I noticed that the dataset is composed of png files in RGBA mode.
I’m planning on implementing the bigger original version of the U-net, with a much larger dataset. So in order to save a bit o storage, how can I change the code in order for it to work with png files in RGB mode?
I just recreated the dataset using Carla Simulator. My code works fine with the original dataset, but for some reason it does not with the new data. No idea why.
I keep getting ‘nan’ in the loss function, so the model can’t learn.
Any help is appreciated.
There is nothing magic about which image format you use. You are importing them images in such a way that they end up with numeric values in memory that represent the pixel colors. You have to study the way they handle the images. Notice how they strip off the Alpha channel and then for the labels they use only one channel which is the categorical value of the label on the given pixel. You have to figure out how to get from whatever your preferred input format is to that numeric in memory format. If you change the files, it’s your job to figure out how to do that. The first step is to study the existing code enough to understand what it is doing and why. Then map that to your new dataset.
Well, that wasn’t really helpful. I figured anyways. The alpha was misleading…
The current Carla Simulator uses 35 classes instead of the previous 23 used in the given dataset. Thanks anyway.