Here, we include the PyTorch implementation for the work submitted to ICML 2023 (https://openreview.net/pdf?id=5d53OeOZHE)
Step 1: Identifying the Sparse Subspace
We use the PASCAL VOC natural image dataset (1472 RGB natural images) to pre-train a shallow U-Net with 64 channels and five scales for 2000 epochs, while retaining 5000 checkpoints, equally-spaced throughout the optimisation trajectory. We downsize the pre-training image size to
Step 2: Natural Images Restoration (Denoising and Deblurring)
We report two additional experimental figures on a denosing and deblurring tasks, where the standard DIP and Sub-DIP (NGD) are compared on three widely used RGB images, namely Airplane F16, House, and Lena. Note that for both restoration task, the image resolution used is
For denoising, at test time 10% white noise is removed from the images below.
For deblurring, at test time a Gaussian blur with standard deviation of 1.6 pixels is used along with 5% Gaussian noise.