From 8e7769610e9c718f0574c341590ab6161e246853 Mon Sep 17 00:00:00 2001 From: Sang Choe Date: Mon, 20 May 2024 14:10:09 -0400 Subject: [PATCH] Update README.md --- examples/mnist/README.md | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-) diff --git a/examples/mnist/README.md b/examples/mnist/README.md index 2a9bed8a..c86c40f8 100644 --- a/examples/mnist/README.md +++ b/examples/mnist/README.md @@ -1,11 +1,20 @@ # MNIST ### Step 1: Train +Since influence functions are typically computed at the final model +weight, we first train the model (MLP) on the (small) MNIST dataset +before performinig influence analyses. + ```bash python train.py ``` -### Step 2: AnaLog +### Step 2: Log extraction & Influence analysis +With the trained model, we first extract and save logs (e.g. +Hessian, gradeint) to disk, and use it to compute influence +scores. Users can specify the gradient projection strategy +(e.g. LoGra) and the Hessian computation strategy. + ```bash -python compute_influences.py +python compute_influences.py --lora none --hessian raw --save grad ```