解决方案：似乎 qiskit 在您重复使用电路时会做一些奇怪的事情，这意味着运行它们需要额外的时间。解决方案是拉
for idx, image in enumerate(this_data):
np.random.seed(42)
quantum_circuits = [generate_random_circuit(depth=10,num_qubits=4,prob_appl_single=0.3,prob_appl_multi=0.7) for _ in range(2)]
对于任何感兴趣的人，进入 for 循环。
我已经看到为什么简单的 Python 程序会随着时间的推移而变慢？和Python 脚本随着它的进展而减慢？正如你所看到的，我已经尝试实施所有我可以从那里学到的东西，但这并没有帮助。
我正在尝试对一组数据执行一些计算，但随着时间的推移它会变得非常慢。具体来说
from quantum_pipe import generate_random_circuit, conv
import pandas as pd
import numpy as np
import gc
train_data = pd.read_csv('./fashion-mnist/fashion-mnist_train.csv')
this_data = train_data[args.start_idx:args.start_idx+args.num_datapoints].drop(['label'],axis=1)
this_data = (this_data / 255).round().astype(np.uint8).values
train_data = None # this is just for garbage collector to deal with
quantum_circuits = [generate_random_circuit(depth=10,num_qubits=4,prob_appl_single=0.3,prob_appl_multi=0.7) for _ in range(12)]
img_outputs = []
for idx, image in enumerate(this_data):
image = image.reshape((28,28))
outputs = [conv(qc, 2, image) for qc in quantum_circuits]
print(f"IMAGE COMPLETED: {idx+1} of {args.num_datapoints}")
img_outputs.append(outputs)
gc.collect()
np.save(f'./quantum_data/start{args.start_idx}to{args.start_idx+args.num_datapoints}.npy',img_outputs)
我的时间是每次在输出中完成列表理解的一个元素。我的 conv 函数定义为
def conv(qc, filter_size, image, mode='threshold'):
''' Write the loops to slide our 'filter' over our image '''
# here filter doesn't actually matter, we just use the flattened binary list as our init
# might as well hard-code 3x3 filters, can happily handle 2^9 = 512 states
start = time.time()
prepped_img = prepare_img(filter_size, image)
print(prepped_img.shape)
img_height, img_width = prepped_img.shape
conv_output = np.zeros(image.shape)
for down_idx in range(img_height - (filter_size-1)):
for across_idx in range(img_width - (filter_size-1)):
section = prepped_img[down_idx:down_idx + filter_size, across_idx: across_idx + filter_size]
init_arr = encoding_function(section,mode)
qc.initialize(init_arr, qc.qubits)
job = execute(qc, BACKEND, shots=500)
results = job.result()
counts = results.get_counts(qc)
output = np.zeros(len(init_arr))
for key, value in counts.items():
keyidx = int(key,2)
output[keyidx] = value
output = output/ np.sqrt(np.sum(output**2))
entropy = shannon_entropy(output)
conv_output[down_idx,across_idx] = entropy
print(f"filter completed in {time.time()-start} s")
return conv_output
查看我的终端输出，我看到类似
filter completed in 5.358206748962402 s
(15, 15)
filter completed in 5.510828733444214 s
(15, 15)
filter completed in 5.9010186195373535 s
(15, 15)
filter completed in 5.649052143096924 s
(15, 15)
filter completed in 5.400753021240234 s
(15, 15)
filter completed in 6.231165647506714 s
(15, 15)
filter completed in 6.202470779418945 s
(15, 15)
filter completed in 5.346660852432251 s
(15, 15)
filter completed in 5.406870365142822 s
(15, 15)
filter completed in 5.373116731643677 s
(15, 15)
filter completed in 6.068220853805542 s
IMAGE COMPLETED: 1 of 150
(15, 15)
filter completed in 9.925729036331177 s
(15, 15)
filter completed in 10.407076597213745 s
(15, 15)
filter completed in 10.152106523513794 s
(15, 15)
filter completed in 10.058736801147461 s
(15, 15)
filter completed in 10.039350509643555 s
(15, 15)
filter completed in 10.171052694320679 s
(15, 15)
filter completed in 10.037626266479492 s
(15, 15)
filter completed in 9.979500770568848 s
(15, 15)
filter completed in 10.973601579666138 s
(15, 15)
filter completed in 10.002075433731079 s
(15, 15)
filter completed in 9.934457540512085 s
(15, 15)
filter completed in 10.39132308959961 s
IMAGE COMPLETED: 2 of 150
(15, 15)
filter completed in 14.923802375793457 s
(15, 15)
filter completed in 14.426809549331665 s
(15, 15)
filter completed in 14.47310185432434 s
(15, 15)
filter completed in 14.773804187774658 s
(15, 15)
filter completed in 14.724096059799194 s
(15, 15)
filter completed in 14.462066650390625 s
