What is the difference between using InferenceRPUConfig vs any presets for inference?

[matifali]

What is the difference between using InferenceRPUConfig vs any presets to be used for inference? e.g.

# 1st case
RPU_CONFIG = InferenceRPUConfig()
RPU_CONFIG.backward.bound_management = BoundManagementType.NONE
RPU_CONFIG.forward.out_res = 2 / (2**8 - 1)
RPU_CONFIG.forward.inp_res = 2 / (2**8 - 1)
RPU_CONFIG.forward.w_noise_type = WeightNoiseType.ADDITIVE_CONSTANT
RPU_CONFIG.forward.w_noise = W_NOISE
RPU_CONFIG.forward.out_noise = O_NOISE
RPU_CONFIG.noise_model = PCMLikeNoiseModel(g_max=G_MAX)
# 2nd case
RPU_CONFIG = ReRamSBPreset()
# 3rd case
RPU_CONFIG = PCMPreset()

I have pre-trained models that I am converting to analog using
I am getting a way better accuracy if I use InferenceRPUConfig with PCMLikeNoiseModel
but very bad performance with ReRamSBPreset and PCMPreset

What is the difference between using an InferenceRPUConfig with PCMLikeNoiseModel and a PCMPreset? How does the drift_analog_weights for both cases?

[maljoras]

The InferenceRPUConfig is used for accuracy estimation for a situation where the resistive crossbar arrays are only used in the inference pass (training is done in FP and only inference is using the resistive devices). The other presets are used for in-memory training with forward / backward and update in "analog" and updating using stochastic pulse trains in a fully parallel manner. This is typically very challenging and might need special optimizers such as MixedPrecision / or TikiTaka for which we also have presets.

For inference only chips, the expected programming noise and drift can be set by the noise_model setting.

[matifali]

Thank you, @maljoras. Can we set d2d and c2c variance for the inference-only cases?
Also, is a noise model for ReRamSBPreset to be used only in inference?

[maljoras]

ReRamSBPreset determines the pulse response of the conductance. It is typically only used for training. However, of course you could try to explicitly program the inference weights using the update and a read-write-verify scheme.

Device-to-device variation for inference is implicitly controlled by setting the programming noise polynomials coefficients (noise_model.prog_coeff). See here for more details of the noise model.