durandtibo / wildcat.pytorch

PyTorch implementation of "WILDCAT: Weakly Supervised Learning of Deep ConvNets for Image Classification, Pointwise Localization and Segmentation", CVPR 2017

wildcat.pytorch

PyTorch implementation of "WILDCAT: Weakly Supervised Learning of Deep ConvNets for Image Classification, Pointwise Localization and Segmentation", CVPR 2017 (http://webia.lip6.fr/~durandt/pdfs/2017_CVPR/Durand_WILDCAT_CVPR_2017.pdf)

Requirements

Please, install the following packages

• numpy
• torch
• torchnet
• torchvision
• tqdm

Options

• k: number of regions for the spatial pooling. If k is larger than 1, k is the number of regions, otherwise k is the proportion of selected regions. k=0.2 means that 20% of the regions are used.
• maps: number of maps for each class
• alpha: weight for minimum regions
• lr: learning rate
• lrp: factor for learning rate of pretrained layers. The learning rate of the pretrained layers is lr * lrp
• batch-size: number of images per batch
• image-size: size of the image
• epochs: number of training epochs

Demo VOC 2007

python3 -m wildcat.demo_voc2007 ../data/voc --image-size 448 --batch-size 16 --lrp 0.1 --lr 0.01 --epochs 20 --k 0.2 --maps 8 --alpha 0.7

Demo MIT67

python3 -m wildcat.demo_mit67 ../data/mit67 --image-size 448 --batch-size 16 --lrp 0.1 --lr 0.001 --epochs 20 --k 0.4 --maps 8

Citing this repository

If you find this code useful in your research, please consider citing us:

@inproceedings{Durand_WILDCAT_CVPR_2017,
author = {Durand, Thibaut and Mordan, Taylor and Thome, Nicolas and Cord, Matthieu},
title = {{WILDCAT: Weakly Supervised Learning of Deep ConvNets for Image Classification, Pointwise Localization and Segmentation}},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2017}
}

Licence

MIT License

wildcat.pytorch's Issues

about def AveragePrecisionMeter.average_precision

hi, I don't figure out the function AveragePrecisionMeter.average_precision which is used to caculate the precision for each class. is that right? the result is different from precision definition "TP / (TP + FP)", could you explain it in detail, thanks very much

Training stuck at 0%

I'm currently using a docker image which include all required pytorch and CUDA 8.0 dependencies.
I just cloned the repo and executed the demo_voc2007 with the default argument values.
The training stuck at 0%, nothing happens.
Is there a known reason why is this happening?

Real Work Or Not?

Anyone uses this code to reproduce the performance reported in the paper?
From my understanding, max+min pooling really doesn't make any sense in deep learning. Any rationality mentioned by author in this paper all redirect to their several own previous work. Why does it work?
From my experiments using this code, I cannot produce any reported performance number in VOC2007. I can get highest MAP-92.3 via different parameter trials.
If anyone get the correct reported number (say VOC2007, MAP 95.0) in experiment, please kindly let me know the magic tricks. Thanks.

No code in _init_.py

Thank you for your implementation.
I find there is no code in init.py, is there anything wrong?

Pretrained Model

Thank you for your effort.Could you please share the pretrained model?

The hyperparameters about ms-coco dataset.

Just like voc2007 we need to run "python3 -m wildcat.demo_voc2007 ../data/voc --image-size 448 --batch-size 16 --lrp 0.1 --lr 0.01 --epochs 20 --k 0.2 --maps 8 --alpha 0.7", the ms-coco dataset's hyperparameters are?

About Learning a Deep ConvNet for Multi-label Classification with Partial Label.

Hello,
sorry for submitting an issue here.
I read this paper recently, and be interested in the topic.
I want to experiment it for:

1. analysis whether the GNN is still useful when the number of classes is small
2. think and try how to apply on object detection tasks.

But, I couldn't find source code on your github.
Could you release the PyTorch code of GNN and Never Ending Leaning parts?

Thanks.

confusion of MultiLabelSoftMarginLoss with label (0,-1,1)

in voc2007 cls task, there are three labels, 0(difficult example), -1(object absence) and 1(object existing). However, MultiLabelSoftMarginLoss only support y in {0,1}.

implementation for the segmentation?

Hi,

I could not find the implementation for the segmentation and its evaluation.... Where is it?

Best,
Satoshi

The mAP in voc2007

I only get best=93.395 when I run “python3 -m wildcat.demo_voc2007 ../data/voc --image-size 448 --batch-size 16 --lrp 0.1 --lr 0.01 --epochs 20 --k 0.2 --maps 8 --alpha 0.7”. How to get the same result as your paper（voc2007 best mAP=95）?

implementation for localization?

hello, I am so interested in how WILDCAT to implement localization, can you share the code of localization with me?

Not learning different modalities?

My experiments using this code so far (using the settings suggested in the readme) seem to be learning, and producing good classification and localization results, but the separate output layers are not learning different modalities as suggested by the paper. All 8 layers appear to be more or less identical to each other. Am I missing something?