Problem in reproducing attention analysis from the paper "What Do They Capture? - A Structural Analysis of Pre-Trained Language Models for Source Code"

[dfighter1312]

Hi,

First of all, thank you for such detailed writing for discussion on pre-trained models for source code.

I am currently trying to reproduce the result, but in compute_edge_features.py, line 133, you are referring to a path ../data/code_new/code_contact_map/noneighbor/train.json, which I could not find anywhere.

I did try to change the path to the train.ast file provided in the Python AST dataset, but another error is raised.

Layers: 12
Heads: 12
Loading dataset
100% 5000/5000 [00:00<00:00, 1458178.28it/s]
  0% 0/5000 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "compute_edge_features.py", line 155, in <module>
    min_attn=min_attn)
  File "compute_edge_features.py", line 64, in compute_mean_attention
    feature_map=item['feature_map']
KeyError: 'feature_map'

I hope you can give me an instruction to resolve the problem.

Many thanks!