Add Predictor, TripletNet, RandomFlip, 2D FCN

[pschubert]

• added class to perform inference with elektronn3 trained models
• added RandomFlip augmentation to randomly and independently flip input (and target) data along spatial dimensions

[mdraw]

Regarding "imposed_patch_size" support in InferenceModel.predict_proba() (splitting inputs to smaller chunks ("tiles") for inference):

To prevent confusion, I think we should call this "spatial chunking" feature "tiling", like in ELEKTRONN2. Proposed definitions/usage:

• A chunk is a slice of a data set that can be efficiently loaded from disk and stored in host memory (disregarding neural network specifics). This is in line with HDF5 and knossos_utils's usage of this term.
• A tile is a slice of a chunk (either the complete chunk or a part of it) to be loaded to GPU memory and to be used for one forward pass (inference). Tiles should be as large as possible, while keeping the following conditions:
  • The GPU doesn't run out of memory when performing a forward pass on it (if it does, the tile shape has to be reduced).
  • The tile shape is supported by the network architecture (e.g. pooling operations prevent certain input shapes).
  • The tile shape is divisible by the chunk shape (we could theoretically relax this constraint, but I'd rather avoid it for sake of simplicity). To my current knowledge, it only makes sense to increase the imposed batch size (batch_size in predict_proba) to anything > 1 if the tile size (imposed_patch_size) is already equal to the chunk size and the GPU memory is still only less than half utilized.

Relevant ELEKTRONN2 code: https://github.com/ELEKTRONN/ELEKTRONN2/blob/c3da2027a895b7aa2425ac567b3c90785204c8e6/elektronn2/neuromancer/node_basic.py#L950-L996.

@pschubert do you have other thoughts on this issue?

[mdraw]

I'm merging this PR and will work on input patch tiling later, as it's more complicated than I initially thought (taking overlapping tiles and interpolation of overlapping regions into consideration).