Modelling Implementation

[guillaume-byte]

Model Weights editing (a.k.a) modelling is partially implemented and available; however:

• Opaque interfaces and API
• Robustness and coverage (currently supports Torch and ONNX but has a limited scope)
• Structural issues (this are symptomatic rather than root cause):
  • Refactor index mapping in neuron operations #6
  • Refactor indices operation functions + - into one function, operating on indices #7
  • Fix hardcoded code for norms & transposed conv & dependency flag & shape conv. #8

A refinement or revamp is necessary in order to enable this new dimension of experimentation:

• Implement API for modelling of the model graph: getters & setter
• getter: model graph information rendering, layer structure and neurons
• setter: or modifiers would represent the actions (freeze, reset, add, prune, perturb)
• Low overhead technique to monitor and hook into the computation graph
• weights insights: trigger rates, triggered sample ids, relative/absolute weights changes; Explainable methods like CAPTUM
• weights lineage tracing
• Weights editing flows examples that we want to support immediately:
• train or load pre-trained, edit data sets, freeze layers, resume training
• train or load pre-trained, freeze layers, add neurons, resume training
• train or load pre-trained, compare 2 samples side by side w.r.t to activation in the model e.g: l0 -> 0.1%, l1 -> 0.1%, l2-> 4.5%.. means l2 is the layers that discriminates the 2 samples we are comparing