Python torch.initial_seed方法代码示例

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def worker_init_fn(worker_id):
    torch_seed = torch.initial_seed()

    random.seed(torch_seed + worker_id)

    if torch_seed >= 2**32:
        torch_seed = torch_seed % 2**32
    np.random.seed(torch_seed + worker_id) 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def worker_init_fn(pid):
    np.random.seed(torch.initial_seed() % (2**31-1)) 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def worker_init(worker_id):
    seed_all(torch.initial_seed() % 2**32) 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def init_np_seed(worker_id):
    seed = torch.initial_seed()
    np.random.seed(seed % 4294967296) 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def _worker_init_fn(worker_id: int) -> None:
    """Sets a unique but deterministic random seed for background workers.

    Only sets the seed for NumPy because PyTorch and Python's own RNGs
    take care of reseeding on their own.
    See https://github.com/numpy/numpy/issues/9650."""
    # Modulo 2**32 because np.random.seed() only accepts values up to 2**32 - 1
    initial_seed = torch.initial_seed() % 2**32
    worker_seed = initial_seed + worker_id
    np.random.seed(worker_seed)


# Be careful from where you call this! Not sure if this is concurrency-safe. 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def set_torch_seed_to_all_gens(_):
    seed = torch.initial_seed() % (2**32 - 1)
    random.seed(seed)
    np.random.seed(seed) 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def test_seed():
    torchfunc.seed(0)
    assert 0 == torch.initial_seed() 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def test_seed_context_manager():
    first_seed = torch.initial_seed()
    with torchfunc.seed(0):
        assert 0 == torch.initial_seed()
    assert torch.initial_seed() == first_seed 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def test_seed_decorator():
    first_seed = torch.initial_seed()

    @torchfunc.seed(0)
    def wrapped():
        assert 0 == torch.initial_seed()

    wrapped()
    assert torch.initial_seed() == first_seed 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def __init__(self, value, cuda: bool = False):
        self.value = value
        self.cuda = cuda

        self._last_seed = torch.initial_seed()
        np.random.seed(self.value)
        torch.manual_seed(self.value)

        if self.cuda:
            torch.backends.cudnn.deterministic = True
            torch.backends.cudnn.benchmark = False 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def __init__(self, sample_len, n_samples, xyz, sigma, b, r, dt=0.01, washout=0, normalize=False, seed=None):
        """
        Constructor
        :param sample_len: Length of the time-series in time steps.
        :param n_samples: Number of samples to generate.
        :param a:
        :param b:
        :param c:
        """
        # Properties
        self.sample_len = sample_len
        self.n_samples = n_samples
        self.xyz = xyz
        self.dt = dt
        self.normalize = normalize
        self.washout = washout
        self.sigma = sigma
        self.b = b
        self.r = r

        # Seed
        if seed is not None:
            torch.initial_seed(seed)
        # end if

        # Generate data set
        self.outputs = self._generate()
    # end __init__

    #############################################
    # OVERRIDE
    #############################################

    # Length 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def __init__(self, sample_len, n_samples, xy, a, b, washout=0, normalize=False, seed=None):
        """
        Constructor
        :param sample_len: Length of the time-series in time steps.
        :param n_samples: Number of samples to generate.
        :param a:
        :param b:
        :param c:
        """
        # Properties
        self.sample_len = sample_len
        self.n_samples = n_samples
        self.a = a
        self.b = b
        self.xy = xy
        self.normalize = normalize
        self.washout = washout

        # Seed
        if seed is not None:
            torch.initial_seed(seed)
        # end if

        # Generate data set
        self.outputs = self._generate()
    # end __init__

    #############################################
    # OVERRIDE
    #############################################

    # Length 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def __getitem__(self, item):
        if self.datapipeline is None:
            # build datapipeline with random seed the first time when __getitem__ is called
            # usually, dataset is already spawned (into subprocess) at this point.
            seed = (torch.initial_seed() + item * self._SEED_STEP +
                    self.ext_seed * self._EXT_SEED_STEP) % self._SEED_DIVIDER
            self.datapipeline = datapipeline_builder.build(self.task,
                                                           self.cfg,
                                                           seed=seed)
            logger.info("AdaptorDataset #%d built datapipeline with seed=%d" %
                        (item, seed))

        training_data = self.datapipeline[item]

        return training_data 

# 需要导入模块: import torch [as 别名]
# 或者: from torch import initial_seed [as 别名]
def init_random(seed: int = None):
    """
    Initializes the random generators to allow seeding.

    Args:
        seed (int): The seed used for all random generators.

    """
    global GLOBAL_SEED  # pylint: disable=global-statement
    if GLOBAL_SEED is not None:
        return

    if seed is None:
        tmp_random = numpy.random.RandomState(None)
        GLOBAL_SEED = tmp_random.randint(2**32-1, dtype='uint32')
    else:
        GLOBAL_SEED = seed

    # initialize random generators
    numpy.random.seed(GLOBAL_SEED)
    random.seed(GLOBAL_SEED)

    try:
        # try to load torch and initialize random generator if available
        import torch
        torch.cuda.manual_seed_all(GLOBAL_SEED)  # gpu
        torch.manual_seed(GLOBAL_SEED)  # cpu
    except ImportError:
        pass

    try:
        # try to load tensorflow and initialize random generator if available
        import tensorflow
        tensorflow.random.set_random_seed(GLOBAL_SEED)
    except ImportError:
        pass

    # check whether all calls to torch.* use the same random generator (i.e. same instance)
    # works in a short test -- MS
    # print(torch.initial_seed())

    # logger.info("Seed is {:d}".format(GLOBAL_SEED))
    return GLOBAL_SEED