htdt / hyp_metric

Hyperbolic Vision Transformers: Combining Improvements in Metric Learning | Official repository

Hyperbolic Vision Transformers: Combining Improvements in Metric Learning

CVPR 2022    arxiv.org/abs/2203.10833    Papers With Code

Code includes

• Proxy-Anchor for datasets and evaluation (uses pytorch_metric_learning);
• hyperbolic-image-embeddings for hyperbolic operations;
• train.py - main training;
• eval_pretrain.py - encoder evaluation without training;
• delta.py - δ-hyperbolicity evaluation.

Run training

python -m torch.distributed.launch --nproc_per_node=4 train.py  # multi GPU
python -m train --help  # single GPU

Configs

python -m train --ds CUB --model vit_small_patch16_224 --num_samples 9 --lr 3e-5 --ep 50 --eval_ep "[50]" --resize 256
python -m train --ds CUB --model dino_vits16 --num_samples 9 --lr 1e-5 --ep 50 --eval_ep "[50]" --resize 256
python -m train --ds CUB --model deit_small_distilled_patch16_224 --num_samples 9 --lr 3e-5 --ep 50 --eval_ep "[50]" --resize 256

python -m train --ds Cars --model vit_small_patch16_224 --num_samples 9 --bs 882 --lr 3e-5 --ep 300 --eval_ep "[300]"
python -m train --ds Cars --model dino_vits16 --num_samples 9 --bs 882 --lr 1e-5 --ep 300 --eval_ep "[300]"
python -m train --ds Cars --model deit_small_distilled_patch16_224 --num_samples 9 --bs 882 --lr 3e-5 --ep 300 --eval_ep "[300]"

python -m train --ds SOP --model vit_small_patch16_224 --lr 3e-5 --ep 200 --eval_ep "[200]"
python -m train --ds SOP --model dino_vits16 --lr 1e-5 --ep 200 --eval_ep "[200]"
python -m train --ds SOP --model deit_small_distilled_patch16_224 --lr 3e-5 --ep 200 --eval_ep "[200]"

python -m train --ds Inshop --model vit_small_patch16_224 --lr 3e-5 --ep 400 --eval_ep "[400]"
python -m train --ds Inshop --model dino_vits16 --lr 1e-5 --ep 400 --eval_ep "[400]"
python -m train --ds Inshop --model deit_small_distilled_patch16_224 --lr 3e-5 --ep 400 --eval_ep "[400]"

# add --hyp_c 0 --t 0.1 for sphere version
# use --clip_r 0 to disable clipping
# use --eval_ep "r(300,410,10)" to evaluate every 10 epoch between 300 and 400

Setup

pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

git clone https://github.com/NVIDIA/apex

pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./apex

pip install tqdm wandb timm typed-argument-parser pytorch_metric_learning

pip uninstall -y scipy && pip install scipy

wandb login

Datasets

• CUB-200
• Stanford Online Products ftp://cs.stanford.edu/cs/cvgl/Stanford_Online_Products.zip
• In-shop Clothes Retrieval (In-shop Clothes Retrieval Benchmark -> Img -> img.zip, Eval/list_eval_partition.txt)
• Cars-196 labels