Edge TPU - Tiny YOLO v3

This repository contains the instructions and scripts to run the Tiny YOLO-v3 on Google's Edge TPU USB Accelerator. Edge TPU can only run full quantized TF-Lite models. If you already have a converted model, simply run inference.py with --quant and --edge_tpu to test it.

usage: Run TF-Lite YOLO-V3 Tiny inference. [-h] --model MODEL --anchors
                                       ANCHORS --classes CLASSES
                                       [-t THRESHOLD] [--edge_tpu]
                                       [--quant] [--cam] [--image IMAGE]
                                       [--video VIDEO]

optional arguments:
  -h, --help            show this help message and exit
  --model MODEL         Model to load.
  --anchors ANCHORS     Anchors file.
  --classes CLASSES     Classes (.names) file.
  -t THRESHOLD, --threshold THRESHOLD
                        Detection threshold.
  --edge_tpu            Whether to delegate to Edge TPU or run on CPU.
  --quant               Indicates whether the model is quantized.
  --cam                 Run inference on webcam.
  --image IMAGE         Run inference on image.
  --video VIDEO         Run inference on video.

Note: The inference code should be run on Tensorflow 1.15.0, even though the conversion below requires TF 2.0 nightly packages. I recommend using separate anaconda environments for inference and conversion.

---

Conversion guide

1 - Convert darknet .weights to Keras model

The network can be trained using either the original darknet implementation (https://github.com/pjreddie/darknet) or one of its forks (e.g. https://github.com/AlexeyAB/darknet).

Important note: The Edge TPU does not support the Leaky ReLU function, so it should be replaced by the regular ReLU. (https://coral.ai/docs/edgetpu/models-intro/#supported-operations)

The darknet weights can be converted to a Keras model using: https://github.com/qqwweee/keras-yolo3. However, this implementation doesn't support the regular ReLU function. Additionally, the TF-Lite conversion later also requires the input shape to be explicit (instead of None). So, I've made a fork with the (stupidly simple) required modifications here: https://github.com/guichristmann/keras-yolo3.

python convert.py tiny-yolo-cfg.cfg darknet-weights.weights keras-filename.h5

2 - Convert Keras to TF-Lite model with full integer quantization.

To run on Edge TPU, we need to convert the Keras model to TF-Lite and apply post-training full integer quantization. https://www.tensorflow.org/lite/performance/post_training_quantization.

python keras_to_tflite_quant.py keras-model.h5 output-filename.tflite

Note: The quantization of the RESIZE_NEAREST_NEIGHBOR op is only supported in Tensorflow 2.0 nightly packages as of now, so you need use that version for thconversion.

pip install tf-nightly

3 - Compile with Edge TPU compiler

Install the Edge TPU library and compiler: https://coral.ai/docs/edgetpu/compiler/.

Run the compiler on the TF-Lite quantized model:

edgetpu_compiler quantized.tflite

If everything is correct you should get a log with every op mapped to Edge TPU:

Edge TPU Compiler version 2.0.267685300
Input: quantized.tflite
Output: quantized_edgetpu.tflite
Operator                       Count      Status
RESIZE_NEAREST_NEIGHBOR        1          Mapped to Edge TPU
MAX_POOL_2D                    6          Mapped to Edge TPU
CONCATENATION                  1          Mapped to Edge TPU
QUANTIZE                       4          Mapped to Edge TPU
CONV_2D                        13         Mapped to Edge TPU

This model can be run on Edge TPU with inference.py script.