I am trying to work on an idea that is centered around predicting weather in the future. It is still in the pre-planning stage.
From the POV of data I realized that I will have to use heatmap kind of images of certain geographical region(s). I was exploring various models that could be helpful for this kind of a task.
The models that were trained on ImageNet didn’t have this type of data.
Can I try working this by removing some layers of, let us say ResNet and using the initial layers that according to the Matt Zeiler paper will be good for recognizing these gradient style images?
People have used ResNets “as is” for many purposes - even when they don’t look like the images from ImageNet.
Search for Audio classification using images as an example.
Adding onto what meanpenguin is saying; that’s what fine-tuning is for! You take the model, and further train it on what you want to predict.
Also, it’s better to begin with trying out ResNet so you have a baseline from which you can gauge whether those changes you’re making to the model are worth it, or whether other models you’re trying out are actually better than ResNet. It may turn out that ResNet is perfectly fine all along.