Results aren't promising, but they're also fishy. From what I can tell all layers have <0.15 robustness, with layer 0 slightly less robust... the heatmaps we have now for Baseline suggest resetting to epoch 0 should give us ~0.4 at least.
If you could give it a go and see what you think that would help. Paste this in colab if you have a few minutes (models are pre-trained):
from winnery_intersection import WinneryIntersection
wi = WinneryIntersection(root_dir='/content/drive/My Drive/globi/',
resource_load_dir='WinneryIntersection/models_9.8.2019',
prune_threshold=0.07)
wi.go()
Implements #85, based on PR #87
Results aren't promising, but they're also fishy. From what I can tell all layers have <0.15 robustness, with layer 0 slightly less robust... the heatmaps we have now for Baseline suggest resetting to epoch 0 should give us ~0.4 at least.
If you could give it a go and see what you think that would help. Paste this in colab if you have a few minutes (models are pre-trained):