In-memory learning algorithms (partially) with pytorch

[maljoras]

Related issues

closes #560

Description

The TransferCompounds that implemts the Tiki-taka learning rules are encapsulated in RPUCuda for speed. However, for ease of experimentation, a python version would be good. Also for hardware implementation, the digital operations are best separated from the analog operations, so that the connection to any HW call is easier.

Here the TorchTransferTile is introduced that implements TTv2, c-TTv2, as well as AGAD in torch. That is the transfer part is explicitly available, while the pulsed update is done with single tiles (in RPUCuda).

Details

Usage:

See example 26. The learning rule is first defined as before using the ChoppedCompound etc. Then the tile_class property of the rpu_config is simply set to the TorchTransferTile instead of the default tiles.AnalogTile (that would be using the RPUCuda library)

rpu_config = build_config('agad', device=SoftBoundsReferenceDevice())
rpu_config.tile_class = TorchTransferTile

convert_to_analog(model, rpu_config)