• Stars
    star
    125
  • Rank 286,335 (Top 6 %)
  • Language Objective-C++
  • License
    GNU General Publi...
  • Created over 6 years ago
  • Updated over 1 year ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

Built with TensorFlow, Yolov2 Tiny and React Native | Flutter | Javascript

中文文档

Not Hotdog

Personal implementation of Jin Yang's Not Hotdog app in HBO's Silicon Valley.

seefood

App developed with:

Model trained on:

Links


Building the Model

1. Collecting data set

I used COCO dataset 2014 to train the model. COCO dataset contains 80 thing classes, one of which is "hot dog".

I created a Python script (./yolo/coco2yolo.py) to extract all the hot dog images (800+ from train and 400+ from val) and convert the annotations to yolo format.

2. Training Yolov2 Tiny

The model is trained with a single class "hotdog" using this fork of darknet.

darknet.exe partial yolov2-tiny.cfg yolov2-tiny.weights yolov2-tiny.conv.13 13

The .cfg file and initial weights can be found in ./yolo directory

darknet.exe detector train data\obj.data yolov2-tiny-hotdog.cfg yolov2-tiny.conv.13

3. Converting weights to TensorFlow protobuf (.pb)

The yolo weights is converted to TensorFlow model using darkflow:

flow --model ../yolov2-tiny-hotdog.cfg --load ../yolov2-tiny-hotdog_final.weights  --savepb

The saved .pb file can be found in ./yolo directory.

4. Quantization

The saved .pb is about 44MB. I used the quantization script in Tensorflow repo to quantize and reduced the size to 11MB.

python3 tensorflow/tools/quantization/quantize_graph.py --input=yolov2-tiny-hotdog.pb --output_node_names=output --output=quantized_yolov2-tiny-hotdog.pb --mode=weights

The quantized .pb file can be found in ./react-native-NotHotdog/ios/NotHotdog/data directory.