packages# LSDA 2018, DIKU - Cph University

Group 4 (Emil, Frederik, Roberta, Steffen)

Data saving and loading

• Download images, resize to 299x299 (done in scraper) and save as HDF5 files
• Figure out how to load HDF5 data into keras

Model

• Use Inception-v3 as template.
• Training: Train dense layers first (save features outputtet by the conv-part, and train the dense layers only on that), then the whole architecture. Possibly implement boosting to extend rare classes.
• Augment data on the fly: Follow Jupyter Notebook (CNNs_In_Practice_Data_Augmentation.ipynb) and possibly use this for extending rare classes
• Validate and create tensorboard graphs of accuracy
• Perform test predictions
• Implement tensorflow micro f1 score function as metric

NOTES

Downloading the data

Data can be downloaded with the scraper. To download all data (expected size: 25GB (with resizing)):
$ cd src
$ python scraper.py

specific directories and the amount of files to download can also be specified. To download the first 500 images from the training dataset do:
$ cd src
$ python scraper.py ../data/labels/train.json ../data/raw_images/train 500

Using the Framework

The framework is hopefully an easier way to access and keep track of our data, and try out classifiers. You can try out the current test version by: $ cd src
$ python framework.py

Kaggle Link:

https://www.kaggle.com/c/imaterialist-challenge-fashion-2018

Team Name: K-NN or GTFO

https://www.kaggle.com/c/imaterialist-challenge-fashion-2018/team

Overleaf Links

• Report 0: https://www.overleaf.com/15811080jqgcpbxscxdj#/60201263/
• Report 1: https://www.overleaf.com/16854886dwykvxvpyxcn#/64493245/

Azure

login: lsda CompetitionGroup4

Connect with ssh [email protected]

Source into tensorflow enviroment: source lsdaenv/bin/activate

!!! DO NOT FORGET TO TURN OFF THE SERVER !!!

Useful libraries

• CV2: https://docs.opencv.org/3.0-beta/doc/py_tutorials/py_imgproc/py_table_of_contents_imgproc/py_table_of_contents_imgproc.html#py-table-of-content-imgproc
• H5PY: http://docs.h5py.org/en/latest/quick.html

Stuff

• Should the "missing"-images be deleted?
• Should (and can) we do the augmentation beforehand, i.e., not on the fly? Yes! https://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html, https://machinelearningmastery.com/image-augmentation-deep-learning-keras/
• Initializer for the on-top NN (https://keras.io/initializers/): Use either Glorot (http://proceedings.mlr.press/v9/glorot10a/glorot10a.pdf, http://andyljones.tumblr.com/post/110998971763/an-explanation-of-xavier-initialization) or He (https://arxiv.org/pdf/1502.01852.pdf), preferably uniform (keras-team/keras#52). (Use He! Remember to avoid dead neurons in the first layer).
• Check if our pre-trained net uses ranges [0;255] or [0;1].
• Find a way to implement real-time (or almost) feedback during training.
• Using a pre-trained model tutorials: https://keras.io/applications/, https://www.learnopencv.com/keras-tutorial-fine-tuning-using-pre-trained-models/.
• Remember to use the Keras function preprocess_input().
• Why SGD is batch-size-dependent: https://stats.stackexchange.com/questions/316464/how-does-batch-size-affect-convergence-of-sgd-and-why

Notes for Sunday:

• Setup framework for loading batches (of 1k or 10k images) - can new batch be loaded while training on previous one?
• Make sure that images and labels match
• Figure out how to load only batches of CNN_features into memory
• Boosting? Not really possible: https://stackoverflow.com/questions/39063676/how-to-boost-a-keras-based-neural-network-using-adaboost

What to include in the report/presentation

• Choice of approach: Incv3 (small, good performance) with new top-layer, then retrain last layer of Inv3.
• We experimented a lot with architecture and optimizers, and went with 2x1024 with BN. One layer didn't attain same performance, while three layers didn't perform better. Well-tuned SGD beat ADam, both of which beat RMSProp by far. Regularization was important to avoid overfitting: Drop-out worked horribly (maybe Keras/Tensorboard didn't turn off the drop-out when validating?). L2-reg. was really bad.
• Data augmentation: We did data augmentation when training the full model. Data augmentation when training the top layer would have been quite difficult (as we would have had to save the augmented images), since we trained it on the CNN features. We looked into a scheme for augmenting low-freq. labels (for "flattening" the distribution of labels)
• Test data augmentation
• Weird "bug" with overfitting when training the top layer from the very beginning - but validation score was still improving.
• We only had 30 GBs of server space, which wasn't enough, so we had to put everything on a temp. server drive, which was deleted as soon as we logged out.