First computer vision project using Kaggle and TPUs

Some personal notes regarding the project

• Make sure to create different TFRecord formats for training (labeled) and test (unlabeled) datasets.

• When loading dataset better make test data ordered (idk why but otherwise I had random shuffled results). Making training set "ignore_order.experimental_deterministic = False" can increase computations.

• When creating data pipelines we are using two separate functions to work with the training dataset: the first one is used it the training itself, and the second one is used in the evaulation and visualization. The main difference -> the second one doesn't use repeat() not to make the dataset infinite and doesn't use augmentation.

• Check class distribution (may use visuals)

• Transfer Learning using Keras using Functional API

• Use warm-up training with few (3) epochs with pre-trained model frozen and decrease the learning rate for the actual learning.

• Set STEPS_PER_EPOCH = NUM_TRAINING_EXAMPLES // BATCH_SIZE

• sklearn.metrics.classification_report for f1, presicion and recall