PaddleClas/ppcls/loss/skdloss.py

73 lines
2.5 KiB
Python

# copyright (c) 2022 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle
import paddle.nn as nn
import paddle.nn.functional as F
class SKDLoss(nn.Layer):
"""
Spherical Knowledge Distillation
paper: https://arxiv.org/pdf/2010.07485.pdf
code reference: https://github.com/forjiuzhou/Spherical-Knowledge-Distillation
"""
def __init__(self,
temperature,
multiplier=2.0,
alpha=0.9,
use_target_as_gt=False):
super().__init__()
self.temperature = temperature
self.multiplier = multiplier
self.alpha = alpha
self.use_target_as_gt = use_target_as_gt
def forward(self, logits_student, logits_teacher, target=None):
"""Compute Spherical Knowledge Distillation loss.
Args:
logits_student: student's logits with shape (batch_size, num_classes)
logits_teacher: teacher's logits with shape (batch_size, num_classes)
"""
if target is None or self.use_target_as_gt:
target = logits_teacher.argmax(axis=-1)
target = F.one_hot(
target.reshape([-1]), num_classes=logits_student[0].shape[0])
logits_student = F.layer_norm(
logits_student,
logits_student.shape[1:],
weight=None,
bias=None,
epsilon=1e-7) * self.multiplier
logits_teacher = F.layer_norm(
logits_teacher,
logits_teacher.shape[1:],
weight=None,
bias=None,
epsilon=1e-7) * self.multiplier
kd_loss = -paddle.sum(F.softmax(logits_teacher / self.temperature) *
F.log_softmax(logits_student / self.temperature),
axis=1)
kd_loss = paddle.mean(kd_loss) * self.temperature**2
ce_loss = paddle.mean(-paddle.sum(
target * F.log_softmax(logits_student), axis=1))
return kd_loss * self.alpha + ce_loss * (1 - self.alpha)