why with torch.no_grad() when i == 0:

[fido20160817]

 if i == 0:
    with torch.no_grad():
        log_p, logdet, _ = model.module(
            image + torch.rand_like(image) / n_bins
        )

        continue

else:
    log_p, logdet, _ = model(image + torch.rand_like(image) / n_bins)

[davitpapikyan]

This is for data dependent initialization of Actnorm.