test.py

import numpy as np
import gc
import memory_profiler

import theano
import theano.tensor as T
from theano.sandbox.cuda.dnn import dnn_conv

X = T.tensor4()
W = T.tensor4()

f = theano.function([X, W], dnn_conv(X, W))

X0 = np.zeros((1, 1, 5, 5)).astype('float32')
W0 = np.zeros((1, 1, 5, 5)).astype('float32')

gc.collect()
gc.collect()
gc.collect()
mem0 = memory_profiler.memory_usage(-1)[0]

for _ in range(100):
    for _ in range(1000):
        _ = f(X0, W0)
    gc.collect()
    gc.collect()
    gc.collect()
    mem = memory_profiler.memory_usage(-1)[0]
    print(mem - mem0)

output:

Using gpu device 0: GeForce GTX TITAN
0.13671875
0.14453125
0.171875
0.26171875
0.35546875
0.4453125
0.53515625
0.62890625
0.71875
0.8125
0.90234375
0.99609375
1.0859375
1.17578125
1.26953125
1.359375
1.453125
1.54296875
1.63671875
1.7265625
1.8203125
1.91015625