{"id":5350,"date":"2025-04-06T10:00:34","date_gmt":"2025-04-06T06:30:34","guid":{"rendered":"https:\/\/parmin.cloud\/blog\/?p=5350"},"modified":"2025-03-29T10:19:01","modified_gmt":"2025-03-29T06:49:01","slug":"%d9%86%d8%b1%d9%85%d8%a7%d9%84%d8%b3%d8%a7%d8%b2%db%8c-%d8%af%d8%b3%d8%aa%d9%87%d8%a7%db%8c-%d8%af%d8%b1-%d8%b4%d8%a8%da%a9%d9%87%d9%87%d8%a7%db%8c-%d8%b9%d8%b5%d8%a8%db%8c","status":"publish","type":"post","link":"https:\/\/parmin.cloud\/blog\/%d9%86%d8%b1%d9%85%d8%a7%d9%84%d8%b3%d8%a7%d8%b2%db%8c-%d8%af%d8%b3%d8%aa%d9%87%d8%a7%db%8c-%d8%af%d8%b1-%d8%b4%d8%a8%da%a9%d9%87%d9%87%d8%a7%db%8c-%d8%b9%d8%b5%d8%a8%db%8c\/","title":{"rendered":"\u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u062f\u0631 \u0634\u0628\u06a9\u0647\u200c\u0647\u0627\u06cc \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc"},"content":{"rendered":"
\u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0627\u0635\u0637\u0644\u0627\u062d\u06cc \u0627\u0633\u062a \u06a9\u0647 \u0645\u0639\u0645\u0648\u0644\u0627\u064b \u062f\u0631 \u0632\u0645\u06cc\u0646\u0647 \u0634\u0628\u06a9\u0647\u200c\u0647\u0627\u06cc \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u0630\u06a9\u0631 \u0645\u06cc\u200c\u0634\u0648\u062f. \u062f\u0631 \u0627\u06cc\u0646 \u0645\u0642\u0627\u0644\u0647\u060c \u0642\u0635\u062f \u062f\u0627\u0631\u06cc\u0645 \u0628\u0631\u0631\u0633\u06cc \u06a9\u0646\u06cc\u0645 \u06a9\u0647 \u0627\u06cc\u0646 \u0645\u0641\u0647\u0648\u0645 \u062f\u0642\u06cc\u0642\u0627\u064b \u0634\u0627\u0645\u0644 \u0686\u0647 \u0645\u0648\u0627\u0631\u062f\u06cc \u0627\u0633\u062a \u0648 \u0686\u0647 \u062a\u0623\u062b\u06cc\u0631\u06cc<\/span> (\u062f\u0631<\/span> \u0635\u0648\u0631\u062a \u0648\u062c\u0648\u062f) \u0628\u0631 \u0639\u0645\u0644\u06a9\u0631\u062f \u06cc\u0627 \u0631\u0641\u062a\u0627\u0631 \u06a9\u0644\u06cc \u0634\u0628\u06a9\u0647\u200c\u0647\u0627\u06cc \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u062f\u0627\u0631\u062f.<\/span><\/div>\n
<\/div>\n
\u067e\u06cc\u0634\u200c\u0646\u06cc\u0627\u0632\u0647\u0627<\/span><\/div>\n
\u2022 \u067e\u0627\u06cc\u062a\u0648\u0646: \u0628\u0631\u0627\u06cc \u0627\u062c\u0631\u0627\u06cc \u06a9\u062f\u0647\u0627\u06cc \u0627\u06cc\u0646 \u0645\u0642\u0627\u0644\u0647\u060c \u0633\u06cc\u0633\u062a\u0645 \u0634\u0645\u0627 \u0628\u0627\u06cc\u062f \u067e\u0627\u06cc\u062a\u0648\u0646 \u0631\u0627 \u0646\u0635\u0628 \u062f\u0627\u0634\u062a\u0647 \u0628\u0627\u0634\u062f. \u062e\u0648\u0627\u0646\u0646\u062f\u06af\u0627\u0646 \u0628\u0627\u06cc\u062f \u062a\u062c\u0631\u0628\u0647\u200c\u06cc \u0627\u0628\u062a\u062f\u0627\u06cc\u06cc \u0628\u0631\u0646\u0627\u0645\u0647\u200c\u0646\u0648\u06cc\u0633\u06cc \u0628\u0627 \u067e\u0627\u06cc\u062a\u0648\u0646 \u0631\u0627 \u062f\u0627\u0634\u062a\u0647 \u0628\u0627\u0634\u0646\u062f.<\/span><\/div>\n
