Introduction
In this article, we will show you how to create random tensors in PyTorch with torch.randint() and torch.randint_like() functions. These functions are very useful to initialize neural networks with random values at the start of the training process. We shall cover these functions with the help of examples for easy understanding for beginners.
RandomÂ PyTorch Tensors with torch.randint()
torch.randint function is used to create a tensor with random integer values between the low (inclusive) and high (exclusive) value that is specified in the function.
To go through the examples of torch randint function let us first import the PyTorch library.
In [0]:
importÂ torch;
Â
Example â€“ 1: Creating 1 Dimensional Random Tensor with torch.randint()
In the first example, we are going to create a 1D tensor of 6 elements with random integers between 0 and 20. Do notice that in this example we have not specified low parameter, as a result, it is assumed to be 0 as default.
In [1]:
tensorÂ =Â torch.randint(highÂ =Â 20,Â sizeÂ =Â (6,)) tensor
Out[1]:
tensor([ 5, 18, 10, 5, 17, 12])
Example â€“ 2: Creating 1 Dimensional Random Tensor with Low and High
In this example, we are specifying both values of low and high to create a random tensor of size 6.
In [2]:
tensorÂ =Â torch.randint(lowÂ =Â 10,Â highÂ =Â 20,Â sizeÂ =Â (6,)) tensor
Out[2]:
tensor([17, 13, 11, 17, 11, 15])
Â
Example â€“ 3: Creating 2 Dimensional Random Tensor
Here we are creating 2D random tensor of size 4Ã—5 with torch randint function.
In [3]:
tensorÂ =Â torch.randint(lowÂ =Â 10,Â highÂ =Â 50,Â sizeÂ =Â (4,5)) tensor
Out[3]:
tensor([[44, 38, 39, 32, 34], [18, 29, 36, 23, 17], [13, 10, 10, 14, 32], [44, 39, 45, 21, 48]])
Example â€“ 4: Creating 3 Dimensional Random Tensor
Next, we are creating a 3D tensor of size 2x4x5 with torch randint function.
In [4]:
tensorÂ =Â torch.randint(lowÂ =Â 10,Â highÂ =Â 50,Â sizeÂ =Â (2,4,5)) tensor
tensor([[[14, 42, 13, 35, 14], [34, 24, 44, 49, 39], [24, 11, 29, 46, 43], [11, 15, 15, 46, 49]], [[36, 48, 13, 26, 20], [16, 35, 20, 47, 16], [39, 44, 15, 25, 23], [42, 18, 29, 12, 34]]])
Example â€“ 5: Using dtype parameter with torch.randint
By default torch randint creates a tensor whose elements are of type int. However, you can use dtype parameter to specify other data types. In this example, we are using dtype parameter as float.
In [5]:
tensorÂ =Â torch.randint(lowÂ =Â 10,Â highÂ =Â 50,Â sizeÂ =Â (4,5),Â dtypeÂ =Â float) tensor
tensor([[27., 11., 42., 33., 48.], [43., 43., 11., 27., 17.], [31., 30., 39., 15., 38.], [10., 20., 18., 29., 35.]], dtype=torch.float64)
Example â€“ 6: Using Seed Value with torch.randint
torch randint function will generate different random values of tensor whenever it is executed. However, we can use a seed value to make sure it generates the same value whenever it is executed.
To start with, we create a generator object by initializing with a seed number with manual_seed. Next, this generator object is used inside torch.randint function that will make sure the same values are generated in the tensor whenever the code is executed.
In[6]:
g_cpuÂ =Â torch.Generator() g_cpu.manual_seed(2542587391) tensorÂ =Â torch.randint(lowÂ =Â 25,Â highÂ =Â 125,Â sizeÂ =Â (4,3),Â generatorÂ =Â g_cpu) tensor
tensor([[ 29, 64, 30], [ 92, 51, 122], [ 45, 122, 35], [ 83, 123, 107]])
Random PyTorch Tensors with torch.randint_like()
There might be a situation where you may have to create a random tensor whose size is same as another tensor. To achieve this you can manually get the size of the other tensor and then create the random tensor using that. Alternatively, there is a more convenient way of using torch.randint_like() function to achieve the same.
To understand this function let us see some examples.
Example â€“ 1: Creating 1D Random Tensor with torch.randint_like()
Let us first create a zero tensor of size 5 and then use it inside torch.randint_like function as a reference for size along with low & high parameters.
It is interesting to note that by default zero tensor is of float type and the same data type is inherited by the random tensor too.
In [7]:
input_tensorÂ =Â torch.zeros(size=(5,)) input_tensor
tensor([0., 0., 0., 0., 0.])
tensorÂ =Â torch.randint_like(inputÂ =Â input_tensor,lowÂ =Â 10,Â highÂ =Â 50) tensor
tensor([12., 10., 10., 45., 17.])
Example â€“ 2: Creating 2D Random Tensor with torch.randint_like()
In [9]:
input_tensorÂ =Â torch.zeros(size=(5,3),Â dtypeÂ =Â int) input_tensor
tensor([[0, 0, 0], [0, 0, 0], [0, 0, 0], [0, 0, 0], [0, 0, 0]])
tensorÂ =Â torch.randint_like(inputÂ =Â input_tensor,lowÂ =Â 10,Â highÂ =Â 50) tensor
tensor([[42, 13, 49], [32, 35, 42], [19, 23, 28], [22, 16, 39], [25, 10, 20]])
Example â€“ 3: Creating 3D Random Tensor with torch.randint_like()
Here we are creating a 3D zeros tensor of size 2x5x3 and then use it to create an integer random tensor of the same size with torch rand int function.
In [11]:
input_tensorÂ =Â torch.zeros(size=(2,5,3)) input_tensor
tensor([[[0., 0., 0.], [0., 0., 0.], [0., 0., 0.], [0., 0., 0.], [0., 0., 0.]], [[0., 0., 0.], [0., 0., 0.], [0., 0., 0.], [0., 0., 0.], [0., 0., 0.]]])
tensorÂ =Â torch.randint_like(inputÂ =Â input_tensor,lowÂ =Â 10,Â highÂ =Â 50,Â dtype=int) tensor
tensor([[[35, 35, 10], [12, 13, 23], [45, 22, 34], [31, 46, 12], [25, 18, 10]], [[41, 44, 42], [32, 20, 46], [15, 19, 14], [16, 34, 13], [32, 39, 11]]])
 Reference: PyTorch Documentation
Â

MLK is a knowledge sharing community platform for machine learning enthusiasts, beginners and experts. Let us create a powerful hub together to Make AI Simple for everyone.