[Good First Issue][NNCF]: Fixing NNCFGraph export for visualization in Netron

[andrey-churkin]

Context

Neural Network Compression Framework (NNCF) uses an instance of NNCFGraph class to represent a framework-specific deep learning model. An instance of NNCFGraph class is used in almost every algorithm and serves as the primary means to obtain information about deep learning model.

It is very beneficial to have a way to visualize such graph. One possible way to visualize it is by using a .dot file.
Please see visualize_graph() method:

nncf/nncf/common/graph/graph.py

Lines 565 to 567 in 481ce9d

	def visualize_graph(self, path: str):
	out_graph = self._get_graph_for_visualization()
	write_dot_graph(out_graph, path)

The main disadvantage of this approach is that it is very difficult to analyze a .dot file for very large models. It is better to use Netron. We already have a method that allows us to export an instance of the NNCFGraph class to a format supported by Netron. Actually, we export NNCFGraph to the OpenVINO IR format. Please see OpenVINO IR format specification. We save only the topology (.XML file).

Please follow the following code snippet to understand how to save NNCFGraph for visualization in Netron

import openvino as ov
from nncf.common.factory import NNCFGraphFactory
from nncf.experimental.common.graph.netron import save_for_netron

ov_model = ov.Core().read_model("<MODEL_PATH>")
graph = NNCFGraphFactory.create(ov_model)
save_for_netron(graph, "graph.xml")

What needs to be done?

• Currently, it is not possible to open the resulting graph.xml in Netron. The main task is to resolve this problem by fixing the export. Netron returns the following error: Error loading OpenVINO model. Unsupported precision 'undefined'
• Add docstrings for all methods (get_graph_desc(), save_for_netron()) and classes (PortDesc, NodeDesc, EdgeDesc) inside nncf.experimental.common.graph.netron module
• Add tests for the following methods
  • get_graph_desc()
  • PortDesc.as_xml_element()
  • NodeDesc.as_xml_element()
  • EdgeDesc.as_xml_element()

Example Pull Requests

N/A

Resources

• Contribution guide - start here!

Contact points

@andrey-churkin @alexsu52

Ticket

N/A

[DaniAffCH]

.take

[github-actions]

Thank you for looking into this issue! Please let us know if you have any questions or require any help.

[DaniAffCH]

Hi @andrey-churkin,
I dug deeper into the issue and found a potential solution, but I want to discuss the details with you.

The dtype should be retrieved from the instances of the class NNCFGraphEdge. However, the class contains only the nncf dtype that could be either FLOAT or INTEGER. This is because the original ov dtype is squished into one of these two classes

https://github.com/openvinotoolkit/nncf/blob/481ce9d9624c4b699878aa337d2e9b78dc25a517/nncf/openvino/graph/nncf_graph_builder.py#L32C1-L63C1

The Netron OpenVino frontend supports only the native OpenVino datatypes

https://github.com/lutzroeder/netron/blob/49e3a04a3210cf655e0abb04016e36b2a3bfcebf/source/openvino.js#L748C1-L772C111

So we can choose to save the original ov dtype into the class NNCFGraphEdge to display the correct layer dtype in Netron.

An alternative solution would be to map the nncf dtype FLOAT into f32 and INTEGER into i32.

I think the first solution would be the correct one as it will allow to display of the correct data type in Netron, but this requires modifying all the NNCFGraphEdge instantiations. Or maybe I could set the parameter as optional.

Let me know your thoughts on this, so I can proceed with the implementation

[andrey-churkin]

@DaniAffCH Hi
I think we can map Dtype.FLOAT to "fp32" and DType.INTEGER to "i32". It allows us to avoid significant validation.

[DaniAffCH]

I've just submitted the PR #2567 solving this issue

[andrey-churkin]

@DaniAffCH Thanks for the contribution!