From b41755a27983d260b0077acdb57aabcf757489ec Mon Sep 17 00:00:00 2001
From: Martin Lange <32666496+lange-martin@users.noreply.github.com>
Date: Thu, 24 Mar 2022 10:58:12 +0100
Subject: [PATCH] Switch harmful and helpful lists

The most helpful samples are those with the "most positive -I_{up, loss}" (page 6 Koh and Liang paper). Since influences contains the I_{up, loss} for each training sample, the most helpful samples should be those indices in ascending order.
---
 pytorch_influence_functions/calc_influence_function.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/pytorch_influence_functions/calc_influence_function.py b/pytorch_influence_functions/calc_influence_function.py
index 8861cd4..5abad63 100644
--- a/pytorch_influence_functions/calc_influence_function.py
+++ b/pytorch_influence_functions/calc_influence_function.py
@@ -343,8 +343,8 @@ def calc_influence_single(model, train_loader, test_loader, test_id_num, gpu,
         influences.append(tmp_influence)
         display_progress("Calc. influence function: ", i, train_dataset_size)
 
-    harmful = np.argsort(influences)
-    helpful = harmful[::-1]
+    helpful = np.argsort(influences)
+    harmful = helpful[::-1]
 
     return influences, harmful.tolist(), helpful.tolist(), test_id_num