# D1.5
Mark Zakharov and Danila Romanov

# What was done

We set up a training script at Kaggle kernel. We first tried to do it at Colab, but met a strange bug with no information in the Internet besides a single github issue thread.
We trained an EfficientNetB5 model on 200x200 images on a Kaggle TPU and achieved decent training accuracies (about 75% in 8 epochs), however, validation accuracy turned out to be unexpectedly low: about 6%. Currently, we are investigating the roots of the problem. It might be due to uniform sampling of validation set, i.e. something to do with label distributions. Obviously, our next goal is to figure out the solution to the problem and then continue improving the model.