Model mapping

Move from a model file to a virtual AKD1500 target, then to physical hardware.

Map without hardware first

A saved Akida model can be loaded from its .fbz path. Using the virtual AKD1500 target lets you test whether the model maps to that hardware class before involving the PCIe stack.

import akida

model = akida.Model('model.fbz')
model.map(akida.AKD1500())
print(model.summary())

Map to a detected device

import akida

devices = akida.devices()
if not devices:
    raise RuntimeError('No Akida hardware detected')

model = akida.Model('model.fbz')
model.map(devices[0])
print(model.summary())

Read the model summary

The current runtime documentation reports sequence count, NP allocation and external-memory usage after mapping. More than one hardware sequence means the runtime must chain-load sequences during inference.

Use the CLI for a quick run

akida run -m model.fbz

The CLI can map a model to an available device, run inference and print mapping/metrics information. For repeatable lab work we still prefer a small Python script committed with the experiment.