\u2022 \u0645\u0641\u0627\u0647\u06cc\u0645 \u067e\u0627\u06cc\u0647 \u06cc\u0627\u062f\u06af\u06cc\u0631\u06cc \u0639\u0645\u06cc\u0642: \u0627\u06cc\u0646 \u0645\u0642\u0627\u0644\u0647 \u0645\u0641\u0627\u0647\u06cc\u0645\u06cc \u0631\u0627 \u067e\u0648\u0634\u0634 \u0645\u06cc\u200c\u062f\u0647\u062f \u06a9\u0647 \u0628\u0631\u0627\u06cc \u0627\u0639\u0645\u0627\u0644 \u062a\u0626\u0648\u0631\u06cc \u06cc\u0627\u062f\u06af\u06cc\u0631\u06cc \u0639\u0645\u06cc\u0642 \u0636\u0631\u0648\u0631\u06cc \u0647\u0633\u062a\u0646\u062f. \u0627\u0632 \u062e\u0648\u0627\u0646\u0646\u062f\u06af\u0627\u0646 \u0627\u0646\u062a\u0638\u0627\u0631 \u0645\u06cc\u200c\u0631\u0648\u062f \u06a9\u0647 \u0628\u0627 \u0627\u0635\u0637\u0644\u0627\u062d\u0627\u062a \u0648 \u0645\u0628\u0627\u0646\u06cc \u0627\u0648\u0644\u06cc\u0647\u200c\u06cc \u0627\u06cc\u0646 \u062d\u0648\u0632\u0647 \u0622\u0634\u0646\u0627\u06cc\u06cc \u062f\u0627\u0634\u062a\u0647 \u0628\u0627\u0634\u0646\u062f.<\/span><\/div>\n
<\/div>\n
\u0627\u0635\u0637\u0644\u0627\u062d \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc<\/span><\/div>\n
<\/div>\n
import torch \u00a0<\/span><\/div>\n
import torch.nn as nn \u00a0<\/span><\/div>\n
import torch.nn.functional as F \u00a0<\/span><\/div>\n
import torchvision \u00a0<\/span><\/div>\n
import torchvision.transforms as transforms \u00a0<\/span><\/div>\n
import torchvision.datasets as Datasets \u00a0<\/span><\/div>\n
from torch.utils.data import Dataset, DataLoader \u00a0<\/span><\/div>\n
import numpy as np \u00a0<\/span><\/div>\n
import matplotlib.pyplot as plt \u00a0<\/span><\/div>\n
import cv2 \u00a0<\/span><\/div>\n
from tqdm.notebook import tqdm \u00a0<\/span><\/div>\n
import seaborn as sns \u00a0<\/span><\/div>\n
from torchvision.utils import make_grid \u00a0<\/span><\/div>\n
<\/div>\n
if torch.cuda.is_available(): \u00a0<\/span><\/div>\n
\u00a0 device = torch.device(‘cuda:0’) \u00a0<\/span><\/div>\n
\u00a0 print(‘Running on the GPU’) \u00a0<\/span><\/div>\n
else: \u00a0<\/span><\/div>\n
\u00a0 device = torch.device(‘cpu’) \u00a0<\/span><\/div>\n
\u00a0 print(‘Running on the CPU’) \u00a0<\/span><\/div>\n
\u00a0\u00a0<\/span><\/div>\n
\u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0631 \u0639\u0644\u0645 \u0622\u0645\u0627\u0631 \u0628\u0647 \u0641\u0631\u0627\u06cc\u0646\u062f \u0645\u062d\u062f\u0648\u062f \u06a9\u0631\u062f\u0646 \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u06cc\u0627 \u06cc\u06a9 \u0645\u062c\u0645\u0648\u0639\u0647 \u0627\u0632 \u0645\u0642\u0627\u062f\u06cc\u0631 \u062f\u0631 \u0628\u0627\u0632\u0647\u200c\u06cc \u06f0 \u062a\u0627 \u06f1 \u0627\u0634\u0627\u0631\u0647 \u062f\u0627\u0631\u062f. \u0627\u0645\u0627 \u0628\u0647\u200c\u0635\u0648\u0631\u062a \u0646\u0627\u062e\u0648\u0634\u0627\u06cc\u0646\u062f\u06cc\u060c \u062f\u0631 \u0628\u0631\u062e\u06cc \u0645\u0646\u0627\u0628\u0639\u060c \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0647\u0645\u0686\u0646\u06cc\u0646 \u0628\u0647 \u0641\u0631\u0627\u06cc\u0646\u062f \u062a\u0646\u0638\u06cc\u0645 \u0645\u06cc\u0627\u0646\u06af\u06cc\u0646 \u06cc\u06a9 \u062a\u0648\u0632\u06cc\u0639 \u062f\u0627\u062f\u0647 \u0631\u0648\u06cc \u0635\u0641\u0631 \u0648 \u0627\u0646\u062d\u0631\u0627\u0641 \u0645\u0639\u06cc\u0627\u0631 \u0622\u0646 \u0631\u0648\u06cc \u06f1 \u0646\u06cc\u0632 \u0627\u0637\u0644\u0627\u0642 \u0645\u06cc\u200c\u0634\u0648\u062f.<\/span><\/div>\n
<\/div>\n
\u062f\u0631 \u062d\u0642\u06cc\u0642\u062a\u060c \u0627\u06cc\u0646 \u0641\u0631\u0627\u06cc\u0646\u062f \u06a9\u0647 \u062f\u0631 \u0622\u0646 \u0645\u06cc\u0627\u0646\u06af\u06cc\u0646 \u062a\u0648\u0632\u06cc\u0639 \u0631\u0648\u06cc \u06f0 \u062a\u0646\u0638\u06cc\u0645 \u0645\u06cc\u200c\u0634\u0648\u062f \u0648 \u0627\u0646\u062d\u0631\u0627\u0641 \u0645\u0639\u06cc\u0627\u0631 \u0622\u0646 \u0628\u0631\u0627\u0628\u0631 \u06f1 \u0642\u0631\u0627\u0631 \u0645\u06cc\u200c\u06af\u06cc\u0631\u062f\u060c \u0627\u0633\u062a\u0627\u0646\u062f\u0627\u0631\u062f\u0633\u0627\u0632\u06cc \u0646\u0627\u0645 \u062f\u0627\u0631\u062f. \u0628\u0627\u0627\u06cc\u0646\u200c\u062d\u0627\u0644\u060c \u0628\u0647 \u062f\u0644\u06cc\u0644 \u0628\u0631\u062e\u06cc \u0645\u0644\u0627\u062d\u0638\u0627\u062a\u060c \u0628\u0647 \u0622\u0646 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u06cc\u0627 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc z-score \u0646\u06cc\u0632 \u06af\u0641\u062a\u0647 \u0645\u06cc\u200c\u0634\u0648\u062f. \u0645\u0647\u0645 \u0627\u0633\u062a \u06a9\u0647 \u0627\u06cc\u0646 \u062a\u0645\u0627\u06cc\u0632 \u0631\u0627 \u06cc\u0627\u062f \u0628\u06af\u06cc\u0631\u06cc\u0645 \u0648 \u0628\u0647 \u062e\u0627\u0637\u0631 \u0628\u0633\u067e\u0627\u0631\u06cc\u0645.<\/span><\/div>\n
<\/div>\n
\u067e\u06cc\u0634\u200c\u067e\u0631\u062f\u0627\u0632\u0634 \u062f\u0627\u062f\u0647\u200c\u0647\u0627<\/span><\/div>\n
<\/div>\n
\u067e\u06cc\u0634\u200c\u067e\u0631\u062f\u0627\u0632\u0634 \u062f\u0627\u062f\u0647 \u0628\u0647 \u0645\u0631\u0627\u062d\u0644\u06cc \u0627\u0634\u0627\u0631\u0647 \u062f\u0627\u0631\u062f \u06a9\u0647 \u0642\u0628\u0644 \u0627\u0632 \u0648\u0631\u0648\u062f\u06cc \u062f\u0627\u062f\u0647 \u0628\u0647 \u06cc\u06a9 \u0627\u0644\u06af\u0648\u0631\u06cc\u062a\u0645 \u06cc\u0627\u062f\u06af\u06cc\u0631\u06cc \u0645\u0627\u0634\u06cc\u0646 \u06cc\u0627 \u06cc\u0627\u062f\u06af\u06cc\u0631\u06cc \u0639\u0645\u06cc\u0642 \u0627\u0646\u062c\u0627\u0645 \u0645\u06cc\u200c\u0634\u0648\u0646\u062f. \u062f\u0648 \u0641\u0631\u0627\u06cc\u0646\u062f \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0648 \u0627\u0633\u062a\u0627\u0646\u062f\u0627\u0631\u062f\u0633\u0627\u0632\u06cc \u06a9\u0647 \u062f\u0631 \u0628\u062e\u0634 \u0642\u0628\u0644 \u0630\u06a9\u0631 \u0634\u062f\u0646\u062f\u060c \u0627\u0632 \u0645\u0631\u0627\u062d\u0644 \u067e\u06cc\u0634\u200c\u067e\u0631\u062f\u0627\u0632\u0634 \u062f\u0627\u062f\u0647 \u0645\u062d\u0633\u0648\u0628 \u0645\u06cc\u200c\u0634\u0648\u0646\u062f.<\/span><\/div>\n
<\/div>\n
\u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0645\u06cc\u0646-\u0645\u0627\u06a9\u0633<\/span> (Min-max<\/span> Normalization)<\/span><\/div>\n
<\/div>\n
\u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0645\u06cc\u0646-\u0645\u0627\u06a9\u0633 \u06cc\u06a9\u06cc \u0627\u0632 \u0631\u0627\u06cc\u062c\u200c\u062a\u0631\u06cc\u0646 \u0631\u0648\u0634\u200c\u0647\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0627\u0633\u062a. \u0647\u0645\u0627\u0646\u200c\u0637\u0648\u0631 \u06a9\u0647 \u0627\u0632 \u0646\u0627\u0645\u0634 \u067e\u06cc\u062f\u0627\u0633\u062a\u060c \u0627\u06cc\u0646 \u0631\u0648\u0634 \u0646\u0642\u0627\u0637 \u062f\u0627\u062f\u0647 \u0631\u0627 \u062f\u0631 \u0628\u0627\u0632\u0647\u200c\u06cc \u06f0 \u062a\u0627 \u06f1 \u0645\u062d\u062f\u0648\u062f \u0645\u06cc\u200c\u06a9\u0646\u062f. \u0628\u0647 \u0627\u06cc\u0646 \u0635\u0648\u0631\u062a \u06a9\u0647 \u0645\u0642\u062f\u0627\u0631 \u062d\u062f\u0627\u0642\u0644<\/span> (min)<\/span> \u062f\u0631 \u0645\u062c\u0645\u0648\u0639\u0647 \u062f\u0627\u062f\u0647 \u0628\u0647 \u06f0 \u062a\u0646\u0638\u06cc\u0645 \u0645\u06cc\u200c\u0634\u0648\u062f\u060c \u0645\u0642\u062f\u0627\u0631 \u062d\u062f\u0627\u06a9\u062b\u0631<\/span> (max)<\/span> \u0628\u0647 \u06f1 \u062a\u0628\u062f\u06cc\u0644 \u0645\u06cc\u200c\u0634\u0648\u062f\u060c \u0648 \u0633\u0627\u06cc\u0631 \u0645\u0642\u0627\u062f\u06cc\u0631 \u0645\u062a\u0646\u0627\u0633\u0628 \u0628\u0627 \u0627\u06cc\u0646 \u0628\u0627\u0632\u0647 \u0645\u0642\u06cc\u0627\u0633\u200c\u0628\u0646\u062f\u06cc \u0645\u06cc\u200c\u0634\u0648\u0646\u062f.<\/span><\/div>\n
<\/div>\n
\u0641\u0631\u0645\u0648\u0644 \u0632\u06cc\u0631 \u062a\u0648\u0635\u06cc\u0641 \u0631\u06cc\u0627\u0636\u06cc \u0627\u06cc\u0646 \u0641\u0631\u0627\u06cc\u0646\u062f \u0631\u0627 \u0627\u0631\u0627\u0626\u0647 \u0645\u06cc\u200c\u062f\u0647\u062f. \u0627\u0633\u0627\u0633\u0627\u064b\u060c \u0627\u06cc\u0646 \u0641\u0631\u0627\u06cc\u0646\u062f \u0634\u0627\u0645\u0644 \u06a9\u0645 \u06a9\u0631\u062f\u0646 \u0645\u0642\u062f\u0627\u0631 \u062d\u062f\u0627\u0642\u0644 \u062f\u0627\u062f\u0647 \u0627\u0632 \u0647\u0631 \u0646\u0642\u0637\u0647 \u062f\u0627\u062f\u0647 \u0648 \u0633\u067e\u0633 \u062a\u0642\u0633\u06cc\u0645 \u0622\u0646 \u0628\u0631 \u062f\u0627\u0645\u0646\u0647\u200c\u06cc<\/span> (\u062d\u062f\u0627\u06a9\u062b\u0631<\/span> – \u062d\u062f\u0627\u0642\u0644) \u0627\u0633\u062a.<\/span><\/div>\n
<\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
<\/div>\n
\u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0627\u0632 \u062a\u0627\u0628\u0639 \u0628\u0631\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0645\u06cc\u0646-\u0645\u0627\u06a9\u0633<\/span><\/div>\n
<\/div>\n
\u0628\u0627 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0627\u0632 \u062a\u0627\u0628\u0639 \u0632\u06cc\u0631\u060c \u0645\u06cc\u200c\u062a\u0648\u0627\u0646\u06cc\u0645 \u0641\u0631\u0627\u06cc\u0646\u062f \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0645\u06cc\u0646-\u0645\u0627\u06a9\u0633 \u0631\u0627 \u0634\u0628\u06cc\u0647\u200c\u0633\u0627\u0632\u06cc \u06a9\u0646\u06cc\u0645. \u0627\u06cc\u0646 \u062a\u0627\u0628\u0639 \u0628\u0647 \u0645\u0627 \u06a9\u0645\u06a9 \u0645\u06cc\u200c\u06a9\u0646\u062f \u062a\u0627 \u062f\u0631\u06a9 \u0628\u0647\u062a\u0631\u06cc \u0627\u0632 \u0622\u0646\u0686\u0647 \u062f\u0631 \u067e\u0634\u062a\u200c\u0635\u062d\u0646\u0647 \u0627\u062a\u0641\u0627\u0642 \u0645\u06cc\u200c\u0627\u0641\u062a\u062f\u060c \u062f\u0627\u0634\u062a\u0647 \u0628\u0627\u0634\u06cc\u0645.<\/span><\/div>\n
<\/div>\n
def min_max_normalize(data_points: np.array):<\/span><\/div>\n
\u00a0 \u00a0<\/span> “””<\/span><\/div>\n
\u00a0 \u00a0 \u0627\u06cc\u0646 \u062a\u0627\u0628\u0639 \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0631\u0627 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u06a9\u0631\u062f\u0647 \u0648 \u0622\u0646\u200c\u0647\u0627 \u0631\u0627 \u062f\u0631 \u0628\u0627\u0632\u0647\u200c\u06cc 0 \u062a\u0627 1 \u0645\u062d\u062f\u0648\u062f \u0645\u06cc\u200c\u06a9\u0646\u062f.<\/span><\/div>\n
\u00a0 \u00a0<\/span> “””<\/span><\/div>\n
\u00a0 \u00a0 # \u062a\u0628\u062f\u06cc\u0644 \u0644\u06cc\u0633\u062a \u0628\u0647 \u0622\u0631\u0627\u06cc\u0647\u200c\u06cc numpy<\/span> (\u062f\u0631<\/span> \u0635\u0648\u0631\u062a \u0644\u0632\u0648\u0645)<\/span><\/div>\n
