Nathan’s Status Report 04/10

This week there were plenty of hurdles for the deblurnet CNN and I’m sure there will be more to come. The training code was finally finished and the data was properly split into frames of 128×128 and all that was left was to deploy the ec2 instance and begin training. The problem is that on aws the p3.2xlarge container that had the GPU capabilities for training a large dataset like the one we have requires 8 vCPU’s and my account, being newly created, was allowed 0 vCPU’s. Because of this I have been dealing with aws customer service where I have been requesting more capacity. Currently my request is under review and I am not sure when I will be able to launch the instance and start training. But with the upcoming demo I want to be able to show off the capabilities of the CNN, so given that the paper I am basing this architecture off of is open source, I reached out to the author of the paper and received the pretrained weights from their results. Once I had these weights I tried to load them into my pytorch model in python, but there is a major lack of support for torch models from lua and models in pytorch. After hours of hopeless trying to convert the lua model to pytorch I realised it was not probable, so I will most likely be building the model in lua, just as the paper did to show some real time results in the demo. Hopefully my aws request will be fulfilled soon and we can begin training the model.