Asymmetric Loss For Multi-Label Classification

Official PyTorch Implementation

Emanuel Ben-Baruch, Tal Ridnik, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, Lihi Zelnik-Manor
DAMO Academy, Alibaba Group

Abstract

In a typical multi-label setting, a picture contains on average few positive labels, and many negative ones. This positive-negative imbalance dominates the optimization process, and can lead to under-emphasizing gradients from positive labels during training, resulting in poor accuracy. In this paper, we introduce a novel asymmetric loss ("ASL"), which operates differently on positive and negative samples. The loss enables to dynamically down-weights and hard-thresholds easy negative samples, while also discarding possibly mislabeled samples. We demonstrate how ASL can balance the probabilities of different samples, and how this balancing is translated to better mAP scores. With ASL, we reach state-of-the-art results on multiple popular multi-label datasets: MS-COCO, Pascal-VOC, NUS-WIDE and Open Images. We also demonstrate ASL applicability for other tasks, such as single-label classification and object detection. ASL is effective, easy to implement, and does not increase the training time or complexity.

Asymmetric Loss (ASL) Implementation

In this PyTorch file, we provide implementations of our new loss function, ASL, that can serve as a drop-in replacement for standard loss functions (Cross-Entropy and Focal-Loss)

For the multi-label case (sigmoids), the two implementations are:

class AsymmetricLoss(nn.Module)
class AsymmetricLossOptimized(nn.Module)

The two losses are bit-accurate. However, AsymmetricLossOptimized() contains a more optimized (and complicated) way of implementing ASL, which minimizes memory allocations, gpu uploading, and favors inplace operations.

For the single-label case (softmax), the implementations is called:

class ASLSingleLabel(nn.Module)

Pretrained Models

In this link, we provide pre-trained models on various dataset.

Validation Code

Thanks to external contribution of @hellbell, we now provide a validation code that repdroduces the article results on MS-COCO:

python validate.py  \
--model_name=tresnet_l \
--model_path=./models_local/MS_COCO_TRresNet_L_448_86.6.pth

Inference Code

We provide inference code, that demonstrate how to load our model, pre-process an image and do actuall inference. Example run of MS-COCO model (after downloading the relevant model):

python infer.py  \
--dataset_type=MS-COCO \
--model_name=tresnet_l \
--model_path=./models_local/MS_COCO_TRresNet_L_448_86.6.pth \
--pic_path=./pics/000000000885.jpg \
--input_size=448

which will result in:

Example run of OpenImages model:

python infer.py  \
--dataset_type=OpenImages \
--model_name=tresnet_l \
--model_path=./models_local/Open_ImagesV6_TRresNet_L_448.pth \
--pic_path=./pics/000000000885.jpg \
--input_size=448

Citation

 @misc{benbaruch2020asymmetric, 
        title={Asymmetric Loss For Multi-Label Classification}, 
        author={Emanuel Ben-Baruch and Tal Ridnik and Nadav Zamir and Asaf Noy and Itamar Friedman and Matan Protter and Lihi Zelnik-Manor}, 
        year={2020}, 
        eprint={2009.14119},
        archivePrefix={arXiv}, 
        primaryClass={cs.CV} }

Contact

Feel free to contact if there are any questions or issues - Emanuel Ben-Baruch (emanuel.benbaruch@alibaba-inc.com) or Tal Ridnik (tal.ridnik@alibaba-inc.com).

y78h11b09 / ASL