thanks for your detailed code!

1. according to your paper and codes, train set only contains 1909 images for coco datasets followed by pareto distribution with max=1200 min=1. i am confused about the params settings, the train set is too samll, why not try to set a larger max param to construct a larger train set.
2. in table 1, for the ERM settings, are all images sampled uniformly without class aware sampler ? just as most common simply baseline(just like training imagenet except for BCEloss)? I have done some experiments on original coco14 datasets(also a long-tailed distribution), but i find using class-aware sampler get a inferior performance than baseline, 2~3 point below. maybe there is something wrong in my implementation.

https://github.com/wutong16/DistributionBalancedLoss/blob/14460d0acda6a3e77a8a6a5999b9746243cb1535/mllt/datasets/loader/sampler.py#L255 I wonder why this lined was commented and introduce the variance "num_samples_cls". In my opion, this line can accomplish the same function as the lines below, is that right? Or could you please tell why use variance "num_samples_cls", was that for code efficiency？ Thanks a lot!

Distribution-Balanced Loss is a very good work, we reproduced the experimental results in coco long tail experiment, but the difference of mAP on VOC dataset is almost 1 point, is there something wrong with our settings, looking forward to your reply~

Hi, thank you for providing your source code of the paper, Actually, I run you training code of VOC dataset for 3 times, but always get the Total mAP at around 77.7mAP. I notice in your paper "Table 1", the mAP of your "DB-focal" is 78.94. I was wondering if I did something wrong and why this is the case ?

Below is the log-out information of your training code. Split: head mAP:0.7344 acc:0.8429 micro:0.6096 macro:0.5392 Split:middle mAP:0.8359 acc:0.9375 micro:0.6400 macro:0.6527 Split: tail mAP:0.7651 acc:0.9400 micro:0.5492 macro:0.5305 Split: Total mAP:0.7771 acc:0.9101 micro:0.6024 macro:0.5698

Thanks in advance.

Traceback (most recent call last): File "tools/train.py", line 11, in from mllt.datasets import build_dataset File "/content/drive/My Drive/DistributionBalancedLoss/mllt/datasets/init.py", line 4, in from .Lvis import LvisDataset File "/content/drive/My Drive/DistributionBalancedLoss/mllt/datasets/Lvis.py", line 2, in from lvis.lvis import LVIS ModuleNotFoundError: No module named 'lvis' when I run python tools/train.py configs/coco/LT_resnet50_pfc_DB.py , this erorr is rasied. I think a lvis file should be existed, however, i can't find it. can you tell how to solve this erorr? thank you!

Is the dataset first released by authors? where can I download the dataset you evaluated in your manuscript? Or is there any widely-accepted public dataset for multi-label long-tailed classification task?

Hi, I tried to run python tools/train.py configs/coco/LT_resnet50_pfc_DB.py --gpus 2

and met some bugs:

File "/data2/cjq/DistributionBalancedLoss/mllt/models/losses/resample_loss.py", line 164, in rebalance_weight repeat_rate = torch.sum( gt_labels.float() * self.freq_inv, dim=1, keepdim=True) RuntimeError: expected device cuda:1 and dtype Float but got device cuda:0 and dtype Float

Does this code support multi-GPU training? Have you ever tried, or I should fix these bugs by myself?

Thanks for you code !

in Eq7, the for positive samples, zik in exponent should be -zik，and for negative samples，-λ *zik should be λ *zik, which is consistent with formulation in Eq5.

please correct me if there is something wrong. thanks~

Hey thanks for the code 😄 but have some questions!!

https://github.com/wutong16/DistributionBalancedLoss/blob/14460d0acda6a3e77a8a6a5999b9746243cb1535/mllt/models/losses/resample_loss.py#L158-L160

Isn't this inconsistent with eq7

1. Like line 159 is dealing with the exp power in eq7 right? (Line 159) why are we directly multiplying it with labels, that anyhow occurs when we pass it to self.cls_criterion right? (same goes for weight)

2. Bit confused in name https://github.com/wutong16/DistributionBalancedLoss/blob/14460d0acda6a3e77a8a6a5999b9746243cb1535/mllt/models/losses/resample_loss.py#L113 Is cls_score here (logits passed through sigmoid/softmax) or just pure raw classifier outputs?, because even though we call it cls_score it's passed the to the same F.binary_cross_entropy_with_logits_loss

Traceback (most recent call last): File "tools/train.py", line 108, in main() File "tools/train.py", line 85, in main train_dataset = build_dataset(cfg.data.train) File "/root/autodl-tmp/DB_loss/DistributionBalancedLoss/mllt/datasets/builder.py", line 36, in build_dataset dataset = build_from_cfg(cfg, DATASETS) File "/root/autodl-tmp/DB_loss/DistributionBalancedLoss/mllt/utils/registry.py", line 74, in build_from_cfg return obj_type(**args) File "/root/autodl-tmp/DB_loss/DistributionBalancedLoss/mllt/datasets/coco.py", line 27, in init super(CocoDataset, self).init(**kwargs) File "/root/autodl-tmp/DB_loss/DistributionBalancedLoss/mllt/datasets/custom.py", line 59, in init self.img_infos = self.load_annotations(ann_file, LT_ann_file) File "/root/autodl-tmp/DB_loss/DistributionBalancedLoss/mllt/datasets/coco.py", line 32, in load_annotations self.coco = COCO(ann_file) File "/root/miniconda3/envs/blast/lib/python3.7/site-packages/pycocotools-2.0-py3.7-linux-x86_64.egg/pycocotools/coco.py", line 84, in init dataset = json.load(open(annotation_file, 'r')) FileNotFoundError: [Errno 2] No such file or directory: '/mnt/SSD/det/coco/annotations/instances_train2017.json'

When I run python tools/train.py configs/coco/LT_resnet50_pfc_DB.py ,this error was rasied. I think this is because the program did not find the location of coco dataset. But I'm not sure. Can you tell me how to solve this error? Thank you!

args.gpus is missing https://github.com/wutong16/DistributionBalancedLoss/blob/a3ecaa9021a920fcce9fdafbd7d83b51bf526af8/tools/train.py#L59

maybe add parser.add_argument('--gpus', default=1) in the parse_args()