PyTorch中验证集与测试集DataLoader是否应开启shuffle？

我认为答案是否定的，应该设置shuffle=False。

官方教程提到：‘训练模型时，我们通常以“小批量”传递样本，每个epoch重新打乱数据以减少模型过拟合’——这说明只有训练集需要设置shuffle=True，但官方示例里却把验证集和测试集也都设成了True，这让我有点困惑，希望能得到明确指导。

我现在遇到的问题是：把测试集的shuffle设为True后，测试准确率低得离谱，训练和测试的指标如下：

Epoch: 1 | Train Loss: 2.1650 | Train Acc: 0.4683 | Test Loss: 6.2742 | Test Acc: 0.0349
Epoch: 2 | Train Loss: 1.2008 | Train Acc: 0.6624 | Test Loss: 6.9259 | Test Acc: 0.0359
Epoch: 3 | Train Loss: 0.9474 | Train Acc: 0.7325 | Test Loss: 7.3948 | Test Acc: 0.0365
Epoch: 4 | Train Loss: 0.7897 | Train Acc: 0.7808 | Test Loss: 7.7263 | Test Acc: 0.0352
Epoch: 5 | Train Loss: 0.6799 | Train Acc: 0.8186 | Test Loss: 8.0628 | Test Acc: 0.0377
Epoch: 6 | Train Loss: 0.5953 | Train Acc: 0.8374 | Test Loss: 8.3828 | Test Acc: 0.0351
Epoch: 7 | Train Loss: 0.5269 | Train Acc: 0.8647 | Test Loss: 8.5854 | Test Acc: 0.0351
Epoch: 8 | Train Loss: 0.4693 | Train Acc: 0.8791 | Test Loss: 8.9052 | Test Acc: 0.0337
Epoch: 9 | Train Loss: 0.4234 | Train Acc: 0.8955 | Test Loss: 9.1290 | Test Acc: 0.0354

我的代码片段如下：

train_ds = ImageFolder(
    root=TRAIN_DIR,
    transform=pretrained_weights.transforms(),
)

test_ds = ImageFolder(
    root=TEST_DIR,
    transform=pretrained_weights.transforms(),
)

train_dl = DataLoader(
    dataset=train_ds, batch_size=BATCH_SIZE, shuffle=True
)

test_dl = DataLoader(
    dataset=test_ds, batch_size=BATCH_SIZE, shuffle=True
)

内容的提问来源于stack exchange，提问作者dimButTries