Examining 3x3 convolutional filters that are common in pretrained object recognition networks (by first rounding each filter weight to the nearest 0.025).
Networks used (from Pytorch): resnet18, alexnet, squeezenet, vgg16, inception
Information about the top 100 filters in filters.txt
- Includes weights, counts, and networks where the filter was used.
Note: only input convs included as of now
- Include description of filters (interpretation of what they do).
- Aggregate multiples of the same filters together.
- Frequency by what layers they occur in (bottom 1/3, middle 1/3, top 1/3)
- Include output filters