NETX library fails to convert simple slayer classification network to an equivalent lava network

[rhendz]

Objective of issue: I trained an IRIS network using the slayer library to near 100% accuracy and converted it over to a lava process using the netx library; however, the accuracy is not the same (about 33%). I need documentation or examples on how to convert a network trained via slayer over to a lava process with minimal accuracy loss if possible.

Lava DL version:

• 0.3.0 (feature release)
• 0.2.1 (bug fixes)
• 0.2.0 (current version)
• 0.1.2

Lava version:

• 0.4.0 (feature release)
• 0.3.1 (bug fixes)
• 0.3.0 (current version)
• 0.2.0
• 0.1.2

I'm submitting a ...

• bug report
• feature request
• documentation request

Current behavior:

• Slayer trained network is not converting over (netx) to an equivalent network in lava.

Expected behavior:

• Each network (slayer and lava) should produce the same inference results.

Steps to reproduce:

# Clone and setup my custom repo
git clone https://github.com/rhendz/lava-apps.git
cd lava-apps
sh setup.sh

# Use lava-dl virtual environment
source lava-dl/.venv/bin/activate

# Setup notebook and kernel
pip install notebook
pip install ipykernel

python -m ipykernel install --user --name=lavadl

# Launch notebook
jupyter notebook

# Notes:
# Use lava-dl-netx-iris to test inference of original network and generate/run an inference test on the netx network
# Use lava-dl-slayer-iris to train IRIS network
# Change kernel to lavadl

Other information:
My objective is to eventually bring in other networks via netx library and be able to produce good inference results in lava. However, I need a good baseline via the slayer library first.

[nskat]

Hi @rhendz, did you manage to find a solution to this? I've found the same issue when loading networks using netx with the latest lava/lava-dl versions.
Interestingly, if I write the net's Process and ProcessModel by hand, it works perfectly fine -- this may help the dev team figure this out.

[Geoforger]

Hi, was this issue ever resolved? I'm having similar issues using lava 0.7.0 and lava-dl 0.3.3. My lava-dl trained classifier is achieving very high accuracies in training but after conversion with netx is seemingly making random decisions.