aihwkit.simulator.tiles.custom module

High level analog tiles (floating point).

class aihwkit.simulator.tiles.custom.CustomRPUConfig(tile_class=<class 'aihwkit.simulator.tiles.custom.CustomTile'>, runtime=<factory>, pre_post=<factory>, tile_array_class=<class 'aihwkit.simulator.tiles.array.TileModuleArray'>, mapping=<factory>, simulator_tile_class=<class 'aihwkit.simulator.tiles.custom.CustomSimulatorTile'>, forward=<factory>, backward=<factory>, update=<factory>)

Bases: MappableRPU, PrePostProcessingRPU

Configuration for resistive processing unit using the CustomTile.

Parameters:

• tile_class (Type)

• runtime (RuntimeParameter)

• pre_post (PrePostProcessingParameter)

• tile_array_class (Type)

• mapping (MappingParameter)

• simulator_tile_class (Type)

• forward (IOParameters)

• backward (IOParameters)

• update (CustomUpdateParameters)

backward: IOParameters

Input-output parameter setting for the backward direction.

forward: IOParameters

Input-output parameter setting for the forward direction.

simulator_tile_class

Simulator tile class implementing the analog forward / backward / update.

alias of CustomSimulatorTile

tile_array_class

Tile class used for mapped logical tile arrays.

alias of TileModuleArray

tile_class

Tile class that corresponds to this RPUConfig.

alias of CustomTile

update: CustomUpdateParameters

Parameter for the update behavior.

class aihwkit.simulator.tiles.custom.CustomSimulatorTile(x_size, d_size, rpu_config, bias=False)

Bases: SimulatorTile, Module

Custom Simulator Tile for analog training.

To implement specialized SGD algorithms here the forward / backward / update are explicitly defined without using auto-grad.

When not overriden, forward and backward use the analog forward pass of the TorchSimulatorTile.

Update is in floating point but adds optionally noise to the gradient.

Parameters:

• x_size (int)

• d_size (int)

• rpu_config (CustomRPUConfig)

• bias (bool)

backward(d_input, bias=False, in_trans=False, out_trans=False, non_blocking=False)

Backward pass.

Note

Ignores additional arguments

Raises:

TileError – in case transposed input / output or bias is requested

Parameters:

• d_input (Tensor)

• bias (bool)

• in_trans (bool)

• out_trans (bool)

• non_blocking (bool)

Return type:

Tensor

extra_repr()

Extra documentation string.

Return type:

str

forward(x_input, bias=False, in_trans=False, out_trans=False, is_test=False, non_blocking=False)

General simulator tile forward.

Note

Ignores additional arguments

Raises:

TileError – in case transposed input / output or bias is requested

Parameters:

• x_input (Tensor)

• bias (bool)

• in_trans (bool)

• out_trans (bool)

• is_test (bool)

• non_blocking (bool)

Return type:

Tensor

get_brief_info()

Returns a brief info

Return type:

str

get_d_size()

Returns output size of tile

Return type:

int

get_learning_rate()

Get the learning rate of the tile.

Returns:

learning rate if exists.

Return type:

float | None

get_meta_parameters()

Returns meta parameters.

Return type:

Any

get_weights()

Get the tile weights.

Returns:

a tuple where the first item is the [out_size, in_size] weight matrix; and the second item is either the [out_size] bias vector or None if the tile is set not to use bias.

Return type:

Tensor

get_x_size()

Returns input size of tile

Return type:

int

set_config(rpu_config)

Updated the configuration to allow on-the-fly changes.

Parameters:

rpu_config (CustomRPUConfig) – configuration to use in the next forward passes.

Return type:

None

set_learning_rate(learning_rate)

Set the learning rate of the tile.

No-op for tiles that do not need a learning rate.

Parameters:

• rate (learning) – learning rate to set

• learning_rate (float | None)

Return type:

None

set_weights(weight)

Set the tile weights.

Parameters:

weight (Tensor) – [out_size, in_size] weight matrix.

Return type:

None

set_weights_uniform_random(bmin, bmax)

Sets the weights to uniform random numbers.

Parameters:

• bmin (float) – min value

• bmax (float) – max value

Raises:

TileError – in case bmin >= bmax

Return type:

None

update(x_input, d_input, bias=False, in_trans=False, out_trans=False, non_blocking=False)

Update with gradient noise.

Note

Ignores additional arguments

Raises:

TileError – in case transposed input / output or bias is requested

Parameters:

• x_input (Tensor)

• d_input (Tensor)

• bias (bool)

• in_trans (bool)

• out_trans (bool)

• non_blocking (bool)

Return type:

Tensor

class aihwkit.simulator.tiles.custom.CustomTile(out_size, in_size, rpu_config=None, bias=False, in_trans=False, out_trans=False)

Bases: TileModule, TileWithPeriphery, SimulatorTileWrapper

Custom tile based on TileWithPeriphery.

Implements a tile with periphery for analog training without the RPUCuda engine.

Raises:

• TileError – in case in-trans / out-trans is used (not supported)

• AnalogBiasConfigError – if analog bias is enabled in the RPUConfig

Parameters:

• out_size (int)

• in_size (int)

• rpu_config (CustomRPUConfig | None)

• bias (bool)

• in_trans (bool)

• out_trans (bool)

forward(x_input, tensor_view=None)

Torch forward function that calls the analog context forward

Parameters:

• x_input (Tensor)

• tensor_view (Tuple | None)

Return type:

Tensor

post_update_step()

Operators that need to be called once per mini-batch.

Note

This function is called by the analog optimizer.

Caution

If no analog optimizer is used, the post update steps will not be performed.

Return type:

None

supports_indexed = False

class aihwkit.simulator.tiles.custom.CustomUpdateParameters(gradient_noise=0.0)

Bases: _PrintableMixin

Custom parameters for the update

Parameters:

gradient_noise (float)

gradient_noise: float = 0.0

Adds Gaussian noise (with this std) to the weight gradient.