brando90 / ultimate-anatome

Ἀνατομή is a PyTorch library to analyze representation of neural networks
MIT License
11 stars 1 forks source link

anatome

Ἀνατομή is a PyTorch library to analyze internal representation of neural networks

This project is under active development and the codebase is subject to change.

Note the name of the pypi package is ultimate-anatome but the python import is done with the original name import anatome. Test via:

python -c "import anatome.my_utils as my_utils;my_utils.hello()"
python -c "import anatome.my_utils as my_utils;my_utils.my_anatome_test()"

Credit to original library: https://github.com/moskomule/anatome and consider citing both of us! Citation at the end of readme.

Installation

If you are going to use a gpu the do this first before continuing (or check the offical website: https://pytorch.org/get-started/locally/):

pip3 install torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html

Otherwise, just doing the follwoing should work.

pip install ultimate-anatome

Manual installation [for Development]

To use code first get the code from this repo (e.g. fork it on github):

git clone git@github.com:brando90/ultimate-anatome.git

Then install it in development mode in your python env with python >=3.9. E.g. create your env with conda:

conda create -n ultimate_anatome_env python=3.9
conda activate ultimate_anatome_env

Then install code in edibable mode and all it's depedencies with pip in the currently activated conda environment:

pip install -e ~/ultimate-anatome

Available Tools

Representation Similarity

To measure the similarity of learned representation, anatome.SimilarityHook is a useful tool. Currently, the following methods are implemented.

import torch
from torchvision.models import resnet18
from anatome import DistanceHook
from anatome.my_utils import remove_hook

model = resnet18()
hook1 = DistanceHook(model, "layer3.0.conv1")
hook2 = DistanceHook(model, "layer3.0.conv2")
model.eval()
with torch.no_grad():
    model(torch.randn(128, 3, 224, 224))
# downsampling to (size, size) may be helpful
hook1.distance(hook2, size=8)
hook1.clear()
hook2.clear()
remove_hook(model, hook1)
remove_hook(model, hook2)

or to test do:

from anatome.my_utils import my_anatome_test
my_anatome_test()

Loss Landscape Visualization

from anatome import landscape2d

x, y, z = landscape2d(resnet18(),
                      data,
                      F.cross_entropy,
                      x_range=(-1, 1),
                      y_range=(-1, 1),
                      step_size=0.1)
imshow(z)

Fourier Analysis

from anatome import fourier_map

map = fourier_map(resnet18(),
                  data,
                  F.cross_entropy,
                  norm=4)
imshow(map)

Citation

If you use this implementation in your research, please conser citing my version of anatome:

@software{miranda2021ultimate_anatome,
    author={Brando Miranda},
    title={Ultimate Anatome, the ultimate PyTorch library to analyze internal representation of neural networks},
    url={https://github.com/brando90/ultimate-anatome},
    year={2021}
}

A permanent link lives here: https://www.ideals.illinois.edu/handle/2142/112796

Also consider citing the original version where we took good inspiration:

@software{hataya2020anatome,
    author={Ryuichiro Hataya},
    title={anatome, a PyTorch library to analyze internal representation of neural networks},
    url={https://github.com/moskomule/anatome},
    year={2020}
}

Related libraries: