# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.

# Copyright (c) 2015-present, Facebook, Inc.
# All rights reserved.
"""
Implements the knowledge distillation loss
Modified from: https://github.com/facebookresearch/deit
"""
import torch
from torch.nn import functional as F
import torch.nn as nn


class DistillationLoss(torch.nn.Module):
    """
    This module wraps a standard criterion and adds an extra knowledge distillation loss by
    taking a teacher model prediction and using it as additional supervision.
    """
    def __init__(self, base_criterion: torch.nn.Module, teacher_model: torch.nn.Module,
                 distillation_type: str, alpha: float, tau: float):
        super().__init__()
        self.base_criterion = base_criterion
        self.teacher_model = teacher_model
        assert distillation_type in ['none', 'soft', 'hard']
        self.distillation_type = distillation_type
        self.alpha = alpha
        self.tau = tau

    def forward(self, inputs, outputs, labels):
        """
        Args:
            inputs: The original inputs that are feed to the teacher model
            outputs: the outputs of the model to be trained. It is expected to be
                either a Tensor, or a Tuple[Tensor, Tensor], with the original output
                in the first position and the distillation predictions as the second output
            labels: the labels for the base criterion
        """
        outputs_kd = None
        if not isinstance(outputs, torch.Tensor):
            # assume that the model outputs a tuple of [outputs, outputs_kd]
            outputs, outputs_kd = outputs

        if isinstance(outputs, tuple):
            base_loss = self.base_criterion(outputs, labels, inputs)
        else:
            base_loss = self.base_criterion(outputs, labels)

        if self.distillation_type == 'none':
            return base_loss

        if outputs_kd is None:
            raise ValueError("When knowledge distillation is enabled, the model is "
                             "expected to return a Tuple[Tensor, Tensor] with the output of the "
                             "class_token and the dist_token")
        # don't backprop throught the teacher
        with torch.no_grad():
            teacher_outputs = self.teacher_model(inputs)

        if self.distillation_type == 'soft':
            T = self.tau
            # taken from https://github.com/peterliht/knowledge-distillation-pytorch/blob/master/model/net.py#L100
            # with slight modifications
            distillation_loss = F.kl_div(
                F.log_softmax(outputs_kd / T, dim=1),
                F.log_softmax(teacher_outputs / T, dim=1),
                reduction='sum',
                log_target=True
            ) * (T * T) / outputs_kd.numel()
        elif self.distillation_type == 'hard':
            distillation_loss = F.cross_entropy(outputs_kd, teacher_outputs.argmax(dim=1))

        loss = base_loss * (1 - self.alpha) + distillation_loss * self.alpha
        return loss