"Label Structure Preserving Contrastive Embedding for Multi-Label Learning with Missing Labels"
Zhongchen Ma†, Lisha Li†, Qirong Mao, Senior Member, IEEE, and Songcan Chen∗, Senior Member, IEEE
https://arxiv.org/abs/2209.01314
you can try:
$ python train_clml.py --dataset ./dataset/coco_train_0.75left.txt --data /home/mscoco --b 64 --loss BCE --lambda_ 1.00 --useclml True --threshold 0.8
and then start training:
creating model...
num_classes = 80
model use imagenet pretained!
done
loading annotations into memory...
Done (t=4.78s)
creating index...
index created!
load class_nums = 80
len(val_dataset)): 40137
len(train_dataset)): 82081
Epoch [0/80], Step [000/1283], LR 4.0e-06, Loss: 3719.9
Epoch [0/80], Step [100/1283], LR 4.0e-06, Loss: 2793.5
···