PyTorch 里面处理的最基本的操作对象就是 Tensor(张量),表示的是一个多维的矩阵,比如零维就是一个点,一维就是向量,二维就是一般的矩阵,多维就相当于一个多维的数组,这和 numpy 是对应的,而且 PyTorch 的 tensor 可以和 numpy 的 ndarray 相互转换,
唯一不同的是 PyTorch 可以在 GPU 上运行
而 numpy 的 ndarray 只能在 CPU 上运行
前言
我们先介绍一下一些常用的不同数据类型的Tensor( torch.Tensor默认的是 torch.FloatTensor数据类型):
- 16位整型 torch.ShortTensor
- 32位整型 torch.IntTensor
- 64位整型 torch.LongTensor
- 16位浮点型 torch.HalfTensor
- 32位浮点型 torch.FloatTensor
- 64位浮点型 torch.DoubleTensor
- torch.Tensor 默认 torch.FloatTensor
类型(Types)
Numpy | PyTorch |
---|
np.ndarray | torch.Tensor |
np.float32 | torch.float32; torch.float |
np.float64 | torch.float64; torch.double |
np.float16 | torch.float16; torch.half |
np.int8 | torch.int8 |
np.uint8 | torch.uint8 |
np.int16 | torch.int16; torch.short |
np.int32 | torch.int32; torch.int |
np.int64 | torch.int64; torch.long |
索引
Numpy | PyTorch |
---|
x[0] | x[0] |
x[:, 0] | x[:, 0] |
x[indices] | x[indices] |
np.take(x, indices) | torch.take(x, torch.LongTensor(indices)) |
x[x != 0] | x[x != 0] |
ones and zeros
Numpy | PyTorch |
---|
np.empty((2, 3)) | torch.empty(2, 3) |
np.empty_like(x) | torch.empty_like(x) |
np.eye | torch.eye |
np.identity | torch.eye |
np.ones | torch.ones |
np.ones_like | torch.ones_like |
np.zeros | torch.zeros |
np.zeros_like | torch.zeros_like |
从已知数据构造
Numpy | PyTorch |
---|
np.array([[1, 2], [3, 4]]) | torch.tensor([[1, 2], [3, 4]]) |
np.array([3.2, 4.3], dtype=np.float16) , np.float16([3.2, 4.3]) | torch.tensor([3.2, 4.3], dtype=torch.float16) |
x.copy() | x.clone() |
np.fromfile(file) | torch.tensor(torch.Storage(file)) |
np.frombuffer | |
np.fromfunction | |
np.fromiter | |
np.fromstring | |
np.load | torch.load |
np.loadtxt | |
np.concatenate | torch.cat |
数值范围
Numpy | PyTorch |
---|
np.arange(10) | torch.arange(10) |
np.arange(2, 3, 0.1) | torch.arange(2, 3, 0.1) |
np.linspace | torch.linspace |
np.logspace | torch.logspace |
构造矩阵
Numpy | PyTorch |
---|
np.diag | torch.diag |
np.tril | torch.tril |
np.triu | torch.triu |
参数
Numpy | PyTorch |
---|
x.shape | x.shape |
x.strides | x.stride() |
x.ndim | x.dim() |
x.data | x.data |
x.size | x.nelement() |
x.dtype | x.dtype |
形状(Shape)变换
Numpy | PyTorch |
---|
x.reshape | x.reshape; x.view |
x.resize() | x.resize_ |
| x.resize_as_ |
x.transpose | x.transpose or x.permute |
x.flatten | x.view(-1) |
x.squeeze() | x.squeeze() |
x[:, np.newaxis]; np.expand_dims(x, 1) | x.unsqueeze(1) |
数据选择
Numpy | PyTorch |
---|
np.put | |
x.put | x.put_ |
x = np.array([1, 2, 3]) x.repeat(2) # [1, 1, 2, 2, 3, 3] | x = torch.tensor([1, 2, 3]) x.repeat(2) # [1, 2, 3, 1, 2, 3] x.repeat(2).reshape(2, -1).transpose(1, 0).reshape(-1) # [1, 1, 2, 2, 3, 3] |
np.tile(x, (3, 2)) | x.repeat(3, 2) |
np.choose | |
np.sort sorted, indices = torch.sort(x, [dim]) | |
np.argsort sorted, indices = torch.sort(x, [dim]) | |
np.nonzero | torch.nonzero |
np.where | torch.where |
x[::-1] | |
数值计算
Numpy | PyTorch |
---|
x.min | x.min |
x.argmin | x.argmin |
x.max | x.max |
x.argmax | x.argmax |
x.clip | x.clamp |
x.round | x.round |
np.floor(x) | torch.floor(x); x.floor() |
np.ceil(x) | torch.ceil(x); x.ceil() |
x.trace | x.trace |
x.sum | x.sum |
x.cumsum | x.cumsum |
x.mean | x.mean |
x.std | x.std |
x.prod | x.prod |
x.cumprod | x.cumprod |
x.all | (x == 1).sum() == x.nelement() |
x.any | (x == 1).sum() > 0 |
数值比较
Numpy | PyTorch |
---|
np.less | x.lt |
np.less_equal | x.le |
np.greater | x.gt |
np.greater_equal | x.ge |
np.equal | x.eq |
np.not_equal | x.ne |