Machine Learning & Food Classification

This repository contains code for fine-tuning a Convolutional Neural Network in order to classify a set of food images. It is based on all the work carried out for the Capstone project for Springboard's Data Science Intensive course.

Dependencies:

Dependencies can be installed with:

pip install requirements.txt

Caffe needs to be compiled (from master branch), including PYTHON wrapper make pycaffe

Instructions for use:

1. Clone the repo

git clone https://github.com/simonb83/food_classification

2. cd into the food_classification folder

3. Get the data and relevant model weights.

If you wish to run both training and testing, use flag -m TRAIN:

python get_data.py -m TRAIN

If you only wish to run testing code, use flag -m TEST:

python get_data.py -m TEST

Training:

1. Run data augmentation:

python run_augment.py

2. Create lmdb databases for caffe

./path-to-caffe-root/build/tools/convert_imageset  \
data/augmented/ \
data/augmented_train.txt \
data/train_lmdb

./path-to-caffe-root/build/tools/convert_imageset  \
data/augmented/ \
data/augmented_val.txt \
data/val_lmdb

3. Generate mean image of training data

./path-to-caffe-root/build/tools/compute_image_mean \
-backend=lmdb \
data/train_lmdb \
data/alexnet_4_mean.binaryproto

4. Update paths in train_val.prototxt

Update path to data_classification directory in mean_file and source lines:

mean_file: "~your_path/food_classification/data/alexnet_4_mean.binaryproto"
source: "~your_path/food_classification/data/train_lmdb"

source: "~your_path/food_classification/data/val_lmdb"

5. Update paths in solver.prototxt

Update path to data_classification directory in net and snapshot_prefix lines:

net: "~your_path/food_classification/model/train_val.prototxt"
snapshot_prefix: "~your_path/food_classification/model/"

6. (Optional) Change flag from GPU to CPU in solver.prototxt

solver_mode: GPU

7. Train the model

./path-to-caffe-root/build/tools/caffe train \
-solver /food_classification/model/solver.prototxt \
-weights /food_classification/model/bvlc_alexnet.caffemodel \
\> /food_classification/data/train.log

Testing:

1. Convert mean image

python convert_mean.py -m data/alexnet_4_mean.binaryproto

2. Make predictions

python run_predict.py

To visualize predictions, use the visualize.ipynb notebook.

Making New Predictions:

The model can also be used for making predictions on new images.

Requirements:

1. If you have not already done so, download data with TEST flag:

python get_data.py -m TEST

2. Convert mean image:

python convert_mean.py -m data/alexnet_4_mean.binaryproto

3. Add some images to the 'data/images' directory; the expected image format is jpg.

4. From the root directory run the prediction script, with -m 0 for GPU, -m -1 for CPU:

python predict_new.py -m 0

5. The predictions will be output to 'data/images/predictions.txt'

See more here