MSWC FSCIL task

[Maxtimer97]

Added cleaned CNN baseline code for MSWC FSCIL task based on a monolingual english version of the dataset. This integrates pre-training and incremental learning over the 10 session with a simple weight consolidation technique.

The implementation also still requires the torch-mate libary and actually employs one of its files that has been modified . Something to discuss with @V0XNIHILI

The cl_utils file that implement simple Continual learning tricks comes from the following github repo github https://github.com/vlomonaco/ar1-pytorch/tree/master and has copyright indications. Is there something to do about that?

Still to be done:

• Implement multilingual version of dataset
• Adapt/optimize baseline
• Add SNN version (under development)

[jasonlyik]

The cl_utils file that implement simple Continual learning tricks comes from the following github repo github https://github.com/vlomonaco/ar1-pytorch/tree/master and has copyright indications. Is there something to do about that?

Re: this point, the copyright license is CC BY 4.0, which shortly states that the code is free use, as long as it is attributed.

Therefore I believe all we need to do is update the current header in the file with a link to the CC BY 4.0 license, and notes if anything was changed within the file.

[Maxtimer97]

Updated CNN baseline with multilingual version of the dataset. New benchmark results on Google doc

Things to figure out for the multilingual set:

• Provide a hosted version of the dataset with csv splits
• OR provide word + sample selection code to create splits from whole dataset

Can we manage something for the first version @jasonlyik ?

[jasonlyik]

List of short things to do before integrating this branch into dev:

• [x] Add comment for attributing the cl_utils files
• [x] MSWC should make data format consistent with (batch, timesteps, channels)
• [x] Merge dev and validate that all other benchmarks run, since we make changes to Benchmark interface