This project is done for Deep Learning for Computer Vision course (WiSe 2017/18) offered by Technical University of Munich, Faculty of Informatics, Computer Vision Group. Our project consist of 2 parts.

1. Using CNNs (convolutional neural networks) to predict tags for restaurant images. This is a kaggle competition which is organized by Yelp. https://www.kaggle.com/c/yelp-restaurant-photo-classification
2. Using RNNs (Recurrent Neural Networks) to generate fake reviews for restaurants. The dataset can be downloaded via https://www.yelp.com/dataset.

In this project, we have used Keras with Tensorflow backend. We have installed these dependencies via anaconda. We have used pandas for dataframe operations.

• keras 2.1.3
• tensorflow 1.4.1
• pandas 0.20.3

The folder structure should be like below:

/split.py
/organize_photos.py
/reorganize_photos.py
/MultilabelGenerator.py
/TopClassifier.py
/utility.py
/VGG16.py
/VGG16_bottleneck.py
/InceptionV3model.py
/InceptionV3_bottleneck.py
/XceptionModel.py
/Xception_bottleneck.py
/train_vgg16.py
/extract_vgg16_bottleneck_features.py
/train_vgg16_bottleneck.py
/train_inceptionv3.py
/extract_inceptionV3_bottleneck_features.py
/train_inceptionV3_bottleneck.py
/train_xception.py
/extract_xception_bottleneck_features.py
/train_xception_bottleneck.py
/predict_inceptionv3.py
/predict_xception.py
/predict_test.py
/parse_reviews.py
/RNNTextGeneration.py
/train_rnn.py
/RNN.py
/generate_text.py
/data/
/data/learning/
/data/learning/review.json
/data/learning/train.csv
/data/learning/train_photo_to_biz_ids.csv
/data/learning/photos - all training photos should be in this folder first
/data/learning/photos/train
/data/learning/photos/validation
/data/learning/models - it should be empty before training
/data/test
/data/test/test_photo_to_biz.csv
/data/test/photos
/data/test/photos/test - all test photos should be in this folder

1. Check that all train.csv and train_photo_to_biz_ids.csv are in the /data/learning/ folder, and all of the training photos are in the /data/learning/photos/ folder.
2. Run python3 split.py to split the learning data to training and validation. Don't forget to set the test_size attribute in line 5 to your wanted rate of training and validation data.
3. Run python3 organize_photos.py to move the training and validation photos to their folders.

1. Check that all train.csv and train_photo_to_biz_ids.csv are in the /data/learning/ folder, and all of the training photos are in the /data/learning/photos/ folder.
2. Run python3 reorganize_photos.py command to move the photos their original folder.
3. Run python3 split.py to split the learning data to training and validation. Don't forget to set the test_size attribute in line 5 to 0 for no validation data.
4. Run python3 organize_photos.py to move the training and validation photos to their folders.

1. Run python3 train_inceptionv3.py command. After this command executes, there should be a weights_inceptionv3.hdf5 file in your /data/learning/models/ folder.
2. Run python3 predict_inceptionv3.py command to get the predictions and F1-Scores for your training (and for hyperparameter research validation) data.

1. Run python3 train_xception.py command. After this command executes, there should be a weights_xception.hdf5 file in your /data/learning/models/ folder.
2. Run python3 predict_xception.py command to get the predictions and F1-Scores for your training (and for hyperparameter research validation) data.

1. Run python3 predict_test.py command. After this command executes, there should be a results.csv file in your /data/test/ folder. You can upload this file to kaggle.

1. Check that your review.json file is in the folder /data/learning/.
2. Run python3 parse_reviews.py command, to generate a better file to feed RNN.

1. Run python3 train_rnn.py command. This command will create rnn.hdf5 file under /data/learning/models/ folder.

1. Run python3 generate_text.py command. This command will generate text for each business label.

https://keras.io/applications https://keras.io/preprocessing/image/ https://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html https://www.youtube.com/watch?v=BhQW2OLzx_c https://medium.com/@chengweizhang2012/an-easy-guide-to-build-new-tensorflow-datasets-and-estimator-with-keras-model-9b0f6b4c1b0d https://github.com/fchollet/deep-learning-with-python-notebooks/blob/master/5.3-using-a-pretrained-convnet.ipynb

https://stackoverflow.com/questions/44164749/how-does-keras-handle-multilabel-classification http://www.kubacieslik.com/extending-keras-imagedatagenerator-handle-multilable-classification-tasks/ https://depends-on-the-definition.com/guide-to-multi-label-multi-class-classification-with-neural-networks-in-python/

https://machinelearningmastery.com/text-generation-lstm-recurrent-neural-networks-python-keras/ https://chunml.github.io/ChunML.github.io/project/Creating-Text-Generator-Using-Recurrent-Neural-Network/