Alibaba-MIIL/ML_Decoder
Alibaba-MIIL/ML_Decoder has 354 GitHub stars and 56 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-21 17:46 UTC. Last push 2023-01-20.
Crecimiento de estrellas medido
- in 24h
- 0
- in 7 days
- Recopilando historial
- in 15 days
- Recopilando historial
- in 1 month
- Recopilando historial
Cambios netos en períodos que se solapan, no totales diarios. Los datos no disponibles se dejan en blanco. Las estrellas indican interés, no calidad.
Acerca de
Official PyTorch implementation of "ML-Decoder: Scalable and Versatile Classification Head" (2021)
Aprendizaje automáticoREADME.md
OriginalML-Decoder: Scalable and Versatile Classification Head
Paper |
Pretrained Models |
Datasets
Official PyTorch Implementation
Tal Ridnik, Gilad Sharir, Avi Ben-Cohen, Emanuel Ben-Baruch, Asaf Noy
DAMO Academy, Alibaba Group
Abstract
In this paper, we introduce ML-Decoder, a new attention-based classification head. ML-Decoder predicts the existence of class labels via queries, and enables better utilization of spatial data compared to global average pooling. By redesigning the decoder architecture, and using a novel group-decoding scheme, ML-Decoder is highly efficient, and can scale well to thousands of classes. Compared to using a larger backbone, ML-Decoder consistently provides a better speed-accuracy trade-off. ML-Decoder is also versatile - it can be used as a drop-in replacement for various classification heads, and generalize to unseen classes when operated with word queries. Novel query augmentations further improve its generalization ability. Using ML-Decoder, we achieve state-of-the-art results on several classification tasks: on MS-COCO multi-label, we reach 91.4% mAP; on NUS-WIDE zero-shot, we reach 31.1% ZSL mAP; and on ImageNet single-label, we reach with vanilla ResNet50 backbone a new top score of 80.7%, without extra data or distillation.
![]() |
![]() |
ML-Decoder Implementation
ML-Decoder implementation is available here. It can be easily integrated into any backbone using this example code:
ml_decoder_head = MLDecoder(num_classes) # initilization
spatial_embeddings = self.backbone(input_image) # backbone generates spatial embeddings
logits = ml_decoder_head(spatial_embeddings) # transfrom spatial embeddings to logits
Inference Code and Pretrained Models
See Model Zoo
![]() |
Training Code
We share a full reproduction code for the article results.
Multi-label Training Code
A reproduction code for MS-COCO multi-label:
python train.py \
--data=/home/datasets/coco2014/ \
--model_name=tresnet_l \
--image_size=448
Single-label Training Code
Our single-label training code uses the excellent timm repo. Reproduction code is currently from a fork, we will work toward a full merge to the main repo.
git clone https://github.com/mrT23/pytorch-image-models.git
This is the code for A2 configuration training, with ML-Decoder (--use-ml-decoder-head=1):
python -u -m torch.distributed.launch --nproc_per_node=8 \
--nnodes=1 \
--node_rank=0 \
./train.py \
/data/imagenet/ \
--amp \
-b=256 \
--epochs=300 \
--drop-path=0.05 \
--opt=lamb \
--weight-decay=0.02 \
--sched='cosine' \
--lr=4e-3 \
--warmup-epochs=5 \
--model=resnet50 \
--aa=rand-m7-mstd0.5-inc1 \
--reprob=0.0 \
--remode='pixel' \
--mixup=0.1 \
--cutmix=1.0 \
--aug-repeats 3 \
--bce-target-thresh 0.2 \
--smoothing=0 \
--bce-loss \
--train-interpolation=bicubic \
--use-ml-decoder-head=1
ZSL Training Code
First download the following files to the root path of the dataset:
benchmark_81_v0.json
wordvec_array.pth
data.csv
Training code for NUS-WIDE ZSL:
python train_zsl_nus.py \
--data=/home/datasets/nus_wide/ \
--image_size=224
New Top Results - Stanford-Cars and CIFAR 100
Using ML-Decoder classification head, we reached a top result of 96.41% on Stanford-Cars dataset, and 95.1% on CIFAR-100 dataset. We will add this result to a future version of the paper.
Citation
@misc{ridnik2021mldecoder,
title={ML-Decoder: Scalable and Versatile Classification Head},
author={Tal Ridnik and Gilad Sharir and Avi Ben-Cohen and Emanuel Ben-Baruch and Asaf Noy},
year={2021},
eprint={2111.12933},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
Explorar proyectos relacionados
Cargando…


