From c9b35c815d1fa8026b336bd9455ecd031f72b76f Mon Sep 17 00:00:00 2001 From: Jakub Kaczmarzyk Date: Fri, 6 Sep 2024 13:40:04 +0200 Subject: [PATCH] clarify relationship between ablated patches and outputs --- README.md | 12 ++++-------- 1 file changed, 4 insertions(+), 8 deletions(-) diff --git a/README.md b/README.md index 4538fb9..1825b90 100644 --- a/README.md +++ b/README.md @@ -221,9 +221,10 @@ def model_probs_fn(features): return probs -# Find the 1% highest effect patches. These are the patches that, -# when removed, drop the probability of metastasis the most. -# The `results` variable is a dictionary with.... results of the search! +# Find the 1% highest effect patches. These are the patches that, when removed, drop the probability +# of metastasis the most. The `results` variable is a dictionary with.... results of the search! +# The model outputs in `results["model_outputs"]` correspond to the results after removing the patches +# in `results["ablated_patches"][:k]`. num_rounds = math.ceil(len(features) * 0.01) results = hippo.greedy_search( features=features, @@ -235,10 +236,6 @@ results = hippo.greedy_search( optimizer=hippo.minimize, ) -# The values of `results["model_outputs"]` and `results["ablated_patches"]` are numpy arrays -# with the same length. The k-th index in `results["model_outputs"]` corresponds to the model -# outputs after the patches `results["ablated_patches"][k+1]` have been removed. - # Now we can test the effect of removing the 1% highest effect patches. patches_not_ablated = np.setdiff1d(np.arange(len(features)), results["ablated_patches"]) with torch.inference_mode(): @@ -262,7 +259,6 @@ plt.ylabel("Probability of metastasis") ``` - # Cite ```bibtex