self_ensembling.py

import copy
from torch.nn import functional as F

from dassl.engine import TRAINER_REGISTRY, TrainerXU
from dassl.metrics import compute_accuracy
from dassl.modeling.ops.utils import sigmoid_rampup, ema_model_update


@TRAINER_REGISTRY.register()
class SelfEnsembling(TrainerXU):
    """Self-ensembling for visual domain adaptation.

    https://arxiv.org/abs/1706.05208.
    """

    def __init__(self, cfg):
        super().__init__(cfg)
        self.ema_alpha = cfg.TRAINER.SE.EMA_ALPHA
        self.conf_thre = cfg.TRAINER.SE.CONF_THRE
        self.rampup = cfg.TRAINER.SE.RAMPUP

        self.teacher = copy.deepcopy(self.model)
        self.teacher.train()
        for param in self.teacher.parameters():
            param.requires_grad_(False)

    def check_cfg(self, cfg):
        assert cfg.DATALOADER.K_TRANSFORMS == 2

    def forward_backward(self, batch_x, batch_u):
        global_step = self.batch_idx + self.epoch * self.num_batches
        parsed = self.parse_batch_train(batch_x, batch_u)
        input_x, label_x, input_u1, input_u2 = parsed

        logit_x = self.model(input_x)
        loss_x = F.cross_entropy(logit_x, label_x)

        prob_u = F.softmax(self.model(input_u1), 1)
        t_prob_u = F.softmax(self.teacher(input_u2), 1)
        loss_u = ((prob_u - t_prob_u)**2).sum(1)

        if self.conf_thre:
            max_prob = t_prob_u.max(1)[0]
            mask = (max_prob > self.conf_thre).float()
            loss_u = (loss_u * mask).mean()
        else:
            weight_u = sigmoid_rampup(global_step, self.rampup)
            loss_u = loss_u.mean() * weight_u

        loss = loss_x + loss_u
        self.model_backward_and_update(loss)

        ema_alpha = min(1 - 1 / (global_step+1), self.ema_alpha)
        ema_model_update(self.model, self.teacher, ema_alpha)

        loss_summary = {
            "loss_x": loss_x.item(),
            "acc_x": compute_accuracy(logit_x, label_x)[0].item(),
            "loss_u": loss_u.item(),
        }

        if (self.batch_idx + 1) == self.num_batches:
            self.update_lr()

        return loss_summary

    def parse_batch_train(self, batch_x, batch_u):
        input_x = batch_x["img"][0]
        label_x = batch_x["label"]
        input_u = batch_u["img"]
        input_u1, input_u2 = input_u

        input_x = input_x.to(self.device)
        label_x = label_x.to(self.device)
        input_u1 = input_u1.to(self.device)
        input_u2 = input_u2.to(self.device)

        return input_x, label_x, input_u1, input_u2