Ezekiel was another attempt at high-resolution photography inputs and regardless of parameters or modifications to the original Cain network, it didn’t seem to produce decent results. Until we have more powerful GPU hardware we are going to put this model to one side and focus on our main strand which Read more…
Now that we have the new OpenStack platform we have used this to generate images from the Daniella network. Results are coming in faster than ever before. The M10 cards work very well and we can naturally break them down into multiple parallel jobs since there are 4 GPU’s per Read more…
Well, we have acquired some Tesla M10 cards and are looking to utilise PCI passthrough with OpenStack to split down the GPU’s into 4 instances. The Tesla M10 has 4 x GPU per card, each with 8GB of memory. This will mean we can start running smaller models (tests) in Read more…
We developed a CNN called Cain and started working with SGD Photos to generate artworks based on their photos. Now, due to the high-resolution nature of the images, it took quite some time to prepare images that fit the profile required. The benefits were that photographs did not need cropping Read more…
We didn’t quite get the results we expected with Becky, regardless of the parameters we set and learning rates and batches we utilised the CNN just didn’t produce the results we expected – She got stuck in a failing loop every time. We have put the results on the main Read more…
Thanks to the guys at Coreix we now have a new private cloud hardware platform on OpenNebula which is running the nVidia RTX 3060 series cards. It isn’t ideal but that is due to our budgetary constraints at this time, nevertheless, we can push forward with this and run the Read more…
We have spoken to Coreix regarding the cards and based on our budgets at this time we are going to install a small Open Nebula solution with some nVidia RTX 3060 series cards. The memory is sufficient to get us started at a low resolution which we need to keep Read more…
We have changed the source and increased the number of images to thousands. This has helped us train the network more efficiently. Whilst it still takes time due to limited hardware we are in discussion with our hosting provider Coreix in relation to what hardware we can upgrade to or Read more…
Well, we have finally started to make some progress around with the network. We have had mixed results when using the datasets with limited sources but we are proposing to keep a track of very basic progress on this blog. The generation is using the works from asbt.ract.me and we Read more…