Practical overview
When it fits
Use it when you need a configurable research pipeline for training and evaluating deep metric-learning models on the datasets and methods listed in the repository.
Who it is for
- Researchers comparing deep metric-learning training strategies.
- Developers extending a PyTorch image-embedding training pipeline.
Documented uses
- Train a metric-learning model with a selected dataset, loss, sampling method and architecture.
- Compare loss and sampling choices across supported image datasets.
- Evaluate models using recall-at-k values and configurable embedding dimensions.
- Adapt the codebase to implement additional metric-learning methods.
Requirements & limits
The public GitHub repository is licensed under Apache-2.0.
- The documented stack is Python 3 and PyTorch 1.0.0/1; the README gives a Python 3.6 Conda setup and describes testing around CUDA 8 and CUDA 9.
- Faiss is used for k-means and nearest-neighbour computation; the README describes a scikit-learn alternative through auxiliaries_nofaiss.py.
- The documented architecture choices are ResNet50 and GoogLeNet, with default embedding dimensions of 128 and 512 respectively.
- PKU Vehicle-ID requires special licensing, according to the README.
- The README identifies a later extended repository with additional criteria, miners, metrics and logging options.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up the Python environment
Create and activate the documented Python environment and install the repository's required libraries.
Before you start
- Python 3.6 environment, using Conda as shown in the README
- Repository files, including requirements.txt and installer.sh
Steps
Create an environment with
conda create -n python=3.6.Activate it with
conda activate.Install the required libraries by running the command below.
bash installer.sh
Expected outcome
A configured environment with the libraries installed through the repository's installer script.
Guide limitations
- The README describes a pipeline built around Python 3 and PyTorch 1.0.0/1, and notes testing around CUDA 8 and CUDA 9.
- The README also says requirements.txt can be checked directly; it does not provide a separate execution or deployment procedure here.
Configure a documented training run
Choose a supported dataset, loss, sampling method, architecture and recall-at-k values for the main training script.
Before you start
- A prepared Python environment with the required libraries
- A dataset arranged in the structure described for that dataset in the README
Steps
Choose a dataset from cub200, cars196, online_products, in-shop or vehicle_id.
Choose a listed loss, sampling method and architecture: marginloss, triplet, npair or proxynca; distance, semihard, random or npair; and resnet50 or googlenet.
Set recall-at-k values and an embedding dimension as needed.
Use the documented
Standard_Training.pyarguments to specify those choices; consult the script's input arguments for other training settings.
Expected outcome
A configured command for a training run using the selected documented options.
Guide limitations
- The README's example default is ResNet50 on CUB200-2011 with Margin Loss and distance sampling.
- PKU Vehicle-ID is mentioned as optional in the repository's dataset overview, and requires special licensing; it is not included in the listed dataset flag values in the usage section.
- This recipe covers configuration, not a verified run or guaranteed result.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 17 October 2026
- Original website title
- Confusezius/Deep-Metric-Learning-Baselines · Official homepage title · retained in its source language
Related options
Revisiting_Deep_Metric_Learning_PyTorch — The README identifies this as a newly released codebase based on an extended version of this repository, with more criteria, miners, metrics and logging options.