dichotomies / proxy-nca

PyTorch Implementation of `No Fuss Distance Metric Learning using Proxies`.

About

This repository contains a PyTorch implementation of No Fuss Distance Metric Learning using Proxies as introduced by Google Research.

The training and evaluation setup is exactly the same as described in the paper, except that Adam was used as optimizer instead of RMSprop.

I have ported the PyTorch BN-Inception model from PyTorch 0.2 to PyTorch >= 0.4. It's weights are stored inside the repository in the directory net.

You need Python3, PyTorch >= 1.1 and torchvision >= 0.3.0 to run the code. I have used CUDA Version 10.0.130.

Note that negative log with softmax is used as ProxyNCA loss. Therefore, the anchor-positive-proxy distance is not excluded in the denominator. In practice, I have not noticed a difference.

The importance of scaling of the normalized proxies and embeddings is mentioned in the ProxyNCA paper (in the theoretical background), but the exact scaling factors are ommitted. I have found that (3, 3) work well for CUB and Cars and (8, 1) work well for SOP (first being for proxies and latter for embeddings).

Reproducing Results

You can adjust most training settings (learning rate, optimizer, criterion, dataset, ...) in the config file.

You'll only have to adjust the root paths for the datasets. Then you're ready to go.

Downloading and Extracting the Datasets

Cars 196

mkdir cars196
cd cars196
wget http://imagenet.stanford.edu/internal/car196/cars_annos.mat
wget http://imagenet.stanford.edu/internal/car196/car_ims.tgz
tar -xzvf car_ims.tgz
pwd # use this path as root path for config file

CUB 200-2011

wget http://www.vision.caltech.edu/visipedia-data/CUB-200-2011/CUB_200_2011.tgz
tar -xzvf CUB_200_2011.tgz
cd CUB_200_2011
pwd # use this path as root path for config file

SOP

wget ftp://cs.stanford.edu/cs/cvgl/Stanford_Online_Products.zip
unzip Stanford_Online_Products.zip
cd Stanford_Online_Products
pwd # use this path as root path for config file

Commands

DATA=cub; SCALING_X=3.0; SCALING_P=3.0; LR=1; python3 train.py --data $DATA \
--log-filename $DATA-scaling_x_$SCALING_X-scaling_p_$SCALING_P-lr_$LR \
--config config.json --epochs=20 --gpu-id 0 --lr-proxynca=$LR \
--scaling-x=$SCALING_X --scaling-p=$SCALING_P --with-nmi

DATA=cars; SCALING_X=3.0; SCALING_P=3.0; LR=1; python3 train.py --data $DATA \
--log-filename $DATA-scaling_x_$SCALING_X-scaling_p_$SCALING_P-lr_$LR \
--config config.json --epochs=50 --gpu-id 1 --lr-proxynca=$LR \
--scaling-x=$SCALING_X --scaling-p=$SCALING_P --with-nmi

DATA=sop; SCALING_X=1; SCALING_P=8; LR=10; python3 train.py --data $DATA \
--log-filename $DATA-scaling_x_$SCALING_X-scaling_p_$SCALING_P-lr_$LR \
--config config.json --epochs=50 --gpu-id 3 --lr-proxynca=$LR \
--scaling-x=$SCALING_X --scaling-p=$SCALING_P

Results

The results were obtained mostly with one Titan X or a weaker GPU.

Reading: This Implementation [Google's Implementation].

	CUB	Cars	SOP
Duration	00:19h	00:24h	01:55h
Epoch	17	15	16
Log	here	here	here
R@1	52.63 [49.21]	72.19 [73.22]	74.07 [73.73]
R@2	64.63 [61.90]	81.31 [82.42]	79.13 [-------]
R@4	75.76 [67.90]	87.54 [86.36]	83.30 [-------]
R@8	84.52 [72.40]	92.54 [88.68]	86.66 [-------]
NMI	60.64 [59.53]	62.45 [64.90]	----------

Referencing this Implementation

If you'd like to reference this ProxyNCA implementation, you can use this bibtex:

@misc{Tschernezki2020,
  author = {Tschernezki, Vadim and Sanakoyeu, Artsiom and Ommer, Bj{\"o}rn,},
  title = {PyTorch Implementation of ProxyNCA},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/dichotomies/proxy-nca}},
}