About your mixup method.

[Lacunainfo]

the method 'mixup' you called, it seems like channel-shuffle.

Can you guide me what you exactly meant for?

def mixup(lq, gt, alpha=1.2):
    if random.random() < 0.5:
       return lq, gt

    v = np.random.beta(alpha, alpha)
    r_index = torch.randperm(lq.size(0)).to(gt.device)

    lq = v * lq + (1 - v) * lq[r_index, :]
    gt = v * gt + (1 - v) * gt[r_index, :]

    return lq, gt

[Zdafeng]

The dimension of input is [B, C, H, W], so 'mixup' is performed in batch dimension.