(15, 15)
filter completed in 14.38855791091919 s
(15, 15)
filter completed in 14.864516735076904 s
(15, 15)
filter completed in 16.536264657974243 s
(15, 15)
filter completed in 15.544280052185059 s
(15, 15)
filter completed in 14.404176235198975 s
(15, 15)
filter completed in 14.43254804611206 s
IMAGE COMPLETED: 3 of 150
(15, 15)
filter completed in 18.811672687530518 s
(15, 15)
filter completed in 19.01636004447937 s
(15, 15)
filter completed in 19.908296823501587 s
(15, 15)
filter completed in 21.753503561019897 s
(15, 15)
filter completed in 20.8498797416687 s
(15, 15)
filter completed in 20.544127941131592 s
(15, 15)
filter completed in 20.662254333496094 s
(15, 15)
filter completed in 20.65923762321472 s
(15, 15)
filter completed in 19.335585355758667 s
(15, 15)
filter completed in 19.945393323898315 s
(15, 15)
filter completed in 20.45564317703247 s
(15, 15)
filter completed in 20.289149522781372 s
IMAGE COMPLETED: 4 of 150
(15, 15)
filter completed in 25.40975260734558 s
(15, 15)
filter completed in 24.785842418670654 s
(15, 15)
filter completed in 24.025460720062256 s
(15, 15)
filter completed in 23.852452516555786 s
(15, 15)
filter completed in 24.38658046722412 s
(15, 15)
filter completed in 23.814111948013306 s
这告诉我，同一图像的每个过滤器花费大约相同的时间，但是每个后续图像花费的时间越来越长，通常是一个恒定的量。鉴于我需要对许多图像执行此操作，这对我来说是不可持续的，并且在 uni 机器上，它从 3s 过滤器变为 70s 过滤器。
如果我缺少任何信息，我的仓库在这里https://github.com/milanleonard/QNN-implementation 。我无法想象为什么这会放缓。
对于一些额外的信息，似乎是导致问题的循环中的步骤，因为减少量子电路的数量，即每个图像的过滤器数量，根本没有帮助。在每一步都进行重构以保存并没有帮助，而且我的记忆在任何地方都没有快速增长。
我们认为可能是附加到 img_outputs 是问题所在，但这个重构表明它不是。
for idx, image in enumerate(this_data):
image = image.reshape((28,28))
image = prepare_img(2,image)
outputs = [conv(qc, 2, image) for qc in quantum_circuits]
print(f"IMAGE COMPLETED: {idx+1} of {args.num_datapoints}")
np.save(f'quantum_data/img{args.start_idx + idx}.npy',outputs)
outputs, image = None, None
在迈克尔建议的一些分析之后，我看到了这个
filter completed in 16.571566104888916 s
filter completed in 16.666788339614868 s
15378121 function calls (15376161 primitive calls) in 33.240 seconds
Ordered by: internal time
ncalls tottime percall cumtime percall filename:lineno(function)
788 4.492 0.006 4.492 0.006 {method 'acquire' of '_thread.lock' objects}
60368 3.586 0.000 9.891 0.000 dagcircuit.py:292(apply_operation_back)
241472 1.336 0.000 3.593 0.000 copy.py:66(copy)
1576240 1.166 0.000 1.188 0.000 {built-in method builtins.isinstance}
930412 0.674 0.000 0.674 0.000 bit.py:75(__hash__)
60368 0.672 0.000 1.220 0.000 assemble_circuits.py:71(<listcomp>)
244216 0.659 0.000 2.253 0.000 {built-in method builtins.all}
60760 0.611 0.000 1.113 0.000 quantumcircuit.py:581(_update_parameter_tabl
第二轮
filter completed in 36.4175808429718 s
filter completed in 36.10355544090271 s
36660244 function calls (36658284 primitive calls) in 72.523 seconds
Ordered by: internal time
ncalls tottime percall cumtime percall filename:lineno(function)
137200 9.124 0.000 25.409 0.000 dagcircuit.py:292(apply_operation_back)
784 7.039 0.009 7.039 0.009 {method 'acquire' of '_thread.lock' objects}
548800 2.966 0.000 8.067 0.000 copy.py:66(copy)
3958024 2.815 0.000 2.836 0.000 {built-in method builtins.isinstance}
2467052 1.752 0.000 1.752 0.000 bit.py:75(__hash__)
137592 1.650 0.000 3.057 0.000 quantumcircuit.py:581(_update_parameter_table)
551544 1.539 0.000 5.519 0.000 {built-in method builtins.all}
984312 1.509 0.000 2.749 0.000 retworkx_dagcircuit.py:62(_add_multi_graph_edge)
137200 1.495 0.000 2.727 0.000 assemble_circuits.py:71(<listcomp>)
489020 1.445 0.000 2.162 0.000 register.py:98(__repr__)
看起来对 ApplyOperationBack 的调用次数随着循环的迭代线性增长
60368 3.497 0.000 9.678 0.000 dagcircuit.py:292(apply_operation_back)
137200 8.993 0.000 25.163 0.000 dagcircuit.py:292(apply_operation_back)
214032 15.435 0.000 41.346 0.000 dagcircuit.py:292(apply_operation_back)
290864 20.282 0.000 54.842 0.000 dagcircuit.py:292(apply_operation_back)
我想我应该去寻找为什么它会增长。