pytorch学习手册【一】

 

一、关于Tensors的一些判断

torch.is_tensor(obj)

torch.is_storage(obj)

torch.set_default_dtype(d) 默认的type为torch.float32

torch.get_default_dtype() → torch.dtype 

torch.set_default_tensor_type(t)

torch.numel(input) → int  返回tensor中所有的元素个数

torch.set_printoptions(precision=None, threshold=None, edgeitems=None, linewidth=None, profile=None)

设置print的相关选项

Parameters:

precision – Number of digits of precision for floating point output (default = 4). threshold – Total number of array elements which trigger summarization rather than full repr(default = 1000). edgeitems – Number of array items in summary at beginning and end of each dimension (default = 3). linewidth – The number of characters per line for the purpose of inserting line breaks (default = 80). Thresholded matrices will ignore this parameter. profile – Sane defaults for pretty printing. Can override with any of the above options. (any one of default, short, full)

torch.set_flush_denormal(mode) → bool

Parameters:	mode (bool) – Controls whether to enable flush denormal mode or not

 

二、创建tensor的一些方法

注意：张量的随机创建会在下面的random sampling里面再说明。

torch.tensor(data, dtype=None, device=None, requires_grad=False) → Tensor

torch.sparse_coo_tensor(indices, values, size=None, dtype=None, device=None, requires_grad=False)→ Tensor

torch.as_tensor(data, dtype=None, device=None) → Tensor

torch.from_numpy(ndarray) → Tensor

torch.zeros(*sizes, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False)→ Tensor

torch.zeros_like(input, dtype=None, layout=None, device=None, requires_grad=False) → Tensor

torch.ones(*sizes, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False)→ Tensor

torch.ones_like(input, dtype=None, layout=None, device=None, requires_grad=False) → Tensor

torch.arange(start=0, end, step=1, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False) → Tensor

torch.range(start=0, end, step=1, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False) → Tensor

torch.linspace(start, end, steps=100, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False) → Tensor

torch.logspace(start, end, steps=100, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False) → Tensor

torch.eye(n, m=None, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False)→ Tensor

torch.empty(*sizes, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False)→ Tensor

torch.empty_like(input, dtype=None, layout=None, device=None, requires_grad=False) → Tensor

torch.full(size, fill_value, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False) → Tensor

torch.full_like(input, fill_value, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False) → Tensor

三、Indexing, Slicing, Joining, Mutating(转变/变化)

torch.cat(tensors, dim=0, out=None) → Tensor

torch.chunk(tensor, chunks, dim=0) → List of Tensors。在某一个维度将一个tensor分成几等份，chunks为int，即需要分成的份数

torch.gather(input, dim, index, out=None) → Tensor。Gathers values along an axis specified by dim.

torch.index_select(input, dim, index, out=None) → Tensor，类似于标准库slice函数的作用

torch.masked_select(input, mask, out=None) → Tensor

torch.narrow(input, dimension, start, length) → Tensor

torch.nonzero(input, out=None) → LongTensor，返回所有非零元素的位置索引，返回的是索引哦

torch.reshape(input, shape) → Tensor

torch.split(tensor, split_size_or_sections, dim=0)

torch.squeeze(input, dim=None, out=None) → Tensor，将维度=1的那个维度（即只包含一个元素的维度）去掉，即所谓的压榨

torch.stack(seq, dim=0, out=None) → Tensor

torch.t(input) → Tensor

torch.take(input, indices) → Tensor

torch.transpose(input, dim0, dim1) → Tensor

torch.unbind(tensor, dim=0) → seq

torch.unsqueeze(input, dim, out=None) → Tensor

torch.where(condition, x, y) → Tensor

四、Random sampling建立随机矩阵

torch.manual_seed(seed)
torch.initial_seed()

torch.get_rng_state()

torch.set_rng_state(new_state)

torch.default_generator = <torch._C.Generator object>

torch.bernoulli(input, *, generator=None, out=None) → Tensor

torch.multinomial(input, num_samples, replacement=False, out=None) → LongTensor

torch.normal()

torch.normal(mean, std, out=None) → Tensor

torch.rand(*sizes, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False)→ Tensor

torch.rand_like(input, dtype=None, layout=None, device=None, requires_grad=False) → Tensor

torch.randint(low=0, high, size, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False) → Tensor

torch.randint_like(input, low=0, high, dtype=None, layout=torch.strided, device=None, requires_grad=False) → Tensor

torch.randn(*sizes, out=None, dtype=None, layout=torch.strided, device=None, requires_grad=False)→ Tensor

torch.randn_like(input, dtype=None, layout=None, device=None, requires_grad=False) → Tensor

torch.randperm(n, out=None, dtype=torch.int64, layout=torch.strided, device=None, requires_grad=False) → LongTensor

