Harish-JHR/DiffuseSegWeights

DiffuseSeg Weights

Used for reproducing results of "Label-Efficient Semantic Segmentation with Diffusion Models" (ICLR 2022)

This repository contains weights and features extracted from a Denoising Diffusion Probabilistic Model (DDPM) for segmentation tasks. DDPM Weights can be found at "https://huggingface.co/Harish-JHR/DDPM_CelebAHQ64". Following the approach in the referenced paper, we extract pixel-level features from the UpBlock layers of the DDPM and train lightweight segmentation heads on them.

Contents

• ddpm_pixel_features_train.pt: Pixel-level feature vectors from DDPM.
• ddpm_pixel_labels_train.pt: Corresponding integer pixel labels for training.
• mlp_X_best.pt: Trained MLP segmentation heads (10 in total).

Each MLP corresponds to a segmentation head trained on different layers/features.

Usage

import torch

features = torch.load("ddpm_pixel_features_train.pt")
labels = torch.load("ddpm_pixel_labels_train.pt")
mlp1 = torch.load("mlp_1_best.pt")