Week 8 Farzain Majeed
Reading Dataset Much of the work involved with training a neural net comes from experiments and actually formatting the data The CommaAI is rather tricky in that the file format is .h5 which we first un-package and save the frames as .jpg’s This makes the dataset much easier to work with A lot of spent was time scripting and automating this whole process.
YOLO Next we needed to use YOLO, preferably in Keras. I found an open source project by the name of YAD2K which provided the YOLO9000 weights and model in a Keras/TF friendly manner. YOLO outputs many classes, while we are looking for just a few (car, pedestrian, traffic light) It may be worth it in the future to retrain YOLO to only predict these parameters. Though, this may actually lead to worse performance because of a lack of data.
Early Results Early results with the new attention based model were interesting. I found that the training loss nicely converged to an amount lower than on the old model. The validation loss jumped around more, but at one point converged to a smaller loss as well. The main problem is that the side task loss didn’t converge at all, which indicates it did not learn
Explanation and Next Week Its possible that the noise from the side task (or even having a side task itself) led to less overfitting and improved results. I need to keep on adjusting parameters to help it converge. I also would like to look more into visualizing the layers to see what the network pays attention to.