I roto i tenei tuhinga, ka korero atu ahau ki a koe me pehea te whakarite i tetahi taiao ako miihini i roto i te 30 meneti, te hanga i tetahi whatunga neural mo te tohu whakaahua, katahi ka whakahaere i taua whatunga ki runga i te tukatuka whakairoiro (GPU).
Tuatahi, me tautuhi he aha te whatunga neural.
I roto i ta maatau, he tauira pangarau tenei, me ona ahuatanga rorohiko, taputapu taputapu ranei, i hangaia i runga i te kaupapa o te whakahaere me te mahi o nga whatunga neural koiora - nga whatunga o nga pūtau nerve o te rauropi ora. I puta mai tenei ariā i te wa e ako ana i nga mahi ka puta i roto i te roro me te ngana ki te whakatauira i enei mahi.
Ko nga whatunga neural kaore i te whakamaherehia i roto i te tikanga o te kupu, kua whakangungua. Ko te kaha ki te ako ko tetahi o nga painga nui o nga whatunga neural i runga i nga algorithm tuku iho. Ko te tikanga, ko te ako ko te kimi i nga whakarea o nga hononga i waenga i nga neurons. I te wa o te whakangungu, ka taea e te whatunga neural te tautuhi i nga whakawhirinakitanga uaua i waenga i nga raraunga whakauru me nga raraunga whakaputa, me te mahi i te whaanui.
Mai i te tirohanga o te ako miihini, ko te whatunga neural he keehi motuhake mo nga tikanga tohu tauira, tātaritanga whakahirahira, tikanga whakahiato me etahi atu tikanga.
Tuhinga
Сначала разберемся с оборудованием. Нам необходим сервер с установленной на нем операционной системой Linux. Оборудование для работы систем машинного обучения требуется достаточно мощное и как следствие дорогое. Тем, у кого нет под рукой хорошей машины, рекомендую обратить внимание на предложение облачных провайдеров. Необходимый сервер можно получить в аренду быстро и платить только за время использования.
I roto i nga kaupapa e tika ana ki te hanga whatunga neural, ka whakamahi ahau i nga kaitoro o tetahi o nga kaiwhakarato kapua a Ruhia. Ka tukuna e te kamupene nga tuutuu kapua mo te riihi motuhake mo te ako miihini me nga miihini whakairoiro Tesla V100 (GPU) mai i NVIDIA. I roto i te poto: te whakamahi i te tūmau me te GPU ka taea te tekau nga wa te pai ake (tere) whakaritea ki te tūmau o te utu rite e whakamahi ana i te PTM (te wae tukatuka pokapū rongonui) mo te tātaitanga. Ka tutuki tenei na nga ahuatanga o te hoahoanga GPU, he tere ake te tarai i nga tatauranga.
Hei whakatinana i nga tauira e whakaahuatia ana i raro nei, i hokona e matou te tūmau e whai ake nei mo etahi ra:
- Kopae SSD 150 GB
- RAM 32 GB
- Tesla V100 16 Gb pūtukatuka ki 4 matua
На машину нам установили Ubuntu 18.04.
Te whakarite i te taiao
Inaianei me whakauru nga mea katoa e tika ana mo te mahi i runga i te tūmau. I te mea ko ta maatau tuhinga te tuatahi mo nga tiimata, ka korero ahau mo etahi waahanga ka whai hua ki a raatau.
Очень много работы при настройке среды выполняется через командную строку. Большинство из пользователей в качестве рабочей ОС используют Windows. Стандартная консоль в этой ОС оставляет желать лучшего. Поэтому мы будем использовать удобный инструмент . Tangohia te putanga paku ka whakahaere Cmder.exe. Whai muri me hono koe ki te tūmau mā SSH:
ssh root@server-ip-or-hostnameEngari i te server-ip-or-hostname, whakapūtā te wāhitau IP, te ingoa DNS rānei o tō tūmau. I muri mai, whakauruhia te kupuhipa me te mea kua angitu te hononga, me whiwhi he karere rite ki tenei.
Welcome to Ubuntu 18.04.3 LTS (GNU/Linux 4.15.0-74-generic x86_64)Основным языком для разработки ML моделей является Python. А наиболее популярной платформой для его использование по Linux ko te reira .
Kia tāuta i runga i to tatou tūmau.
