I lowered the requirements by editing yolov3_train.cfg to:
batch=32
subdivisions=8
Thus making training faster. Although it’s certainly not as effective. Another thing I could do is change the image size but I think that this is enough optimization.
width=416
height=416
It allowed the training to get much further on Google Colsb, by saving more often. I also noticed that the time left is variying more often, not staying fixed or just decreasing which may mean that training further is pointless.