Everytime I run the training, I get a very strange behaviour of the accuracy, is it normal?
Sometimes that “spike” is near the end of the curve
Here’s mine:
Hi @Franz90,
The initial accuracies suggest that you are using a model that already has a good idea about the dataset. However, the sudden drop in the training curve is not at all normal 
After epoch five, the curve you get matches the curve of a model that has been trained from scratch, matching the one by @TMosh.
Could you provide more information on whether you used a model with saved weights or trained it from scratch?
Hope this helps.