Examples on Google Colab

Before running the notebooks, make sure that the runtime type is set to "GPU", by going to the Runtime menu, and clicking on "Change runtime type".

Click "Open in playground" on the Colab header to interact with the notebook.

Simple examples

Notebook	Description	Colab Link
MNIST using TripletMarginLoss	Train with TripletMarginLoss, evaluate with AccuracyCalculator.
MNIST using SubCenterArcFaceLoss	Train with SubCenterArcFaceLoss, view outlier samples.
MoCo on CIFAR10	Self-supervision using MoCo with CrossBatchMemory
Multiprocessing with DistributedDataParallel	An example using pytorch_metric_learning.utils.distributed

Training/testing workflows with logging and model saving

The following notebooks are meant to show entire training/testing workflows. (If you want to use just a loss or miner in your own code, see the notebooks above.) They generally go through the following steps:

• initialize models, optimizers, and transforms
• creates train/validation splits
• initialize a loss, miner, sampler, trainer, and tester
• train the model, record accuracy, and plot the embedding space

Notebook	Description	Colab Link
MetricLossOnly	Use just a metric loss.
A scRNAseq Metric Embedding	An example using canonical single-cell RNAseq cell types.
TrainWithClassifier	Use a metric loss + classification loss and network.
CascadedEmbeddings	Use multiple sub-networks and mine their outputs.
DeepAdversarialMetricLearning	Use a generator to create hard negatives during training.
TwoStreamMetricLoss	For use with two-stream datasets, where anchors and positives/negatives come from different sources.
Inference	Use the inference module after you're done training.