Multi-Glimpse Network
Multi-Glimpse Network: A Robust and Efficient Classification Architecture based on Recurrent Downsampled Attention
Requirements
Python ≥ 3.8
Installation
For example, venv + pip:
$ python3 -m venv env
$ source env/bin/activate
(env) $ python3 -m pip install -r requirements.txtEvaluation
Accuracy on clean images
- Create ImageNet100 from ImageNet (using symbolic links).
$ python3 tools/create_imagenet100.py tools/imagenet100.txt \
/path/to/ImageNet /path/to/ImageNet100-
Download checkpoints from Google Drive.
-
Test accuracy.
$ export dataset="--train_dir /path/to/ImageNet100/train \
--val_dir /path/to/ImageNet100/val \
--dataset imagenet --num_class 100"
# Baseline
$ python3 main.py $dataset --test --n_iter 1 --scale 1.0 --model resnet18 \
--checkpoint resnet18_baseline
# Ours
$ python3 main.py $dataset --test --n_iter 4 --scale 2.33 --model resnet18 \
--checkpoint resnet18_ours --alpha 0.6 --s 0.02Add the flag --flop_count to count the approximate FLOPs for the inference of an image. (using fvcore)
Accuracy on adversarial attacks (PGD)
- Test adversarial accuracy.
# Baseline
$ python3 main.py $dataset --test --n_iter 1 --scale 1.0 --adv --step_k 10 \
--model resnet18 --checkpoint resnet18_baseline
# Ours
$ python3 main.py $dataset --test --n_iter 4 --scale 2.33 --adv --step_k 10 \
--model resnet18 --checkpoint resnet18_ours --alpha 0.6 --s 0.02Accuracy on common corruptions
- Create ImageNet100-C from ImageNet-C (using symbolic links).
$ python3 tools/create_imagenet100c.py \
tools/imagenet100.txt /path/to/ImageNet-C/ /path/to/ImageNet100-C/- Test for a single corruption.
$ export dataset="--train_dir /path/to/ImageNet100/train \
--val_dir /path/to/ImageNet100-C/pixelate/5 \
--dataset imagenet --num_class 100"
# Baseline
$ python3 main.py $dataset --test --n_iter 1 --scale 1.0 --model resnet18 \
--checkpoint resnet18_baseline
# Ours
$ python3 main.py $dataset --test --n_iter 4 --scale 2.33 --model resnet18 \
--checkpoint resnet18_ours --alpha 0.6 --s 0.02- A simple script to test all corruptions and collect results.
# Modify tools/eval_imagenet100c.py and run it to generate script
$ python3 tools/eval_imagenet100c.py /home2/ImageNet100-C/ > run.sh
# Evaluate
$ bash run.sh
# Collect results
$ python3 tools/collect_imagenet100c.pyTraining
$ export dataset="--train_dir /path/to/ImageNet100/train \
--val_dir /path/to/ImageNet100/val \
--dataset imagenet --num_class 100"
# Baseline
$ python3 main.py $dataset --epochs 400 --n_iter 1 --scale 1.0 \
--model resnet18 --gpu 0,1,2,3
# Ours
$ python3 main.py $dataset --epochs 400 --n_iter 4 --scale 2.33 \
--model resnet18 --alpha 0.6 --s 0.02 --gpu 0,1,2,3Check tensorboard for the logs. (When training with multiple gpus, the log value may be scaled by the number of gpus except for the validation accuracy)
tensorboard --logdir=logsNote that we left our exploration in the code for further study, e.g., self-supervised spatial guidance, dynamic gradient re-scaling operation.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent