DEV Community

Super Kai (Kazuya Ito)
Super Kai (Kazuya Ito)

Posted on • Edited on

empty and empty_like in PyTorch

#python #pytorch #empty #emptylike

Buy Me a Coffee☕

*My post explains empty_strided().

empty() can create the 0D or more D tensor of the zero or more floating-point numbers(Default), integers, complex numbers or boolean values from uninitialized memory which are called uninitialized data as shown below:

*Memos:

  • empty() can be used with torch but not with a tensor.
  • The 1st or more arguments with torch are size(Required-Type:int, tuple of int, list of int or size()).
  • There is dtype argument with torch(Optional-Default:None-Type:dtype): *Memos:
  • There is device argument with torch(Optional-Default:None-Type:str, int or device()): *Memos:
  • There is requires_grad argument with torch(Optional-Default:False-Type:bool): *Memos:
    • requires_grad= must be used.
    • My post explains requires_grad argument.
  • There is out argument with torch(Optional-Default:None-Type:tensor): *Memos:
    • out= must be used.
    • My post explains out argument.
  • You can use torch.Tensor() or torch.FloatTensor() like torch.Tensor(3, 2, 4) or torch.FloatTensor(3, 2, 4) because they can do the same job as empty(). *torch.Tensor() is the alias of torch.FloatTensor() by default.
  • Uninitialized memory has data but the data is unknown. 
import torch torch.empty(size=()) torch.empty(size=torch.tensor(8).size()) # tensor(3.6404e-27)  torch.empty(size=(0,)) torch.empty(0) torch.empty(size=torch.tensor([]).size()) # tensor([])  torch.empty(size=(3,)) torch.empty(3) torch.empty(size=torch.tensor([8, 3, 6]).size()) # tensor([-1.3610e+13, 4.4916e-41, -1.3610e+13])  torch.empty(size=(3, 2)) torch.empty(3, 2) torch.empty(size=torch.tensor([[8, 3], [6, 0], [2, 9]]).size()) # tensor([[-1.3610e+13, 4.4916e-41], # [5.7850e-23, 3.1100e-41], # [4.4842e-44, 0.0000e+00]])  torch.empty(size=(3, 2, 4)) torch.empty(3, 2, 4) # tensor([[[3.8848e-23, 3.1100e-41, 0.0000e+00, 0.0000e+00], # [3.3892e-23, 3.1100e-41, 3.0224e-26, 3.1100e-41]], # [[-6.0464e-34, 4.4914e-41, 0.0000e+00, 0.0000e+00], # [0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00]], # [[0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00], # [0.0000e+00, 0.0000e+00, 1.4013e-45, 0.0000e+00]]])  torch.empty(size=(3, 2, 4), dtype=torch.int64) torch.empty(3, 2, 4, dtype=torch.int64) # tensor([[[136263006428688, 96270204571280, 1, 96270203986320], # [0, 0, 96270208839376, 96270118417696]], # [[136257315028352, 0, 0, 0], # [0, 0, 0, 1]], # [[0, 0, 0, 0], # [0, 1, 352951805673479, 2542620672001]]])  torch.empty(size=(3, 2, 4), dtype=torch.complex64) torch.empty(3, 2, 4, dtype=torch.complex64) # tensor([[[1.4167e-07+4.4458e-41j, 1.4167e-07+4.4458e-41j, # 4.4842e-44+0.0000e+00j, 1.5695e-43+0.0000e+00j], # [-1.4883e+19+3.1404e-41j, 0.0000e+00+0.0000e+00j, # 1.4013e-45+0.0000e+00j, -4.9888e-15+3.1409e-41j]], # [[-2.4481e+37+4.4456e-41j, -4.9888e-15+3.1409e-41j, # 9.1477e-41+0.0000e+00j, 8.9683e-44+0.0000e+00j], # [3.5873e-43+0.0000e+00j, -2.6273e+37+4.4456e-41j, # 0.0000e+00+0.0000e+00j, 0.0000e+00+0.0000e+00j]], # [[0.0000e+00+0.0000e+00j, 2.4803e-43+0.0000e+00j, # -4.6535e-15+3.1409e-41j, -3.2145e-15+3.1409e-41j], # [0.0000e+00+0.0000e+00j, 1.4013e-45+0.0000e+00j, # -1.7014e+38+1.1515e-40j, 4.5919e-41+8.2957e-43j]]])  torch.empty(size=(3, 2, 4), dtype=torch.bool) torch.empty(3, 2, 4, dtype=torch.bool) # tensor([[[True, True, True, True], # [True, False, False, False]], # [[True, True, True, True], # [True, True, False, False]], # [[False, True, False, False], # [False, False, False, False]]])
Enter fullscreen mode Exit fullscreen mode

empty_like() can replace the zero or more numbers of a 0D or more D tensor with the zero or more floating-point numbers, integers, complex numbers or boolean values from uninitialized memory which are called uninitialized data as shown below:

*Memos:

  • empty_like() can be used with torch but not with a tensor.
  • The 1st argument with torch is input(Required-Type:tensor of int, float, complex or bool).
  • There is dtype argument with torch(Optional-Default:None-Type:dtype): *Memos:
    • If it's None, it's inferred from input.
    • dtype= must be used.
    • My post explains dtype argument.
  • There is device argument with torch(Optional-Default:None-Type:str, int or device()): *Memos:
    • If it's None, it's inferred from input.
    • device= must be used.
    • My post explains device argument.
  • There is requires_grad argument with torch(Optional-Default:False-Type:bool): *Memos:
    • requires_grad= must be used.
    • My post explains requires_grad argument.
  • Uninitialized memory has data but the data is unknown. 
import torch my_tensor = torch.tensor(7.) torch.empty_like(input=my_tensor) # tensor(-1.3610e+13)  my_tensor = torch.tensor([7., 4., 5.]) torch.empty_like(input=my_tensor) # tensor([2.8244e+23, 4.4787e-41, -5.7316e-07])  my_tensor = torch.tensor([[7., 4., 5.], [2., 8., 3.]]) torch.empty_like(input=my_tensor) # tensor([[-4.7415e-07, 3.1221e-41, -6.4098e-07], # [3.1221e-41, 1.1210e-43, 0.0000e+00]])  my_tensor = torch.tensor([[[7., 4., 5.], [2., 8., 3.]], [[6., 0., 1.], [5., 9., 4.]]]) torch.empty_like(input=my_tensor) # tensor([[[-6.6094e-07, 3.1221e-41, -3.9661e-07], # [3.1221e-41, 8.9683e-44, 0.0000e+00]], # [[1.1210e-43, 0.0000e+00, -8.9451e+02], # [3.1228e-41, 1.7282e-04, 1.2471e+16]]])  my_tensor = torch.tensor([[[7, 4, 5], [2, 8, 3]], [[6, 0, 1], [5, 9, 4]]]) torch.empty_like(input=my_tensor) # tensor([[[137273168313840, 95694909291296, 1], # [95694912519088, 95694842532640, 0]], # [[95694862074384, 95694820258896, 137269160918960], # [0, 0, 0]]])  my_tensor = torch.tensor([[[7.+4.j, 4.+2.j, 5.+3.j], [2.+5.j, 8.+1.j, 3.+9.j]], [[6.+9.j, 0.+3.j, 1.+8.j], [5.+3.j, 9.+4.j, 4.+6.j]]]) torch.empty_like(input=my_tensor) # tensor([[[6.7127e-07+1.7183e-04j, # 1.6519e-04+1.0187e-11j, # 2.0661e+20+6.8629e-07j], # [1.8077e-43+0.0000e+00j, # -4.3084e-07+3.1221e-41j, # -3.8936e-07+3.1221e-41j]], # [[4.4842e-44+0.0000e+00j, # 4.4842e-44+0.0000e+00j, # -8.7266e+02+3.1228e-41j], # [2.8026e-45+0.0000e+00j, # 4.2039e-45+0.0000e+00j, # 9.1084e-44+0.0000e+00j]]])  my_tensor = torch.tensor([[[True, False, True], [False, True, False]], [[True, False, True], [False, True, False]]]) torch.empty_like(input=my_tensor) # tensor([[[True, True, True], # [True, True, False]], # [[False, False, True], # [True, True, True]]])
Enter fullscreen mode Exit fullscreen mode

Top comments (0)