Since the results of the blob inference time documented in the previous log was a bit disappointing, I thought I'd try retraining the MobileSSD network with no 'pretrained model' as i thought that this model was probably what prevented me reducing the number of classes to train option. All the solverstate files and caffemodels had to be deleted in the 'results' folder as the software tried to carry on from the previous train. Initially the software complained that their was no pretrained model, but setting: 'pretrain_model = ""' fooled the python script into carrying on regardless. 'num_classes must be set to a minimum of 2 as their always needs to be a 'dont care' class.

The model is now being trained on the Jetson TX2, which is a bit slow for this kind of work, but it's all I've got other than messing about on AWS. It'll take at least 2 days, maybe a lot longer .... Time will tell!