issues
search
chr5tphr
/
zennit
Zennit is a high-level framework in Python using PyTorch for explaining/exploring neural networks using attribution methods like LRP.
Other
188
stars
32
forks
source link
Tests: Added tests for Attributors
#119
Closed
chr5tphr
closed
2 years ago
chr5tphr
commented
2 years ago
Tests: Added tests for Attributors
Fixtures:
add fixture for common grad_output functions, with an extra function one_hot_max
add fixture for a simple model, which uses any of the linear module fixture, followed by a relu, followed by a dense layer with 4 outputs
add a fixture for the simple model's gradient
add a fixture for the simple model's output
Helpers:
add the IdentityLogger helper in test_attribution.py, which is a torch.nn.Module which clones, logs and returns the input
Tests:
add various tests for all attributors, which currently are Gradient, IntegratedGradients, SmoothGrad and Occlusion
Tests: Move composite fixtures to conftest
move composite fixtures to conftest such that they can be used by other tests, and to conform to the 'fixtures in conftest.py' structure
rename fixture 'composite_model' to 'model_vision'
add alexnet to the model_vision fixture
Tests: Added tests for Attributors
Fixtures:
Helpers:
Tests:
Tests: Move composite fixtures to conftest