aihwkit.utils.analog_info module¶

Analog Information utility.

This module prints relevant information about the model and its analog execution.

class aihwkit.utils.analog_info.AnalogInfo(model, input_size=None, rpu_config=None)[source]¶

Bases: object

Class for computing and storing results of the analog summary.

Parameters

model (torch.nn.modules.module.Module) –
input_size (Any) –
rpu_config (Optional[Union[FloatingPointRPUConfig, SingleRPUConfig, UnitCellRPUConfig, InferenceRPUConfig, DigitalRankUpdateRPUConfig]]) –

calculate_num_analog()[source]¶

Calculate the number of analog layers.

Return type: int

calculate_num_tiles()[source]¶

Calculate the total number of tiles needed by the model.

Return type: int

create_layer_summary()[source]¶

Create the layer summary list.

This list contains LayerInfo elements that corresponds to each layer of the model.

Return type: List[aihwkit.utils.analog_info.LayerInfo]

register_hooks_recursively(module, hook)[source]¶

Hooks the function into all layers with no children.

Parameters

module (torch.nn.modules.module.Module) –
hook (Any) –

Return type

None

class aihwkit.utils.analog_info.LayerInfo(module, rpu_config=None, input_size=None, output_size=None)[source]¶

Bases: object

Class for storing layer statistics and information.

Parameters

module (torch.nn.modules.module.Module) –
rpu_config (Optional[Union[FloatingPointRPUConfig, SingleRPUConfig, UnitCellRPUConfig, InferenceRPUConfig, DigitalRankUpdateRPUConfig]]) –
input_size (Any) –
output_size (Any) –

calculate_reuse_factor()[source]¶

Compute the reuse factor.

The reuse factor is the number of vector matrix multiplication a layer computes.

Return type: None

input_size: Any¶

isanalog: bool¶

kernel_size: Any¶

layer_summary_dict()[source]¶

Return a dictionary with all layer’s information.

Return type: dict

module: torch.nn.modules.module.Module¶

name: str¶

num_tiles: int¶

output_size: Any¶

reuse_factor: int¶

set_kernel_size()[source]¶

Set kernel size attribute.

Return type: None

set_tiles_info()[source]¶

Create TileInfo objects for each tile of the layer.

Return type: List[aihwkit.utils.analog_info.TileInfo]

tiles_info: List[aihwkit.utils.analog_info.TileInfo]¶

tiles_summary_dict()[source]¶

Return a dictionary with all tiles information.

Return type: dict

class aihwkit.utils.analog_info.TileInfo(tile, is_mapped)[source]¶

Bases: object

Class for storing tile statistics and information.

Parameters

tile (aihwkit.simulator.tiles.base.BaseTile) –
is_mapped (bool) –

log_in_size: Any¶

log_out_size: Any¶

phy_in_size: Any¶

phy_out_size: Any¶

tile_summary_dict()[source]¶

Return a dictionary with the tile info.

Return type: dict

utilization: float¶

aihwkit.utils.analog_info.analog_summary(model, input_size=None, rpu_config=None)[source]¶

Summarize the given PyTorch model.

Summarized information includes:

Layer names,

input/output shapes,

kernel shape,

# of digital parameters,

# of analog parameters,

# of analog tiles

reuse factor

Parameters

model (torch.nn.modules.module.Module) – PyTorch model to run on the analog platform.
input_size (Optional[Any]) – required to run a forward pass of the model.
rpu_config (Optional[Union[FloatingPointRPUConfig, SingleRPUConfig, UnitCellRPUConfig, InferenceRPUConfig, DigitalRankUpdateRPUConfig]]) – resistive processing unit configuration.

Returns

AnalogInfo Object.

Return type

aihwkit.utils.analog_info.AnalogInfo