Pytorch Dataloader内部随机数

pytorch的dataloader的__getitem__()方法，内部定义随机数，采用numpy.random.rand的方法，每次迭代从dataloader里取数据，都会产生拿到相同的随机数序列，相反，torch.rand的方法，每次迭代拿数据时，拿到的是不同的随机数序列。

	# numpy生成随机数
    def get_train_sample(self, ):
        track_idx = np.random.randint(0, len(self._tracks), (1, ))[0]
        separated = self._tracks[track_idx]
        patch_length = int(self._sample_length * self._sample_rate)
        if self._incoherent_rate <= np.random.rand(1)[0]:   # coherent remix
            start_t = np.random.randint(0, separated.shape[-1]-patch_length, (1, ))
            separated = separated[:, :, start_t: start_t+patch_length]
		return separated

	# torch生成随机数
    def get_train_sample(self, ):
        track_idx = torch.randint(0, len(self._tracks), (1,)).item()
        separated = self._tracks[track_idx]
        patch_length = int(self._sample_length * self._sample_rate)
        if self._incoherent_rate <= np.random.rand(1)[0]:   # coherent remix
            start_t = torch.randint(0, separated.shape[-1]-patch_length, (1, )).item()
            separated = separated[:, :, start_t: start_t+patch_length]
		return separated

用第一种方法生成的数据片段，每个loader里面都相同，用第二种方法，每个loader依旧是随机。