Rethinking Class Activation Mapping for Weakly Supervised Object Localization [Paper][Project Page]

Datasets

Create symlinks for datasets: CUB-200-2011 and ImageNet-1k. Note that the name of the directory for CUB should be 'CUB_200_2011'.

cd /path/to/this/repo
mkdir data && cd data
ln -s /path/to/cub/dataset
ln -s /path/to/imagenet/dataset

Setup

Setup the pipeline by installing dependencies, pretrained models and utils.

cd /path/to/this/repo
bash scripts/setup.sh --data_setup

Train

To train a model, use the command below. Tag can be an arbitrary string.

cd /path/to/this/repo
bash scripts/run_train.sh --config_path configs/vgg16_tap.yml --tag vgg16_tap --train_dir /path/to/train/dir

Eval

To eval a model, use the command below. Note that checkpoint dir has to be synchronized with train_dir used for training.

cd /path/to/this/repo
bash scripts/run_eval.sh --config_path configs/baseline/vgg16_tap.yml --tag vgg16_tap --checkpoint_dir /path/to/checkpoint/dir

NOTE: NWC and PaS can be used in evaluation time. In configs/vgg16_tap.yml, NWC and PaS can be employed simply changing paraameters under 'eval'. For NWC, change 'truncate' to True. For PaS, change 'percentile' to 90 and 'loc_threshold' to 0.35.

TODO

Model and configs for the other backbones such as Resnet50-SE, Inception need to be added.

Citation

If you use this code or model for your research, please cite:

@InProceedings{bae2020rethinkingCAM,
  author = {Wonho Bae and Junhyug Noh and Gunhee Kim},
  title = {Rethinking Class Activation Mapping for Weakly Supervised Object Localization},
  booktitle = {The European Conference on Computer Vision (ECCV)},
  month = {August},
  year = {2020}
}

Acknowledgment

Many of core functions are borrowed from ADL.

RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.

Hi, my name is Yuen.

Thank you for sharing this wonderful work.

So far I followed steps and trained the vgg16 model.

However, when I tried "eval" command, it showed the following errors:

python eval.py --config_path=configs/vgg16_tap.yml --tag=vgg16_tap --checkpoint_dir=data
[2022-03-27 15:36:24,223 - data\eval_vgg16_tap] GPU is not available.
[2022-03-27 15:36:24,268 - data\eval_vgg16_tap] VGG network structure: [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 1024]
[2022-03-27 15:36:24,452 - data\eval_vgg16_tap] Checkpointer is built.
[2022-03-27 15:36:24,453 - data\eval_vgg16_tap] Loading checkpoint from data\checkpoint_191_1150014.pth
Traceback (most recent call last):
  File "eval.py", line 31, in <module>
    engine.evaluate()
  File "D:\xxxx\rethinkingCAM\src\engine.py", line 175, in evaluate
    epoch, num_step, **self.eval_config)
  File "D:\xxxx\rethinkingCAM\src\engine.py", line 208, in _eval_one_epoch
    top1_cls, top5_cls = metrics.topk_accuracy(predictions, labels, topk=(1,5))
  File "D:\xxxx\rethinkingCAM\src\utils\metrics.py", line 22, in topk_accuracy
    correct_k = correct[:k].view(-1).float().sum(0, keepdim=False)
RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.

Can I know how to resolve it ?

I am using the pytorch 1.11

Thank you

Best regads,
Yuen

won-bae / rethinkingcam Goto Github PK

rethinkingcam's Introduction

Rethinking Class Activation Mapping for Weakly Supervised Object Localization [Paper][Project Page]

Datasets

Setup

Train

Eval

TODO

Citation

Acknowledgment

rethinkingcam's People

Contributors

Stargazers

Watchers

Forkers

rethinkingcam's Issues

Recommend Projects

Recommend Topics

Recommend Org