Ka timata tatou ma te whakahōu i te kaiwhakahaere mōkihi ā-rohe:
sudo apt-get updateTāutahia te curl (whakamahinga raina whakahau):
sudo apt-get install curlTangohia te putanga hou o Anaconda Distribution:
cd /tmp
curl –O https://repo.anaconda.com/archive/Anaconda3-2019.10-Linux-x86_64.shMe timata te whakaurunga:
bash Anaconda3-2019.10-Linux-x86_64.shI te wa o te whakaurunga, ka tonohia koe ki te whakaū i te kirimana raihana. Ina oti te whakaurunga me kite koe i tenei:
Thank you for installing Anaconda3!He maha nga anga kua hangaia inaianei mo te whanaketanga o nga tauira ML; mahi matou me nga mea rongonui: и .
Ma te whakamahi i te anga ka taea e koe te whakanui ake i te tere o te whakawhanaketanga me te whakamahi i nga taputapu kua rite mo nga mahi paerewa.
I tenei tauira ka mahi tahi tatou me PyTorch. Kia tāuta tatou:
conda install pytorch torchvision cudatoolkit=10.1 -c pytorchInaianei me whakarewa tatou i te Jupyter Notebook, he taputapu whanaketanga rongonui mo nga tohunga ML. Ka taea e koe te tuhi waehere me te kite tonu i nga hua o tana mahi. Kei te whakauruhia a Jupyter Notebook ki a Anaconda a kua whakauruhia ki runga i ta maatau tūmau. Me hono koe ki a ia mai i ta maatau punaha papamahi.
Hei mahi i tenei, ka whakarewahia e matou a Jupyter i runga i te tūmau e tohu ana i te tauranga 8080:
jupyter notebook --no-browser --port=8080 --allow-rootWhai muri, ka whakatuwhera i tetahi atu ripa i roto i ta maatau papatohu Cmder (tahua runga - korero papatohu hou) ka hono atu ma te tauranga 8080 ki te tūmau ma te SSH:
ssh -L 8080:localhost:8080 root@server-ip-or-hostnameKa uru ana matou ki te whakahau tuatahi, ka tukuna he hononga ki a matou ki te whakatuwhera i a Jupyter i roto i to maatau tirotiro:
To access the notebook, open this file in a browser:
file:///root/.local/share/jupyter/runtime/nbserver-18788-open.html
Or copy and paste one of these URLs:
http://localhost:8080/?token=cca0bd0b30857821194b9018a5394a4ed2322236f116d311
or http://127.0.0.1:8080/?token=cca0bd0b30857821194b9018a5394a4ed2322236f116d311
Me whakamahi te hono mo localhost:8080. Tāruatia te ara katoa ka whakapiri ki te pae wāhitau o te pūtirotiro paetata o tō PC. Ka tuwhera a Jupyter Notebook.
Me hanga he pukatuhi hou: Hou - Pukatuhituhi - Python 3.
Kia tirohia te mahi tika o nga waahanga katoa i whakauruhia e matou. Whakauruhia te tauira PyTorch waehere ki Jupyter ka whakahaere i te mahi (Pātene Whakahaere):
from __future__ import print_function
import torch
x = torch.rand(5, 3)
print(x)
Me penei te hua:
Mena he rite to hua, katahi ka whirihora tika i nga mea katoa ka taea e taatau te tiimata ki te whakawhanake i tetahi whatunga neural!
Te hanga whatunga neural
Ka hangahia e matou he whatunga neural mo te tohu whakaahua. Me tango tenei hei kaupapa .
Ka whakamahia e matou te huinga raraunga CIFAR10 e waatea ana ki te iwi katoa hei whakangungu i te whatunga. He karaehe: "rererangi", "motika", "manu", "ngeru", "tia", "kuri", "poroka", "hoiho", "kaipuke", "taraka". Ko nga whakaahua kei CIFAR10 he 3x32x32, ara, 3-hongere whakaahua tae o te 32x32 pika.
Mo te mahi, ka whakamahia e matou te kete i hangaia e PyTorch mo te mahi me nga whakaahua - torchvision.
Ka mahi maatau i nga waahanga e whai ake nei:
- Te uta me te whakarite i nga huinga raraunga whakangungu me te whakamatautau
- Whakamaramatanga Whatunga Neural
- Whakangungu whatunga mo nga raraunga whakangungu
- Te whakamatautau whatunga mo nga raraunga whakamatautau
- Kia tuaruatia te whakangungu me te whakamatautau ma te whakamahi GPU
Ka mahia e matou nga waehere katoa i raro nei i te Pukatuhipoka Jupyter.
Te uta me te whakarite CIFAR10
Tāruahia me te whakahaere i te waehere e whai ake nei ki Jupyter:
import torch
import torchvision
import torchvision.transforms as transforms
transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=4,
shuffle=True, num_workers=2)
testset = torchvision.datasets.CIFAR10(root='./data', train=False,
download=True, transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size=4,
shuffle=False, num_workers=2)
classes = ('plane', 'car', 'bird', 'cat',
'deer', 'dog', 'frog', 'horse', 'ship', 'truck')Me penei te whakautu:
Downloading https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to ./data/cifar-10-python.tar.gz
Extracting ./data/cifar-10-python.tar.gz to ./data
Files already downloaded and verifiedMe whakaatu etahi whakaahua whakangungu mo te whakamatautau:
import matplotlib.pyplot as plt
import numpy as np
# functions to show an image
def imshow(img):
img = img / 2 + 0.5 # unnormalize
npimg = img.numpy()
plt.imshow(np.transpose(npimg, (1, 2, 0)))
plt.show()
# get some random training images
dataiter = iter(trainloader)
images, labels = dataiter.next()
# show images
imshow(torchvision.utils.make_grid(images))
# print labels
print(' '.join('%5s' % classes[labels[j]] for j in range(4)))
Whakamaramatanga Whatunga Neural
Me whakaaro tuatahi me pehea te mahi a te whatunga neural mo te tohu whakaahua. He whatunga tohu-ki-ira ngawari tenei. Ka tango i nga raraunga whakauru, ka tukuna ma te maha o nga paparanga kotahi, ka mutu ka whakaputa raraunga whakaputa.
Me hanga he whatunga rite ki to tatou taiao:
import torch.nn as nn
import torch.nn.functional as F
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(3, 6, 5)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(6, 16, 5)
self.fc1 = nn.Linear(16 * 5 * 5, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 10)
def forward(self, x):
x = self.pool(F.relu(self.conv1(x)))
x = self.pool(F.relu(self.conv2(x)))
x = x.view(-1, 16 * 5 * 5)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
net = Net()
Ka tautuhia hoki e matou he mahi ngaro me te moohio
import torch.optim as optim
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)Whakangungu whatunga mo nga raraunga whakangungu
Me timata tatou ki te whakangungu i to tatou whatunga neural. Kia mohio koe i muri i to whakahaere i tenei waehere, me tatari koe kia oti ra ano te mahi. E 5 meneti ahau. He wa roa ki te whakangungu i te whatunga.
for epoch in range(2): # loop over the dataset multiple times
running_loss = 0.0
for i, data in enumerate(trainloader, 0):
# get the inputs; data is a list of [inputs, labels]
inputs, labels = data
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
# print statistics
running_loss += loss.item()
if i % 2000 == 1999: # print every 2000 mini-batches
print('[%d, %5d] loss: %.3f' %
(epoch + 1, i + 1, running_loss / 2000))
running_loss = 0.0
print('Finished Training')
Ka whiwhi tatou i te hua e whai ake nei:
Ka tiakina e matou to maatau tauira kua whakangungua:
PATH = './cifar_net.pth'
torch.save(net.state_dict(), PATH)Te whakamatautau whatunga mo nga raraunga whakamatautau
I whakangungua e matou te whatunga ma te whakamahi i te huinga raraunga whakangungu. Engari me tirotiro mena kua akohia e te whatunga tetahi mea.
Ka whakamatauhia tenei ma te matapae i te tapanga o te karaehe ka whakaputahia e te whatunga neural me te whakamatautau kia kitea mena he pono. Mena he tika te matapae, ka taapirihia te tauira ki te rarangi o nga matapae tika.
Me whakaatu he whakaahua mai i te huinga whakamatautau:
dataiter = iter(testloader)
images, labels = dataiter.next()
# print images
imshow(torchvision.utils.make_grid(images))
print('GroundTruth: ', ' '.join('%5s' % classes[labels[j]] for j in range(4)))
Inaianei me patai ki te whatunga neural kia korero mai he aha kei roto i enei pikitia:
net = Net()
net.load_state_dict(torch.load(PATH))
outputs = net(images)
_, predicted = torch.max(outputs, 1)
print('Predicted: ', ' '.join('%5s' % classes[predicted[j]]
for j in range(4)))
He pai te ahua o nga hua: i tohu tika te whatunga e toru o nga pikitia e wha.
Kia kite e pehea te mahi a te whatunga puta noa i te huingararaunga katoa.
correct = 0
total = 0
with torch.no_grad():
for data in testloader:
images, labels = data
outputs = net(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
print('Accuracy of the network on the 10000 test images: %d %%' % (
100 * correct / total))
Te ahua nei kei te mohio te whatunga ki tetahi mea kei te mahi. Mena ka whakatauhia e ia nga karaehe i runga i te matapōkere, ka 10% te tika.
Inaianei kia kite ko wai nga karaehe e tohu pai ake ana te whatunga:
class_correct = list(0. for i in range(10))
class_total = list(0. for i in range(10))
with torch.no_grad():
for data in testloader:
images, labels = data
outputs = net(images)
_, predicted = torch.max(outputs, 1)
c = (predicted == labels).squeeze()
for i in range(4):
label = labels[i]
class_correct[label] += c[i].item()
class_total[label] += 1
for i in range(10):
print('Accuracy of %5s : %2d %%' % (
classes[i], 100 * class_correct[i] / class_total[i]))
Te ahua nei he pai rawa atu te whatunga ki te tautuhi waka me nga kaipuke: 71% tika.
Na kei te mahi te whatunga. Inaianei me ngana ki te whakawhiti i ana mahi ki te tukatuka whakairoiro (GPU) ka kite he aha nga huringa.
Te whakangungu i te whatunga neural i runga i te GPU
Tuatahi, ka whakamarama poto ahau he aha te CUDA. Ko te CUDA (Compute Unified Device Architecture) he papa rorohiko whakarara i hangaia e NVIDIA mo te rorohiko whanui i runga i nga waeine tukatuka whakairoiro (GPU). Ma te CUDA, ka taea e nga kaihanga te whakatere i nga tono rorohiko ma te whakamahi i te mana o nga GPU. Kua whakauruhia tenei turanga ki runga i ta maatau tūmau i hokona e matou.
Me tautuhi i to tatou GPU hei taputapu cuda tuatahi ka kitea.
device = torch . device ( "cuda:0" if torch . cuda . is_available () else "cpu" )
# Assuming that we are on a CUDA machine, this should print a CUDA device:
print ( device )
Te tuku i te whatunga ki te GPU:
net.to(device)Ka tukuna ano e matou nga whakaurunga me nga whaainga i ia taahiraa ki te GPU:
inputs, labels = data[0].to(device), data[1].to(device)Kia whakangungu ano tatou i te whatunga i runga i te GPU:
import torch.optim as optim
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
for epoch in range(2): # loop over the dataset multiple times
running_loss = 0.0
for i, data in enumerate(trainloader, 0):
# get the inputs; data is a list of [inputs, labels]
inputs, labels = data[0].to(device), data[1].to(device)
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
# print statistics
running_loss += loss.item()
if i % 2000 == 1999: # print every 2000 mini-batches
print('[%d, %5d] loss: %.3f' %
(epoch + 1, i + 1, running_loss / 2000))
running_loss = 0.0
print('Finished Training')I tenei wa, ka roa te whakangungu whatunga mo te 3 meneti. Kia maumahara tatou e 5 meneti te roa o te atamira kotahi i runga i te tukatuka tikanga. Ehara i te mea nui te rereketanga, ka tupu tenei na te mea kaore i te rahi to taatau whatunga. A, no te whakamahi i nga raupapa nui mo te whakangungu, ka piki ake te rereketanga i waenga i te tere o te GPU me te tukatuka tuku iho.
Ko te ahua o tera. He aha i taea e matou te mahi:
- I titiro matou he aha te GPU me te whiriwhiri i te tūmau i whakauruhia ai;
- Kua whakaturia e matou he taiao rorohiko hei hanga i te whatunga neural;
- I hanga e matou he whatunga neural mo te tohu ahua me te whakangungu;
- I tukuna ano e matou te whakangungu whatunga ma te whakamahi i te GPU ka piki ake te tere.
Ka koa ahau ki te whakautu i nga paatai i roto i nga korero.
Source: will.com
