Is it possible to create a FIFO queue with pyTorch?

Question

I need to create a fixed length Tensor in pyTorch that acts like a FIFO queue.

I have this fuction to do it:

def push_to_tensor(tensor, x):
    tensor[:-1] = tensor[1:]
    tensor[-1] = x
    return tensor

For example, I have:

tensor = Tensor([1,2,3,4])

>> tensor([ 1.,  2.,  3.,  4.])

then using the function will give:

push_to_tensor(tensor, 5)

>> tensor([ 2.,  3.,  4.,  5.])

However, I was wondering:

• Does pyTorch have a native method for doing this?
• If not, is there a more clever way of doing it?

Answer 1

A more generic version of @Bruno_Lubascher's answer, with control over the deque size, support of batched insertion, and control over the push dimension:

def push_to_deque(deque, x, deque_size=None, dim=0):
    """Handling `deque` tensor as a (set of) deque/FIFO, push the content of `x` into it."""
    if deque_size is None:
        deque_size = deque.shape[dim]
    deque_dims = deque.dim()
    input_size = x.shape[dim]
    dims_right = deque_dims - dim - 1
    deque_slicing = (
        (slice(None),) * dim
        + (
            slice(
                input_size - deque_size
                if input_size < deque_size
                else deque.shape[dim],
                None,
            ),
        )
        + (slice(None),) * dims_right
    )
    input_slicing = (
        (slice(None),) * dim + (slice(-deque_size, None),) + (slice(None),) * dims_right
    )
    deque = torch.cat((deque[deque_slicing], x[input_slicing]), dim=dim)
    return deque

Examples:

>>> # Consider batched deques containing vectors of shape (2,):
>>> batch_size, vector_size = 1, 2  
>>> deque_size = 4
>>> # Initialize the empty deques:
>>> deques = torch.empty((batch_size, 0, vector_size)) 
>>> # Push at once more vectors than the batched FIFOs can contain:
>>> vals = torch.arange(10).view((batch_size, 5, vector_size))
>>> deque = push_to_deque(deque, vals, deque_size=deque_size, dim=1) 
>>> deque
tensor([[[2., 3.],
         [4., 5.],
         [6., 7.],
         [8., 9.]]])
>>> # Push some more:
>>> vals = torch.arange(10, 20).view((batch_size, 5, vector_size))
>>> deque = push_to_deque(deque, vals, deque_size=deque_size, dim=1) 
>>> deque
tensor([[[12., 13.],
         [14., 15.],
         [16., 17.],
         [18., 19.]]])
>>> vals = torch.arange(20, 24).view((batch_size, 2, vector_size))
>>> deque = push_to_deque(deque, vals, deque_size=deque_size, dim=1) 
>>> deque
tensor([[[16., 17.],
         [18., 19.],
         [20., 21.],
         [22., 23.]]])
>>> # Verify the method can also handle oversized FIFOs:
>>> deque = torch.zeros(batch_size, 10, vector_size)
>>> vals = torch.arange(4).view((batch_size, 2, vector_size))
>>> deque = push_to_deque(deque, vals, deque_size=deque_size, dim=1)
>>> deque
tensor([[[0., 0.],
         [0., 0.],
         [0., 1.],
         [2., 3.]]])

Answer 2

Maybe a little late but I found another way to do this and save some time. In my case, I needed a similar FIFO structure but I only needed to actually parse the FIFO tensor once every N iterations. i.e. I needed a FIFO structure to hold n integers, and every n iterations I needed to parse that tensor thourgh my model. I found it is way faster to implement a collections.deque structure and cast that deque to a tensor torch.

import time
import torch
from collections import deque
length = 5000

que = deque([0]*200)

ten = torch.tensor(que)

s = time.time()
for i in range(length):
    for j in range(200):  
        que.pop()      
        que.appendleft(j*10)        
    torch.tensor(que)
    # after some appending/popping elements, cast to tensor
print("finish deque:", time.time()-s)


s = time.time()
for i in range(length):
    for j in range(200):
        newelem = torch.tensor([j*10])
        ten = torch.cat((ten[1:], newelem))
        #using tensor as FIFO, no need to cast to tensor
print("finish tensor:", time.time()-s)

the results are the following:

finish deque: 0.15857529640197754
finish tensor: 9.483643531799316

I also noticed that when using a deque and always casting to a torch.tensor instead of using push_alternative it can give you a ~20% boost in time.

s = time.time()
for j in range(length):    
        que.pop()      
        que.appendleft(j*10)        
        torch.tensor(que)    
print("finish queue:", time.time()-s)


s = time.time()
for j in range(length):    
        newelem = torch.tensor([j*10])
        ten = torch.cat((ten[1:], newelem))
print("finish tensor:", time.time()-s)

results:

finish queue: 8.422480821609497
finish tensor: 11.169137477874756

Answer 3

I implemented another FIFO queue:

def push_to_tensor_alternative(tensor, x):
    return torch.cat((tensor[1:], Tensor([x])))

The functionality is the same, but then I checked their performance in speed:

# Small Tensor
tensor = Tensor([1,2,3,4])

%timeit push_to_tensor(tensor, 5)
>> 30.9 µs ± 1.26 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

%timeit push_to_tensor_alternative(tensor, 5)
>> 22.1 µs ± 2.25 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

# Larger Tensor
tensor = torch.arange(10000)

%timeit push_to_tensor(tensor, 5)
>> 57.7 µs ± 4.88 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

%timeit push_to_tensor_alternative(tensor, 5)
>> 28.9 µs ± 570 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)

Seems like this push_to_tensor_alternative which uses torch.cat (instead of shifting all items to the left) is faster.