Free memory of cupy.ndarray

[Yusuke Watanabe]

Hello.
I would like to know whether there is an explicit way to "free" memory of cupy.ndarray, which is allocated on a GPU.


I tried the following code:

import numpy as np
import cupy

x = np.random.rand(1000,1000)
cupy.array(x)
cupy.array(x)
cupy.array(x)
cupy.array(x)
cupy.array(x)
...

Everytime I execute `cupy.array(x)`, GPU memory usage increases by about 8MB.
(I checked it using nvidia-smi command.)
But if execute `y = cupy.array(x)` many times, GPU memory usage does not increase.
(Maybe, memory is freed by some reference counting mechanism.)

My problem is that, if allocate cuda array on iterator, like:

def get_iterator():
    z = cupy.array(np.random.rand(1000,1000))
    for i in range(10):
        yield z[100*i:100*(i+1)]
    del z # this del seems not free memory

iterator = get_iterator()
for d in iterator:
    print d[0]
# memory usage increases after this loop...

the memory is not freed.
Are there any good way to free z?

Any suggestions would be helpful.
Thanks.

[Ryosuke Okuta]

Hello.

`z[100*i:100*(i+1)]` slice make shallow copy object.

Please try code below.

def get_iterator():
    z = cupy.array(np.random.rand(1000,1000))
    for i in range(10):


        yield z[100*i:100*(i+1)].copy()  # make deep copy object


    del z # this del seems not free memory 

iterator = get_iterator()
for d in iterator:
    print d[0]
# memory usage increases after this loop...


2015年11月25日水曜日 13時06分57秒 UTC+9 Yusuke Watanabe:

[Yusuke Watanabe]

Oh, great! The memory is freed after iterator is used.
Thank you!

Here is my understanding:
In the original code, there were some reference to z, because z[100*i:100*(i+1)] is shallow copy.
In the new code, there are no reference to z, so it freed.

2015年11月25日水曜日 13時17分17秒 UTC+9 Ryosuke Okuta:

[Ryosuke Okuta]

Yes, you are right.

2015年11月25日水曜日 13時29分53秒 UTC+9 Yusuke Watanabe:

[Igor S]

What if i do not use iterator? Any other ways to free memory?

>>> x = np.array([img_a.reshape((3,224,224))] * 200)

>>> x.shape

(200, 3, 224, 224)

>>> xp = chainer.cuda.cupy

>>> X = xp.asarray(x, dtype=xp.float32)

>>> del X

This one remains in GPU memory after del, maybe i'm missing something?

[Ryosuke Okuta]

Memory will be freed when it is no longer referenced .

In addition to using `del X`, also it will be freed by overwriting `X = None` and exiting from the local scope.

2015年11月30日月曜日 0時00分59秒 UTC+9 Igor S: