feymanpriv / pymetric

pytorch metric learning tools and pycls

pymetric

pymetric tries to build a metric learning and retrieval codebase based on PyTorch. It refers from pycls On Network Design Spaces for Visual Recognition project and fast-reid from (https://github.com/JDAI-CV/fast-reid).

Introduction

• 2nd place in Google Landmark Retrieval 2020 Competition. (https://drive.google.com/file/d/1XnzxMOHhzua9tjrAjo-X55ieKVJzRJw_/view?usp=sharing)
• Includes pycls (https://github.com/facebookresearch/pycls), refer to pycls.md.
• pymetric is written by DistributedDataParallel which is different from fast-reid. Now mainly includes features such as arcface loss and circle loss, ongoing.
• Add multicards feature extraction, searching topk and computing mAP.

Installation

Requirements:

• NVIDIA GPU, Linux, Python3(tested on 3.6.10)
• PyTorch, various Python packages; Instructions for installing these dependencies are found below

Notes:

• pymetric does not currently support running on CPU; a GPU system is required
• pymetric has been tested with CUDA 10.2 and cuDNN 7.1

PyTorch

To install PyTorch with CUDA support, follow the installation instructions from the PyTorch website.

pymetric

Clone the pymetric repository:

# PYMETRIC=/path/to/clone/pymetric
git clone https://github.com/feymanpriv/pymetric $PYMETRIC

Install Python dependencies:

pip install -r $PYMETRIC/requirements.txt

Set PYTHONPATH:

cd $PYMETRIC && export PYTHONPATH=`pwd`:$PYTHONPATH

Datasets

Same with pycls, pymetric finds datasets via symlinks from metric/datasets/data to the actual locations where the dataset images and annotations are stored. Refer to DATA.md.

Getting Started

Training a metric model:

python tools/train_metric.py \
    --cfg configs/metric/R-50-1x64d_step_8gpu.yaml \
    OUT_DIR ./output \
    PORT 12001 \
    TRAIN.WEIGHTS path/to/pretrainedmodel

Resume training:

python tools/train_metric.py \
    --cfg configs/metric/R-50-1x64d_step_8gpu.yaml \
    OUT_DIR ./output \
    PORT 12001 \
    TRAIN.AUTO_RESUME True

Extracting features(labels) and evaluation

set ${total_num} = n*(gpu_cards)
sh tools/metric/eval/infer.sh
python search.py search_gpu ${queryfea_path}, ${referfea_path}, ${output}

Convert to tensorflow2.3 (please refer onnx and onnx-tensorflow)
examples: tools/convert/torch2onnx.py tools/convert/onnx2tf.py

Pretrained weights

-resnet50 (c3l3)

-resnet101 (t3ln)

-resnest269 (3c5a)

Results

2nd place on Google Landmark Retrieval Challenge 2020

Backbone	Scale	Margin	Size	Public Score	Private Score
ResNeSt269	30	0.15	224/224	0.35129	0.30819
ResNeSt269	30	0.15	448/448	0.36972	0.33015
ResNeSt269	30	0.15	640/448	0.39040	0.34718

License

pymetric is released under the MIT license. Please see the LICENSE file for more information.

pymetric's Issues

how can i get a pretrained model ?

how can i get a pretrained model ?

The provided resnest269 checkpoint does not have the embedding_layer weights

Nice work.

I downloaded the provided resnest269 pretrained checkpoint, but find it only has the weights for backbone and a 1000-class classificaiton header, but not the expected embedding layer weights. Any chance to share a version with the embedding layer weights as well?

For reference, here is the name of the last few layers in the state_dict loaded from the provided checkpoint

layer4.7.conv2.bn1.running_var                                                                                                 
layer4.7.conv2.bn1.num_batches_tracked                         
layer4.7.conv2.fc2.weight                                                                                                      
layer4.7.conv2.fc2.bias
layer4.7.conv3.weight                                                                                                          
layer4.7.bn3.weight                                                                                                            
layer4.7.bn3.bias                                              
layer4.7.bn3.running_mean                                      
layer4.7.bn3.running_var                                                                                                       
layer4.7.bn3.num_batches_tracked                               
fc.weight                                                                                                                      
fc.bias      

The fc.weight has shape [1000,2048]

How to perform inference?

I want to ask how to perform inference?