# meter.nres = residual[n_mask].mean()
# meter.nsum = self.sum_mem[ids][n_mask].mean()
# meter.tres = residual[n_mask.logical_not()].mean()
# meter.tsum = self.sum_mem[ids][n_mask.logical_not()].mean()
self.acc_precise_(self.pred_mem[ids].argmax(dim=1),
ys,
meter,
name='sacc')
return meter
def to_logits(self, xs) -> torch.Tensor:
return self.model(xs)
if __name__ == '__main__':
params = NoisyParams()
params.optim.args.lr = 0.06
params.epoch = 400
params.device = 'cuda:0'
params.filter_ema = 0.999
params.mixup = True
params.ideal_mixup = True
params.worst_mixup = False
params.noisy_ratio = 0.8
params.from_args()
params.initial()
trainer = MixupEvalTrainer(params)
trainer.train()
self.optim.zero_grad()
meter.Lall.backward()
self.optim.step()
self.acc_precise_(logits.argmax(dim=1),
ys,
meter=meter,
name='true_acc')
self.acc_precise_(logits.argmax(dim=1),
nys,
meter=meter,
name='noisy_acc')
return meter
if __name__ == '__main__':
params = NoisyParams()
params.ema = True # l2r have no ema for model
params.epoch = 120
params.batch_size = 100
params.device = 'cuda:3'
params.optim.args.lr = 0.1
params.meta_optim = {
'lr': 0.1,
'momentum': 0.9,
}
params.from_args()
trainer = L2RTrainer(params)
trainer.train()
# self.noisy_cls[self.noisy_cls >= 0.5].clamp_min_(params.gmm_w_sche(params.eidx))
m2 = self.acc_mixture_(true_cls, (self.count_mem >= 0).cpu().numpy(), pre='con')
meter.update(m)
self.logger.info(m2)
if __name__ == '__main__':
# for retry,params in enumerate(params.grid_range(5)):
# for retry in range(5):
# retry = retry + 1
params = NoisyParams()
params.right_n = 10
params.use_right_label = False
params.optim.args.lr = 0.06
params.epoch = 500
params.device = 'cuda:2'
params.filter_ema = 0.99
params.burnin = 20
params.mix_burnin = 20
params.targets_ema = 0.3
params.pred_thresh = 0.9
params.feature_mean = False
params.local_filter = True # 局部的筛选方法
params.mixt_ema = True # 是否对 BMM 的预测结果用 EMA 做平滑
params.from_args()
params.initial()
trainer = MultiHeadTrainer(params)
if params.ss_pretrain:
ckpt = torch.load(params.ss_pretrain_fn)
trainer.model.load_state_dict(ckpt)
self.noisy_cls = self.noisy_cls_mem.clone()
self.noisy_cls[self.noisy_cls < 0.5] = 0
# 随时间推移，越难以区分的样本越应该直接挂掉，而不是模糊来模糊去的加权（或许）
self.noisy_cls[self.noisy_cls >= 0.5].clamp_min_(
params.gmm_w_sche(params.eidx))
if __name__ == '__main__':
# for retry,params in enumerate(params.grid_range(5)):
# for retry in range(5):
# retry = retry + 1
params = NoisyParams()
params.right_n = 10
params.use_right_label = False
params.optim.args.lr = 0.06
params.epoch = 300
params.device = 'cuda:2'
params.filter_ema = 0.999
params.burnin = 2
params.gmm_burnin = 20
params.targets_ema = 0.3
# params.tolerance_type = 'exp'
params.pred_thresh = 0.9
# params.widen_factor = 10
params.from_args()
params.initial()
trainer = MultiHeadTrainer(params)
trainer.train()