Food or non-food? Image Classification with TensorFlow ANN
git clone https://github.com/enyangxxx/tfFoodImageClassifier.git
-
You can download the dataset Food-5K here by using e.g. Cyberduck to access via FTP: https://mmspg.epfl.ch/downloads/food-image-datasets/
-
Create a folder 'images' in project root folder:
cd foodImageClassfier && mkdir images
- Copy the sub-folders 'training', 'evaluation', 'validation' into the 'images' folder.
The images with name starting with 1 are food images, names of non-food images start with 0. The training set contains 3000 images (50:50), the evaluation and validation sets contain both 1000 examples with an equal distribution between food images and non-food images.
In this project, I learned to use TensorFlow to build an Deep Neural Network and also how to choose different hyperparameters. The following values are chosen based on given circumstances, e.g. the small size of dataset.
- Number of epochs = 2000
- Size of mini-batches = 1000
- Learning rate = 0.0001
- Side length of an image = 20
- Number of layers = 8
- Dimensions of the layers (or call it number of units per layer) = [1200, 500, 100, 80, 50, 40, 10, 2]
The costs after each 100th epoch are the following:
They were also plotted in a learning graph, together with the training, cross-validation and test accuracies:
This reveals an effect of overfitting, because the model fits the training set very well, but less to the cross-validation set due to small size of dataset overall. A possible solution could be data augmentation to enlarge the dataset or regularization.
This is my first time ever showing the test result with single test images.. I am very very proud to achieve this result although I know that the model can still be improved. But my goal is definitely achieved and a big thanks to Andrew Ng, deeplearning.ai and Coursera for teaching me how to develop my own Deep Learning project. P.S. I used openCV, so BGR is used per default instead of RGB, hopefully that is fine :-)
The next step would be implementation of regularization to reduce variance. What I also want to try is the data augmentation. The goal is to achieve reduction of variance to increase accuracy for dev set.