Hi(ve)! 🐝

This repo is the result of some work done in the Startup Weekend AI in Paris.

It contains two models:

• The first one is a very simple model based on CNN up-to-date best practice, reaching 98% percent accuracy
• The second one is a fine-tuned model fron vgg-19 which took too long to be retrained (no kidding...)

DISCLAIMER This repo does not contains the trainings/dev/test sets due to proprietary concerns.

• The simple model which take 3MB of memories and 6ms (on titan X) to compute an image is god damn accurate!

Completely blowing up previous bees larvae detections i know of using OpenCV, and this was achieved thanks to only 2000 training samples which is a very small dataset. Also, it took less than 10 minutes to train it 🚀

This validate again and gain the fact that deep learning is very well suited to handle real life data and its variability.

• The second model is not really useful for bees larvae detection, yet it shows how it is easy to fine-tune a model using TensorFlow (The VGG-19 model was taken from this site).

The training phase is interesting in terms of overfitting:

We can see that we reach 0% (:scream:) error on the training set which means we completely overfit the data, yet the generalization on the dev set keeps improving until no learning is possible anymore.

This is a clear indicator that more data would improve even more the accuracy, also we probably can simplify it even further and improve performance for this simple binary classifier.

• Run the ./vgg/download.sh script to download pretrained vgg weights

• Run python vgg/vgg.py to use a proper Saver to save the graph and weights (You can run python vgg/tf-vgg.py to check that results are the same)

• Finally you can check the file models/bee.py to see how i add my personnal classifier on top of the CNN and run python train.py --model complex to train it

• If you want to train the simple model, jsut use python train.py

virtualenv env -p python3.5
source env/bin/activate
pip install -r requirements.txt
# To install TensorFlow: https://www.tensorflow.org/versions/r0.11/get_started/index.html

You can test both models by running python test.py script.

And finally you can even export a frozen model using python freeze.py. if you want to use it in production with TensorFlow in a more convenient way.

MIT

(Check the LICENCE file)