aai-institute/veriflow

Normalizing flows for neuro-symbolic AI

veriflow's Issues

Improve REAMDE.md

• Implementation based on pyro
• Installation
• Getting started
• Layers

Missing import in explib/base.py

deepcopy import missing

Create Build Docs Pipeline

Fix pipelines

Make custom solve_triangluar optional

The custom implementation of the solve_triangular function is only necessary if onnx export is needed but is a performance bottleneck. We should make it optional to use the implementation in torch.linalg.

• Extend solve_triangular with another argument, e.g. backend.
  • Depending on the value of the new argument, the function should either use the building implementation or call the torch implementation.
  • Set the default backend to torch.

Improve Testing

VeriFlow

• linalg
• networks
• transforms
• flows
  • Onnx export

ExpLib

• config_parser
• base
• hyperopt
• datasets

Investigate different sampling behaviour of the norm distribution when the cpu is used as device.

When sampling the norm distribution in sample we need to unsqueeze the dimensions when the device is set to gpu. However if the device is set to cpu this unsqueeze needs to be removed otherwise we will get the following error:

r = r.repeat(*[1 for _ in sample_shape], self.dim)
RuntimeError: Number of dimensions of repeat dims can not be smaller than number of dimensions of tensor

ONNX export

Implement convenience function for onnx support

Rename Library

Rename to veriflow

Refactor documentation

Problem

• Some of the code documentation is outdated/incomplete/inconsistent format

Todos

VeriFlow

• linalg
• networks
• transforms
• flows

ExpLib

• config_parser
• base
• hyperopt
• datasets

Low GPU utilization CUDA

GPU utilization on coda machines is low (5-15%).
There is potentially a bottleneck somewhere

• Data handling?
• All operations performed on gpu?
• ???

HyperOptExperiment: improve evaluation

• Final evaluation of best model on test set.
• Options for visualizing intermediate results, e.g. plotting samples.
• Improve final reporting

Onnx export with LU transform not supported

Description

Calling to_onnx on a flow with LU layers raises a NotImplementedError. Reason is that the function linalg_inv is not implemented in onnx

Suggestion

A possible workaround could be to export a surrogate model with precomputed inverse matrix instead.

HyperoptExperiment with NiceFlow on mps

Problem

HyperoptExperiment fails with NiceFlow on mps devices.

Analysis

• Verified for permutation == "half"
• Error: Permutations seem to be not on device?!?
• works with NiceMaskedConvFlow

Refactor Naming conventions

We should unify the class names, especially for the flow classes. Current names carry a lot of legacy baggage. Ideally, names are in line with package names.

Outdated experiment configs

Several experiment configs are outdated and cannot be run currently.

Implement Additive Autoregressive Layer

Implement a layer that replaces additive coupling with additive autoregression.

Improve Data Preprocessing

• Simple configuration in yaml
  • Downscaling
  • Reshaping
  • Dequantization
• Remove any hardcoded downscaling

Low GPU usage when training without Softtraining

Apparently, DenseNN of pyro is slow???

Implement Test Pipeline

Basic MNIST Experiment with training once with all flow architectures

Bug in Sampling(?) method

There seems to be an inconsistency between manually sampling from the base distribution and pushing the result through the flow and calling the sampling method.

• If one computes the norm of a sample that was pushed backwards through the flow. The obtained norms don't seem to follow the specified norm distribution of the base distribution.
• Sampling from the base distribution and pushing back and forth through the flow manually, however, gives the expected result.

Implement LU-layers

Implement LU layers as non url preserving layers

Remove hardcoded base distributions in HyperOptExperiment

Proper list handling of config parser

Proper handling of lists with atomic types during recursion