Dev flow.utils.data part2

[Flowingsun007]

dataloader最小实现（part2）

• pr-part1见：#5406
• 原始pr见：#5074
  

---

此pr主要改动：

1.导出dataloader，dataset基础接口

外部可以直接使用导出后的接口，如，flow.utils.data.Dataset,flow.utils.data.DataLoader等：

dataset = flow.utils.data.TensorDataset(features, labels)
data_iter = flow.utils.data.DataLoader(dataset, batch_size, shuffle=True, num_workers=0)

2.引入torchvision包下的部分datasets和transforms转换工具

torchvision下的transforms和datasets两个文件夹中的部分常用功能，譬如：

• datasets引入了cv里常用的demo数据集——mnist/fashion-mnist/cifar10/cifar100，支持对这些数据集的自动下载，解压格式处理；
• transforms主要是对dataloader加载后的数据进行totensor，resize，normalize等操作。

在多个网络/数据集上测试，数据均可以正常加载，训练loss可正常下降。test case在oneflow/python/utils/vision/：

• test_mlp_flow.py # 测试fashion-mnist
• test_lenet_flow.py # 测试fashion-mnist
• test_cifar_flow.py # 测试cifar10

(oneflow-dev) [luyang@oneflow-15 vision]$ python test_mlp_flow.py
epoch 1, loss 0.0034, train acc 0.714, test acc 0.784, cost >>>>>>> 132.84298586845398(s)
epoch 2, loss 0.0022, train acc 0.809, test acc 0.818, cost >>>>>>> 130.11517643928528(s)
epoch 3, loss 0.0019, train acc 0.829, test acc 0.793, cost >>>>>>> 131.28324055671692(s)
epoch 4, loss 0.0018, train acc 0.841, test acc 0.796, cost >>>>>>> 130.2774519920349(s)

(oneflow-dev) [luyang@oneflow-15 vision]$ python test_lenet_flow.py
training on  cuda:0
epoch 1, loss 1.3111, train acc 0.515, test acc 0.684, time 129.4 sec
epoch 2, loss 0.6790, train acc 0.741, test acc 0.745, time 141.3 sec
epoch 3, loss 0.5768, train acc 0.780, test acc 0.782, time 167.0 sec
epoch 4, loss 0.5070, train acc 0.813, test acc 0.800, time 132.8 sec
epoch 5, loss 0.4620, train acc 0.829, test acc 0.826, time 134.6 sec

注：test case由于训练比较慢，暂时没有移入test/dataloader作为CI的test case，后续看情况是否需要跑CI

[github-actions]

CI failed, removing label automerge

[github-actions]

Speed stats:

GPU Name: GeForce GTX 1080 

PyTorch resnet50 time: 139.6ms (= 6980.0ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 124.6ms (= 6230.2ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
Relative speed: 1.12 (= 139.6ms / 124.6ms)

PyTorch resnet50 time: 82.7ms (= 4137.5ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 74.3ms (= 3714.1ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
Relative speed: 1.11 (= 82.7ms / 74.3ms)

PyTorch resnet50 time: 58.4ms (= 2922.5ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 57.2ms (= 2857.6ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
Relative speed: 1.02 (= 58.4ms / 57.2ms)

PyTorch resnet50 time: 48.2ms (= 2411.1ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 51.7ms (= 2586.7ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
Relative speed: 0.93 (= 48.2ms / 51.7ms)

PyTorch resnet50 time: 43.0ms (= 2148.6ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 53.4ms (= 2670.1ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
Relative speed: 0.80 (= 43.0ms / 53.4ms)