How to deal with quantization operation during training DCVC-TCM?

[SYSU-HQH]

Thanks for sharing your great works!
I ran DCVC-TCM with the released weight without update them, and added noise sampled from a uniform distribution to test the performance of the released weight using the training set(Vimeo-90K) mentioned in the paper, like this:

def quant(self, x):
if self.training:
return x + torch.nn.init.uniform_(torch.zeros_like(x), -0.5, 0.5)
else:
return torch.round(x)

However, I found the loss of adding uniform noise(0.295bpp with 39.335psnr) is quite larger than using round operation(0.112bpp with 42.291psnr).
Is this a normal phenomenon? Or did I use the wrong quantization processing method？