Create a binary pytorch tensor with n% ones

Question

I want to create a binary tensor with 0 and 1s only, but the number of 1s should be n%.

I am using this code, but it takes 7+ seconds and has 19926MB of memory.

Is there any faster way to do that?

import torch
import time
start_time = time.time()
def create_random_binary_tensor(shape, percentage, device):
    size = shape[0] * shape[1]
    num_ones = int(size * (percentage / 100.0))
    
    # Create a tensor filled with zeros
    tensor = torch.zeros(size, dtype=torch.float32, device=device)
    
    # Set random indices to ones until the desired number of ones is reached
    indices = torch.randperm(size, device=device)[:num_ones]
    tensor[indices] = 1
    
    # Reshape the tensor to the desired shape
    tensor = tensor.view(shape)
    
    return tensor

# Set the shape and percentage of ones
shape = (19000, 19000)
percentage = 0.5

# Check if CUDA is available and use it if possible
device = torch.device("cuda:2" if torch.cuda.is_available() else "cpu")

# Create the random binary tensor
tensor = create_random_binary_tensor(shape, percentage, device)

# Print the tensor
execution_time = time.time() - start_time
print("Execution Time: {:.4f} seconds".format(execution_time))

Answer 1

This snippet should take less on the ground that it does fewer operations.

p=0.5
D=[19000,19000]
torch.rand(D)>p

As D increases, the mean approaches p.