In-place random sampling

There are a few more in-place random sampling functions defined on Tensors as well. Click through to refer to their documentation:

• torch.Tensor.bernoulli_() - in-place version of torch.bernoulli()
• torch.Tensor.cauchy_() - numbers drawn from the Cauchy distribution
• torch.Tensor.exponential_() - numbers drawn from the exponential distribution
• torch.Tensor.geometric_() - elements drawn from the geometric distribution
• torch.Tensor.log_normal_() - samples from the log-normal distribution
• torch.Tensor.normal_() - in-place version of torch.normal()
• torch.Tensor.random_() - numbers sampled from the discrete uniform distribution
• torch.Tensor.uniform_() - numbers sampled from the continuous uniform distribution

五、Serialization序列化

torch.save(obj, f, pickle_module=<module 'pickle' from '/private/home/soumith/anaconda3/lib/python3.6/pickle.py'>, pickle_protocol=2)

将一个tensor保存带磁盘

Parameters:	obj – saved object f – a file-like object (has to implement write and flush) or a string containing a file name pickle_module – module used for pickling metadata and objects pickle_protocol – can be specified to override the default protocol

torch.load(f, map_location=None, pickle_module=<module 'pickle' from '/private/home/soumith/anaconda3/lib/python3.6/pickle.py'>)

六、Parallelism并行

torch.get_num_threads() → int

Gets the number of OpenMP threads used for parallelizing CPU operations

torch.set_num_threads(int)

Sets the number of OpenMP threads used for parallelizing CPU operations

七、设置是否可求梯度Locally disabling gradient computation

 torch.no_grad()

torch.enable_grad()

torch.set_grad_enabled()

八、Math operations数学运算操作

8.1 逐点运算Pointwise Ops

torch.abs(input, out=None) → Tensor

torch.acos(input, out=None) → Tensor

torch.add()

torch.add(input, value, out=None)

torch.add(input, value=1, other, out=None)

torch.addcdiv(tensor, value=1, tensor1, tensor2, out=None) → Tensor。首先求tensor1除以tensor2，然后用得到的结果乘以value，然后再加到tensor上面去。

torch.addcmul(tensor, value=1, tensor1, tensor2, out=None) → Tensor。首先求tensor1乘以ensor2，然后用得到的结果乘以value，然后再加到tensor上面去。

torch.asin(input, out=None) → Tensor

torch.atan(input, out=None) → Tensor

torch.atan2(input1, input2, out=None) → Tensor

torch.ceil(input, out=None) → Tensor

torch.clamp(input, min, max, out=None) → Tensor。将tensor中所有小于min的数字用min代替，所有大于max的数字用max代替

torch.clamp(input, *, min, out=None) → Tensor

torch.clamp(input, *, max, out=None) → Tensor

torch.cos(input, out=None) → Tensor

torch.cosh(input, out=None) → Tensor

torch.div()

torch.div(input, value, out=None) → Tensor

torch.div(input, other, out=None) → Tensor

torch.digamma(input, out=None) → Tensor

torch.erf(tensor, out=None) → Tensor

torch.erfc(input, out=None) → Tensor

torch.erfinv(input, out=None) → Tensor

torch.exp(input, out=None) → Tensor

torch.expm1(input, out=None) → Tensor

torch.floor(input, out=None) → Tensor

torch.fmod(input, divisor, out=None) → Tensor

torch.frac(input, out=None) → Tensor

torch.lerp(start, end, weight, out=None)

torch.log(input, out=None) → Tensor

torch.log10(input, out=None) → Tensor

torch.log1p(input, out=None) → Tensor

torch.log2(input, out=None) → Tensor

torch.mul()

torch.mul(input, value, out=None)

torch.mul(input, other, out=None)

torch.mvlgamma(input, p) → Tensor

torch.neg(input, out=None) → Tensor

torch.pow()

torch.pow(input, exponent, out=None) → Tensor

torch.pow(base, input, out=None) → Tensor

torch.reciprocal(input, out=None) → Tensor

torch.remainder(input, divisor, out=None) → Tensor

torch.fmod(),

torch.rsqrt(input, out=None) → Tensor

torch.sigmoid(input, out=None) → Tensor

torch.sign(input, out=None) → Tensor

torch.sin(input, out=None) → Tensor

torch.sinh(input, out=None) → Tensor

torch.sqrt(input, out=None) → Tensor

torch.tan(input, out=None) → Tensor

torch.tanh(input, out=None) → Tensor

torch.trunc(input, out=None) → Tensor