Fix swin dataloader import bug

scripts/swin_dataloader_compare_speed_with_pytorch.py

@@ -3,13 +3,6 @@
 import numpy as np
 import argparse
 
-import oneflow as flow
-from oneflow.utils.data import DataLoader
-
-from flowvision import datasets, transforms
-from flowvision.data import create_transform
-
-
 ONEREC_URL = "https://oneflow-static.oss-cn-beijing.aliyuncs.com/ci-files/dataset/nanodataset.zip"
 MD5 = "7f5cde8b5a6c411107517ac9b00f29db"
 
@@ -64,99 +57,84 @@ def ensure_dataset():
     shutil.unpack_archive(absolute_file_path)
     return str(pathlib.Path.cwd() / "nanodataset")
 
+
 swin_dataloader_loop_count = 200
 
+
 def print_rank_0(*args, **kwargs):
     rank = int(os.getenv("RANK", "0"))
     if rank == 0:
         print(*args, **kwargs)
 
 
-class SubsetRandomSampler(flow.utils.data.Sampler):
-    r"""Samples elements randomly from a given list of indices, without replacement.
-    Arguments:
-        indices (sequence): a sequence of indices
-    """
-
-    def __init__(self, indices):
-        self.epoch = 0
-        self.indices = indices
-
-    def __iter__(self):
-        return (self.indices[i] for i in flow.randperm(len(self.indices)).tolist())
-
-    def __len__(self):
-        return len(self.indices)
-
-    def set_epoch(self, epoch):
-        self.epoch = epoch
-
-
-def build_transform():
-    # this should always dispatch to transforms_imagenet_train
-    transform = create_transform(
-        input_size=224,
-        is_training=True,
-        color_jitter=0.4,
-        auto_augment="rand-m9-mstd0.5-inc1",
-        re_prob=0.25,
-        re_mode="pixel",
-        re_count=1,
-        interpolation="bicubic",
-    )
-    return transform
-
-
-# swin-transformer imagenet dataloader
-def build_dataset(imagenet_path):
-    transform = build_transform()
-    prefix = "train"
-    root = os.path.join(imagenet_path, prefix)
-    dataset = datasets.ImageFolder(root, transform=transform)
-    return dataset
-
-
-def build_loader(imagenet_path, batch_size, num_wokers):
-    dataset_train = build_dataset(imagenet_path=imagenet_path)
-
-    indices = np.arange(
-        flow.env.get_rank(), len(dataset_train), flow.env.get_world_size()
-    )
-    sampler_train = SubsetRandomSampler(indices)
-
-    data_loader_train = DataLoader(
-        dataset_train,
-        sampler=sampler_train,
-        batch_size=batch_size,
-        num_workers=num_wokers,
-        drop_last=True,
-    )
-
-    return dataset_train, data_loader_train
-
-
 def run(mode, imagenet_path, batch_size, num_wokers):
     if mode == "torch":
         import torch as flow
         from torch.utils.data import DataLoader
 
         from torchvision import datasets, transforms
         from timm.data import create_transform
-
-    dataset_train, data_loader_train = build_loader(
-        args.imagenet_path, args.batch_size, args.num_workers
-    )
-    data_loader_train_iter = iter(data_loader_train)
-
-    # warm up
-    for idx in range(5):
-        samples, targets = data_loader_train_iter.__next__()
-
-    start_time = time.time()
-    for idx in range(swin_dataloader_loop_count):
-        samples, targets = data_loader_train_iter.__next__()
-    total_time = time.time() - start_time
-    return total_time
+    else:
+        import oneflow as flow
+        from oneflow.utils.data import DataLoader
+
+        from flowvision import datasets, transforms
+        from flowvision.data import create_transform
+
+    def build_transform():
+        # this should always dispatch to transforms_imagenet_train
+        transform = create_transform(
+            input_size=224,
+            is_training=True,
+            color_jitter=0.4,
+            auto_augment="rand-m9-mstd0.5-inc1",
+            re_prob=0.25,
+            re_mode="pixel",
+            re_count=1,
+            interpolation="bicubic",
+        )
+        return transform
+
+    # swin-transformer imagenet dataloader
+    def build_dataset(imagenet_path):
+        transform = build_transform()
+        prefix = "train"
+        root = os.path.join(imagenet_path, prefix)
+        dataset = datasets.ImageFolder(root, transform=transform)
+        return dataset
+
+    def build_loader(imagenet_path, batch_size, num_wokers):
+        dataset_train = build_dataset(imagenet_path=imagenet_path)
+
+        indices = np.arange(0, len(dataset_train), 1)
+
+        data_loader_train = DataLoader(
+            dataset_train,
+            shuffle=True,
+            batch_size=batch_size,
+            num_workers=num_wokers,
+            drop_last=True,
+        )
+
+        return dataset_train, data_loader_train
+
+    def get_time():
+        dataset_train, data_loader_train = build_loader(
+            args.imagenet_path, args.batch_size, args.num_workers
+        )
+        data_loader_train_iter = iter(data_loader_train)
+
+        # warm up
+        for idx in range(5):
+            samples, targets = data_loader_train_iter.__next__()
+
+        start_time = time.time()
+        for idx in range(swin_dataloader_loop_count):
+            samples, targets = data_loader_train_iter.__next__()
+        total_time = time.time() - start_time
+        return total_time
+
+    return get_time()
 
 
 if __name__ == "__main__":
@@ -173,11 +151,21 @@ def run(mode, imagenet_path, batch_size, num_wokers):
     pytorch_data_loader_total_time = run(
         "torch", args.imagenet_path, args.batch_size, args.num_workers
     )
-    oneflow_data_loader_time = oneflow_data_loader_total_time / swin_dataloader_loop_count
-    pytorch_data_loader_time = pytorch_data_loader_total_time / swin_dataloader_loop_count
+    oneflow_data_loader_time = (
+        oneflow_data_loader_total_time / swin_dataloader_loop_count
+    )
+    pytorch_data_loader_time = (
+        pytorch_data_loader_total_time / swin_dataloader_loop_count
+    )
 
     relative_speed = oneflow_data_loader_time / pytorch_data_loader_time
 
-    print_rank_0(f"OneFlow swin dataloader time: {oneflow_data_loader_time:.3f}s (= {oneflow_data_loader_total_time:.3f}s / {swin_dataloader_loop_count}, num_workers={args.num_workers})")
-    print_rank_0(f"PyTorch swin dataloader time: {pytorch_data_loader_time:.3f}s (= {pytorch_data_loader_total_time:.3f}s / {swin_dataloader_loop_count}, num_workers={args.num_workers})")
-    print_rank_0(f"Relative speed: {pytorch_data_loader_time / oneflow_data_loader_time:.3f} (= {pytorch_data_loader_time:.3f}s / {oneflow_data_loader_time:.3f}s)")
+    print_rank_0(
+        f"OneFlow swin dataloader time: {oneflow_data_loader_time:.3f}s (= {oneflow_data_loader_total_time:.3f}s / {swin_dataloader_loop_count}, num_workers={args.num_workers})"
+    )
+    print_rank_0(
+        f"PyTorch swin dataloader time: {pytorch_data_loader_time:.3f}s (= {pytorch_data_loader_total_time:.3f}s / {swin_dataloader_loop_count}, num_workers={args.num_workers})"
+    )
+    print_rank_0(
+        f"Relative speed: {pytorch_data_loader_time / oneflow_data_loader_time:.3f} (= {pytorch_data_loader_time:.3f}s / {oneflow_data_loader_time:.3f}s)"
+    )