I need some advice on how to manage and create datasets for a photo editing model I’m working on. Im fairly new to tensorflow so please be patient My training data is photos where each composition folder includes raw images in .dng format (with exif data) and a professionally edited final target image. We will call this a batch. Raw images consist of bracketed exposures with 5 or so ambient and 2-4 flash. Currently these images are blended together manually in photoshop to create the final target image. So I’m looking at how to structure my data and build architecture for this model. Essentially my goal is train a model that learns how to blend these together using pixel based attention. And I would love to know the best practice for building these data sets, and what architecture I should use to accomplish this. I’ve tried tfrdatatset and ragged tensors. But I’m running into a lot of challenges with variable length stacks. The other issue is I need to align the raw images and target images and this is problematic because the raw dimensions are much larger than the target image. So I’ve attempted to resize the raw images to the target size and align them to the target images. I’ve so far been successful at this, but I’m unsure how dimensions play into the bigger picture. Since my goal is to be able to pass any size image through the model. So if I align and train with smaller dimension raw images and target images, how do I manage full size images when it comes time to use the model. Continue reading...
