不同batch size时inference结果不同 about tanet HOT 1 CLOSED

您好！感谢您对这项工作的持续关注，刚才我试了一下用不同的batch size，在TAD66K上测试了推理的性能，结果如下：
# test_loader = DataLoader(test_ds, batch_size=64, num_workers=opt['num_workers'], shuffle=False)
# lcc_mean: 0.5266745930806195, srcc_mean: 0.5069025614121254, validate_losses: 0.015697210137563315

# test_loader = DataLoader(test_ds, batch_size=32, num_workers=opt['num_workers'], shuffle=False)
# lcc_mean: 0.5264128589068133, srcc_mean: 0.5066027710957562, validate_losses: 0.015706565844425466

test_loader = DataLoader(test_ds, batch_size=1, num_workers=opt['num_workers'], shuffle=False)
# lcc_mean: 0.5258770707185415, srcc_mean: 0.5068998155002807, validate_losses: 0.01672296673651432

inference的结果几乎没有差别，细微的差别可能来自于推理设备不同和网络中相关随机层参数没有被固定的原因。当时给公司做开发，所以用来推理的batch size等一些参数，都是甲方根据需求来定，一般batch size不会是1。