2面向 Numpy 用户的 PyTorch 速查表.md

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

构造器（Constructor）

零和一（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

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]

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

pytorch与tensorflow API速查表

方法名称	pytroch	tensorflow	numpy
裁剪	torch.clamp(x, min, max)	tf.clip_by_value(x, min, max)	np.clip(x, min, max)
取最小值	torch.min(x, dim)[0]	tf.min(x, axis)	np.min(x , axis)
取两个tensor的最大值	torch.max(x, y)	tf.maximum(x, y)	np.maximum(x, y)
取两个tensor的最小值	torch.min(x, y)	torch.minimum(x, y)	np.minmum(x, y)
取最大值索引	torch.max(x, dim)[1]	tf.argmax(x, axis)	np.argmax(x, axis)
取最小值索引	torch.min(x, dim)[1]	tf.argmin(x, axis)	np.argmin(x, axis)
比较(x > y)	torch.gt(x, y)	tf.greater(x, y)	np.greater(x, y)
比较(x < y)	torch.le(x, y)	tf.less(x, y)	np.less(x, y)
比较(x==y)	torch.eq(x, y)	tf.equal(x, y)	np.equal(x, y)
比较(x!=y)	torch.ne(x, y)	tf.not_equal(x, y)	np.not_queal(x , y)
取符合条件值的索引	torch.nonzero(cond)	tf.where(cond)	np.where(cond)
多个tensor聚合	torch.cat([x, y], dim)	tf.concat([x,y], axis)	np.concatenate([x,y], axis)
堆叠成一个tensor	torch.stack([x1, x2], dim)	tf.stack([x1, x2], axis)	np.stack([x, y], axis)
tensor切成多个tensor	torch.split(x1, split_size_or_sections, dim)	tf.split(x1, num_or_size_splits, axis)	np.split(x1, indices_or_sections, axis)
-	torch.unbind(x1, dim)	tf.unstack(x1,axis)	NULL
随机扰乱	torch.randperm(n) 1	tf.random_shuffle(x)	np.random.shuffle(x) 2 np.random.permutation(x ) 3
前k个值	torch.topk(x, n, sorted, dim)	tf.nn.top_k(x, n, sorted)	NULL

1. 该方法只能对0~n-1自然数随机扰乱，所以先对索引随机扰乱，然后再根据扰乱后的索引取相应的数据得到扰乱后的数据
2. 该方法会修改原值，没有返回值
3. 该方法不会修改原值，返回扰乱后的值