Nuke 15.0 & 14.1 | Faster CopyCat training
78CopyCat brings the power of machine learning into artists' hands in an accessible way. We’ve seen some amazing and varied applications for CopyCat, such as complex clean up work and style transfers on CG renders. We’re excited to see how artists are using this new tool to reduce repetitive tasks and allow them to be more creative.
Up until now, it’s only been possible to train CopyCat on a single user machine. We’re introducing a much-requested feature to help artists continue to iterate faster - and that’s the ability to share the CopyCat training load across multiple machines.
We wanted to create a familiar interface so that distributing training would feel just like distributing rendering. You can now distribute training across multiple machines, using standard render farm applications such as OpenCue and Deadline.
We are changing the way copycat trains by introducing a new training strategy.
This speeds up CopyCat training by up to two times, whilst maintaining the same level of quality, and it’ll do it automatically.
How does it work? Under the hood, CopyCat lowers the resolution of the source and ground truth images, and trains on those first. This gives CopyCat a headstart when it then moves on to complete the training at higher resolutions, and finally the full resolution. This new strategy allows CopyCat to get to the final result faster.
With the distributed training updates, we needed one place to quickly visualise training progress, so that you can work on the rest of your shot, or even another shot, and come back to check the training progress using this UI when you need to.
There’s a new layout in the Progress tab (which was previously the Graphs tab) on the CopyCat Properties panel, adjusted with distributed training in mind. The Runs table is now at the top, and collapsible in a twirly dropdown, as are the graphs, if you just want to view the Contact Sheet.
The runs table displays run data for all training in the specified Data Directory. Note that the Runs table will include all runs in the data directory, whether these were produced by local or remote Nuke sessions.
You can select which run’s data you’re looking at for the Loss graph, and the contact sheet and preview panels by clicking on the checkboxes in the runs table.
The label for the contact sheet is colour-coded to match the run that’s currently chosen in the runs table.
The new “Live Updates” checkbox lets you check on the progress of a distributed training job. This will fetch the latest loss data every 30 seconds from the network location of the training. It’s off by default so that it’s not unnecessarily using the network to fetch data when it’s not needed.
Distributing your CopyCat training gives you a productivity boost. It allows for faster iterations, whilst you can carry on comping.
About Us:
We are the creators of industry-standard visual effects, computer graphics and 3D design software for the Digital Design, Media and Entertainment industries. Since 1996, Foundry has strived to bring artists and studios the best tools for their workflows so they can battle industry constraints whilst staying creative.
Subscribe to our channel and get the latest news, tutorials, webinars and updates from the Foundry team.
Follow Us:
/ thefoundryteam
/ foundryteam
/ thefoundryteam
/ foundryteam