Generalization in Metric Learning: Should the Embedding Layer be the Embedding Layer?

This reproduces the result from our paper https://arxiv.org/abs/1803.03310

The code isn't too clean for now, but it should work

Benchmarks

R@1 retrieval performance

Dataset	Cars-196	CUB-200-2011	Stanford Online Product
Lifted structure [17]	53.0	47.2	62.5
HDC [40]	73.7	53.6	70.9
N-pair [27]	71.1	51.0	67.7
Proxy-NCA [16]	73.2	49.2	73.7
Our (layer pool5.3)	87.8	66.4	74.8

Prerequisite

Python, pytorch, numpy, scipy, etc.

Produce the result on Cars-196

Download cars-196 dataset from http://ai.stanford.edu/~jkrause/cars/car_dataset.html (cars_annos.mat & unzip car_ims.tgz)

Edit train.py, update the value of dataset_path

Run the training: python train.py

To reproduce result on CUB and OnlineProduct dataset, refer to this #1

Notes

We used pytorch version '0.3.0.post4' and scipy '0.19.1'

We noticed scipy.misc.imresize behaviour could vary, so using a different version of scipy might result in slightly worse/different performance than what's reported in our paper.

>>> scipy.__version__
'0.19.1'
>>> scipy.misc.imresize(np.array([[0, 100, 200]]), [1,6])
array([[  0,  32,  95, 159, 223, 255]], dtype=uint8)

>>> scipy.__version__
'1.0.0'
>>> scipy.misc.imresize(np.array([[0, 100, 200]]), [1,6])
array([[  0,   0,  63, 127, 191, 255]], dtype=uint8)