First steps with deep learning. Dog breed recognition

Some time ago, I found great course about Deep learning (fast.ai). I watched few videos and was fascinated about results that could be achieved by using DL for visual recognition.
I wanted to try it how it is working by myself, and found active competition/playground on kaggle.
The goal of competition to identify breed of the dog on the picture.
Train data contains approximately 10K photos of dog of 120 different breeds. Test data - 10K images as well.
I started with playing to identify only 2 different breed - doberman vs pomeranian. For training I had 157 images, for test - 30 (15 for each breed)
First I tried simple architecture with only one convolution layer. I run training for 3-fold cross validation with 100 epoch for each fold.
Time of training each epoch - 4s
I got accuracy of 0.8 on my validation set. Full training took approximately 20 mins.
Whole source code can be found on github