oneflow.nn.CombinedMarginLoss

class oneflow.nn.CombinedMarginLoss(m1: float = 1.0, m2: float = 0.0, m3: float = 0.0)

The operation implements “margin_softmax” in InsightFace: https://github.com/deepinsight/insightface/blob/master/recognition/arcface_mxnet/train.py The implementation of margin_softmax in InsightFace is composed of multiple operators. We fuse them for speed up.

Applies the function:

\[\begin{split}{\rm CombinedMarginLoss}(x_i, label) = \left\{\begin{matrix} \cos(m_1\cdot\arccos x_i+m_2) - m_3 & {\rm if} \ i == label \\ x_i & {\rm otherwise} \end{matrix}\right.\end{split}\]
Parameters

  • x (oneflow.Tensor) – A Tensor

  • label (oneflow.Tensor) – label with integer data type

  • m1 (float) – loss m1 parameter

  • m2 (float) – loss m2 parameter

  • m3 (float) – loss m3 parameter

Note

Here are some special cases:

  • when \(m_1=1, m_2\neq 0, m_3=0\), CombineMarginLoss has the same parameter as ArcFace .

  • when \(m_1=1, m_2=0, m_3\neq 0\), CombineMarginLoss has the same parameter as CosFace (a.k.a AM-Softmax) .

  • when \(m_1\gt 1, m_2=m_3=0\), CombineMarginLoss has the same parameter as A-Softmax.

Returns

A Tensor

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow
>>> np_x = np.array([[-0.7027179, 0.0230609], [-0.02721931, -0.16056311], [-0.4565852, -0.64471215]])
>>> np_label = np.array([0, 1, 1])
>>> x = flow.tensor(np_x, dtype=flow.float32)
>>> label = flow.tensor(np_label, dtype=flow.int32)
>>> loss_func = flow.nn.CombinedMarginLoss(0.3, 0.5, 0.4)
>>> out = loss_func(x, label)
>>> out
tensor([[-0.0423,  0.0231],
        [-0.0272,  0.1237],
        [-0.4566, -0.0204]], dtype=oneflow.float32)