then this is the problem. You need to use the bare tflite model for the CPU inference and the TPU-compiled model for the inference with TPU. You can take a look at this repo where you can find the code I was mentioned before (it's not well documented but it's working, look at the inference CPU and inference CORAL files).

The process, as @JaesungChung answered is well done.

My problem was on the application which was running the .tflite model. I quantized my model output to uint8, so I had to reescale my obtained values to get the right results.

I.e. I had 10 objects because I was requesting all the detected objects with an score above 0.5. My results were no scaled, so the detected objects scores could be perfectly 104. I had to reescale that number dividing by 255.

The same happened when graphing my results. So I had to divide that number and multiplicate by the height and width.

The prebuilt tflite_runtime package set from the above site does not cover armv6 architecture yet.

Alternatively, you can choose some other options.

(1) Install the TensorFlow pip package.

TensorFlow Lite features are a part of TensorFlow package and the prebuilt TensorFlow pip packages support armv6. See https://www.tensorflow.org/install/pip

(2) Build your own tflite_runtime through Bazel or CMake.

If there is a need for installing the tflite_runtime only, it is possible to build the tflite_runtime by yourself. The following document describes the differences between Bazel and CMake and how to build the tflite_runtime through them.

https://www.tensorflow.org/lite/guide/build_arm

This is the correct output for a SSD model. The TFLite_Detection_PostProcess is the custom op that is not run on the EdgeTPU. If you run netron on one of our default SSD models on https://coral.ai/models/, you'll see the PostProcess runs on CPU in that case.

In the case of your model, every part of the of the model has been successfully converted. The last stage (which takes the model output and converts it to various usable outputs) is a custom implementation in TFLite that is already optimized for speed but is generic compute, not TFLite ops that the EdgeTPU accelerates.

This is a bug in the coral software. According to this thread https://github.com/google-coral/edgetpu/issues/260 they messed up a few things in the newest version (at the time 2.5.0). Starting with a fresh virtual environment and using the release 2.1.0 and corresponding driver for Python 3.7 (3.8, 3.9 not supported as of 2.1.0) fixed the issue.

From that thread:

For now, I suggest rolling back to the older driver: https://dl.google.com/coral/edgetpu_api/edgetpu_runtime_20200728.zip

And you also need to remove your current tflite_runtime and downgrade it to an older version (make sure to change to the right python version):

pip3 install https://dl.google.com/coral/python/tflite_runtime-2.1.0.post1-cp36-cp36m-win_amd64.whl pip3 install https://dl.google.com/coral/python/tflite_runtime-2.1.0.post1-cp37-cp37m-win_amd64.whl Apologies, we are working to get this fixed ASAP

I figured out the answer, it was pretty stupid, the library has to be built as a shared library: