fix the modification of set_expected_place

[zhiqiu]

PR types

Bug fixes

PR changes

Others

Describe

revert the modification of set_expected_place

I add cudaSetDevice() in SetExpectedPlace in dygraph mode in PR #30338, and the logic is updated in #30868.

This PR revert the changes since in multi-process dataloader, using CUDA APIs in subprocess may result in error since the
cuda runtime is not initialized.

So this PR revert that changes, and the problem in PR #30338 can be fixed by adding DeviceGuard in SetTensorFromPyArrayT, which is already done in #30868.

Reproduce code:

import os
import numpy as np
import paddle
import paddle.nn.functional as F
from paddle.io import Dataset, DataLoader, DistributedBatchSampler

class FakeDataset(Dataset):
    def __init__(self, num_samples=1000):
        self.num_samples = num_samples

    def __getitem__(self, idx):
        data1 = np.random.rand(3, 32, 32).astype("float32")
        data2 = np.random.rand(3, 32, 32).astype("float32")
        return [data1, data2]
        
    def __len__(self):
        return self.num_samples


def collate_fn(batch):
    with paddle.fluid.dygraph.guard(place=paddle.CPUPlace()):
        width_list = np.array(
            [image_set_i[0].shape[2] for image_set_i in batch])
        avg_width = int(width_list.mean() / 8 / 2) * 8
        r_imgs = [[] for _ in batch[0]]
        for img_set in batch:
            for j, img in enumerate(img_set):
                img = paddle.to_tensor(
                    np.expand_dims(img, 0), place=paddle.CPUPlace())
                img = F.upsample(img, size=(32, avg_width), mode='bilinear')
                img = paddle.squeeze(img, axis=0)
                r_imgs[j].append(img.numpy())
    return r_imgs

if __name__ == "__main__":
    place = paddle.set_device('gpu')
    dataset = FakeDataset()
    num_workers = 1

    loader = DataLoader(
            dataset,
            batch_size=8,
            collate_fn=collate_fn,
            places=place,
            return_list=True,
            num_workers=num_workers)

    print("finish init dataloader")
    for batch in loader:
        for data in batch:
            print(data.shape)
        exit()
    

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[wangxicoding]

LGTM