MLL - Metric Learning Layers

MLL is a simple PyTorch package that includes the most common metric learning layers. MLL only includes layers that are not dependent on negative sample mining and therefore drop in replacements for the final linear layer used in classification problems. All layers aim to achieve greater inter-class variance and minimizing intra-class variance. Moreover, all MLL-layers can be used in conjunction with soft-targets (e.g. with Mixup).

The basis of all these layers is the scaled cosine similarity $$y = xW * s$$ between the $d$-dimensional input vectors (features) $x \in \mathbb{R}^{1 \times d}$ and the $c$ class weights (prototypes, embeddings) $W \in \mathbb{R}^{d \times c}$ where $||x|| = 1$ and $||W_{*, j}|| = 1 ,, \forall j= 1\dots c$ and $s \in \mathbb{R}^+$.

Supported Layers

We currently support the following layers:

• NormalizedLinear and ScaledNormalizedLinear
• CosFace
• ArcFace
• AdaCos and FixedAdaCos
• DeepNCM

You can use multiple sub-centers for all layers except for DeepNCM. If you do not specify a scale, MLL will use the heuristic from AdaCos $s = \sqrt{2} * \log{c-1}$.

Install MLL

Simply run:

pip install metric-learning-layers

Example

import torch
import metric_learning_layers as mll


rnd_batch  = torch.randn(32, 128)
rnd_labels = torch.randint(low=0, high=10, size=(32, ))

arcface = mll.ArcFace(in_features=128, 
                      out_features=10, 
                      num_sub_centers=1, 
                      scale=None, # defaults to AdaCos heuristic
                      trainable_scale=False
                      )

af_out = arcface(rnd_batch, rnd_labels)  # ArcFace requires labels (used to apply the margin)
# af_out: torch.Size([32, 10])

adacos = mll.AdaCos(in_features=128, 
                    out_features=10, 
                    num_sub_centers=1 
                    )

ac_out = arcface(rnd_batch)  # AdaCos does not require labels
# af_out: torch.Size([32, 10])