[Dy2stat] Fix test_mobile_net random failed

[Aurelius84]

PR types

Others

PR changes

Others

Describe

Fix test_mobile_net random failed using local_random = np.random.RandomState(SEED) instead of np.random.seed(SEED).

I found that the numpy.random outputs unstable random data when run twice independent if wrapping it into DataLoader.

So，the random behavior of numpy in DataLoader is not stable or unsafe. We should use local_random = np.random.RandomState(SEED) instead of it.

See reason in detail: #24727

The repetition code as followed：

1. It's ok if only use numpy.random

import time
import numpy as np
import unittest

SEED = 2020

def fake_data_reader(batch_size, label_size):
    def reader():
        batch_data = []
        while True:
            img = np.random.random([3, 224, 224]).astype('float32')
            label = np.random.randint(0, label_size, [1]).astype('int64')
            
            batch_data.append([img, label])
            if len(batch_data) == batch_size:
                yield batch_data
                batch_data = []

    return reader

def load_data():
    np_data = []
    np.random.seed(SEED)

    reader = fake_data_reader(batch_size=1, label_size=3)

    for data in reader():
        img, label = data[0]

        np_data.append(img[0,0,:6].tolist())
        break
        
    return np.array(np_data)


class TestNumpyRandom(unittest.TestCase):
    def test_data(self):
        for i in range(100):
            data1 = load_data()
            data2 = load_data()
            # print(data1)
            # print(data2)
            self.assertTrue(np.array_equal(data1, data2), msg="data1\n: {}\n, data2:\n {}\n".format(data1, data2))

if __name__ == '__main__':
    unittest.main()

**2. It failed if only use numpy.random + DataLoader. But when I add time.sleep(1), it's ok. **

import time
import numpy as np
import paddle.fluid as fluid
import unittest


SEED = 2020


def fake_data_reader(batch_size, label_size):
    def reader():
        batch_data = []
        while True:
            img = np.random.random([3, 224, 224]).astype('float32')
            label = np.random.randint(0, label_size, [1]).astype('int64')
            
            batch_data.append([img, label])
            if len(batch_data) == batch_size:
                yield batch_data
                batch_data = []

    return reader

def train():

    place = fluid.CPUPlace()
    with fluid.dygraph.guard(place):
        np.random.seed(SEED)

        train_reader = fake_data_reader(4, 3)
        train_data_loader = fluid.io.DataLoader.from_generator(capacity=16)
        train_data_loader.set_sample_list_generator(train_reader, place)
        
        np_data = []
        for img, label in train_data_loader():

            img_data = img.numpy()[0,0,0,:6].tolist()
            print("img: ", img_data)
            print("label: ", label.numpy().flatten().tolist())

            np_data.append(img_data)
            time.sleep(1)
            break
            
    return np.array(np_data)

class testNumpyRandom2(unittest.TestCase):
    def test_data(self):
        for i in range(10):
            print("data1: ")
            data1 = train()
            print("data2:")
            data2 = train()
            self.assertTrue(np.array_equal(data1, data2), msg="data1\n: {}\n, data2:\n {}\n".format(data1, data2))

if __name__ == '__main__':
    unittest.main()

[zhhsplendid]

LGTM