This repo contains the sample code of our proposed Unleashing Mask (UM)
and its variant Unleashing Mask Adopt Pruning (UMAP)
to adjust the given well-trained model for OOD detection in our paper: Unleashing Mask: Explore the Intrinsic Out-of-Distribution Detection Capability (ICML 2023).
Figure. Framework overview of UM.
Our work reveal an intermediate training stage with better out-of-distribution (OOD) discriminative capability of the well pre-trained model (on classifying the in-distribution (ID) data). By exploration from various perspective, we show the data-level characterisitics of ID data is also important for the OOD detection performance of the model. We propose Unleashing Mask to restore it of the given well-trained model for OOD detection, by fine-tuning with the estimated loss contraint to forget those relatively atypical ID samples.
- Set up a virtualenv with python 3.7.4. You can use pyvenv or conda for this.
- Run
pip install -r requirements.txt
to get requirements - Create a data directory as a base for all datasets. For example, if your base directory is
/mnt/datasets
then imagenet would be located at/mnt/datasets/ImageNet
and CIFAR-10 would be located at/mnt/datasets/cifar10
UM is quite a easy-to-adopt method to use in your own pipeline for enhancing the OOD discriminative capability. The key point is to add the loss constraint to the CrossEntropyLoss:
class CustomLoss(nn.Module):
def __init__(self, criterion=None):
super(CustomLoss, self).__init__()
self.criterion = criterion
self.CrossEntropyLoss = nn.CrossEntropyLoss()
def forward(self, outputs, labels):
loss = args.beta * self.CrossEntropyLoss(outputs, labels)
if self.criterion != "CrossEntropyLoss":
# args.UM is the estimated loss constriant
loss = (loss - args.UM).abs() + args.UM
return loss
To use UM/UMAP in your pipeline:
...
from utils.custom_loss import CustomLoss
...
criterion = CustomLoss("UM")
...
for input, target in data:
...
loss = criterion(model(input), target) # use this to calculate UM loss
loss.backward()
...
Or apply UM in a more straightforward way:
...
import torch.nn as nn
...
criterion = nn.CrossEntropyLoss()
...
for input, target in data:
...
loss = criterion(model(input), target)
loss = (loss - args.UM).abs() + args.UM # args.UM is the estimated loss constriant
loss.backward()
...
We use config files located in the configs/
folder to organize our experiments. The basic setup for any experiment is:
python main.py --config <path/to/config> <override-args>
Common example override-args
include --multigpu=<gpu-ids seperated by commas, no spaces>
to run on GPUs. Run python main --help
for more details.
We provide pretrained DenseNet-101 models in runs/pretrained_models
, which are trained on CIFAR-10/CIFAR-100 respectively. Note that for each setting, we provide the best (middle-stage) model and the last (final-stage) model. The pretrained model are named as densenet_<cifar10/cifar100>_<best/last>.pth
.
Before you begin experiment, please arrange your dataset directory as follows:
<path/to/data-dir>:
|-cifar10
|-cifar-10-batches-py
|-cifar100
|-cifar-100-python
|-dtd
|-ImageNet
|-iNaturalist
|-iSUN
|-LSUN
|-LSUN-resize
|-MNIST
|-Places
|-SUN
|-svhn
|-tinyimagenet
To estimate the loss constraint for UM/UMAP. Note that you can either estimate the loss constraint by estimate_loss.py
or manual tuning.
python estimate_loss.py --config configs/estimate_loss/estimate_loss_cifar10.yaml \
--data <path/to/data-dir>
To experiment with the post-hoc OOD detection methods. Use flag--msp
, --odin
, --energy
, --mahalanobis
to control the scoring functions respectively. If more than one function is chosen, OOD performance will be measured under these functions respectively. To evaluate OOD performance, you can either use flag --final
to evaluate right after training, or use flag --evaluate
to evaluate a loaded trained model.
python main.py --config configs/smallscale/densenet_cifar10.yaml \
--multigpu 0 \
--name cifar10_UM_post_hoc \
--data <path/to/data-dir> \
--UM <estimated_loss>\
--energy
python main.py --config configs/smallscale/densenet_cifar10.yaml
To experiment with the Outlier Exposure (OE) OOD detection methods. Use flag --oe-ood-method
to control the OE-based methods.
python main_OE.py --config configs/OE/oe-baseline.yaml \
--multigpu 0 \
--name cifar10_UM_oe \
--data <path/to/data-dir> \
--oe-ood-method <choose from [oe, poem, enregy, doe]> \
--UM <estimated_loss>
--energy
When your experiment is done, your experiment base directory will automatically be written to runs/<config-name>/prune-rate=<prune-rate>/<experiment-name>
with settings.txt
, <ID-Dataset>_log.txt
, checkpoints/
and logs/
subdirectories. If your experiment happens to match a previously created experiment base directory then an integer increment will be added to the filepath (eg. /0
, /1
, etc.). Checkpoints by default will have the first, best, and last models. To change this behavior, use the --save-every
flag.
Method | AUROC |
AUPR |
FPR95 |
ID-ACC |
w./w.o. |
|
---|---|---|---|---|---|---|
CIFAR-10 | Energy | 92.07 (0.22) | 92.72 (0.39) | 42.69 (1.31) | 94.01 (0.08) | |
CIFAR-10 | Energy+UM | 93.73 (0.36) | 94.27 (0.60) | 33.29 (1.70) | 92.80 (0.47) | |
CIFAR-10 | OE | 97.07 (0.01) | 97.31 (0.05) | 13.80 (0.28) | 92.59 (0.32) | |
CIFAR-10 | OE+UM | 97.60 (0.03) | 97.87 (0.02) | 11.22 (0.16) | 93.66 (0.12) |
Python 3.7.4, CUDA Version 10.1 (also works with 9.2 and 10.0):
absl-py==0.8.1
Markdown==3.1.1
numpy==1.17.3
Pillow==6.2.1
protobuf==3.10.0
PyYAML==5.1.2
six==1.12.0
tensorboard==2.0.0
torch==1.3.0
torchvision==0.4.1
tqdm==4.36.1
Werkzeug==0.16.0
-
hidden-networks: https://github.com/allenai/hidden-networks
-
POEM: https://github.com/deeplearning-wisc/poem
-
Energy: https://github.com/wetliu/energy_ood
-
ODIN: https://github.com/JoonHyung-Park/ODIN
If you find our paper and repo useful, please cite our paper:
@inproceedings{zhu2023unleashing,
title ={Unleashing Mask: Explore the Intrinsic Out-of-distribution Detection Capability},
author ={Jianing Zhu and Hengzhuang Li and Jiangchao Yao and Tongliang Liu and Jianliang Xu and Bo Han},
booktitle ={International Conference on Machine Learning},
year ={2023}
}