\u00a0 \u00a0 if type(data_points) == list:<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 data_points = np.array(data_points)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 # \u0627\u06cc\u062c\u0627\u062f \u0644\u06cc\u0633\u062a\u06cc \u0628\u0631\u0627\u06cc \u0646\u06af\u0647\u062f\u0627\u0631\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0634\u062f\u0647<\/span><\/div>\n
\u00a0 \u00a0 normalized =<\/span> []<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 # \u0645\u062d\u0627\u0633\u0628\u0647\u200c\u06cc \u062d\u062f\u0627\u0642\u0644 \u0648 \u062d\u062f\u0627\u06a9\u062b\u0631 \u0645\u0642\u062f\u0627\u0631 \u062f\u0631 \u062f\u0627\u062f\u0647\u200c\u0647\u0627<\/span><\/div>\n
\u00a0 \u00a0 minimum = data_points.min()<\/span><\/div>\n
\u00a0 \u00a0 maximum = data_points.max()<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 # \u062a\u0628\u062f\u06cc\u0644 \u0622\u0631\u0627\u06cc\u0647 \u0628\u0647 \u0644\u06cc\u0633\u062a \u0628\u0631\u0627\u06cc \u067e\u06cc\u0645\u0627\u06cc\u0634<\/span><\/div>\n
\u00a0 \u00a0 data_points = list(data_points)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 # \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627<\/span><\/div>\n
\u00a0 \u00a0 for value in data_points:<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 normalize =<\/span> (value<\/span> – minimum) \/<\/span> (maximum<\/span> – minimum)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 normalized.append(round(normalize, 2))<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 return np.array(normalized)<\/span><\/div>\n
\u00a0 \u00a0\u00a0<\/span><\/div>\n
\u0627\u06cc\u062c\u0627\u062f \u06cc\u06a9 \u0645\u062c\u0645\u0648\u0639\u0647 \u062a\u0635\u0627\u062f\u0641\u06cc \u0627\u0632 \u0645\u0642\u0627\u062f\u06cc\u0631 \u0648 \u0627\u062c\u0631\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0645\u06cc\u0646-\u0645\u0627\u06a9\u0633<\/span><\/div>\n
<\/div>\n
\u062d\u0627\u0644 \u06cc\u06a9 \u0645\u062c\u0645\u0648\u0639\u0647 \u0627\u0632 \u0645\u0642\u0627\u062f\u06cc\u0631 \u062a\u0635\u0627\u062f\u0641\u06cc \u0631\u0627 \u0628\u0627 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0627\u0632 NumPy \u0627\u06cc\u062c\u0627\u062f \u06a9\u0631\u062f\u0647 \u0648 \u0622\u0646\u200c\u0647\u0627 \u0631\u0627 \u0628\u0627 \u062a\u0627\u0628\u0639 min_max_normalize \u06a9\u0647 \u062f\u0631 \u0628\u0627\u0644\u0627 \u062a\u0639\u0631\u06cc\u0641 \u0634\u062f\u060c \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0645\u06cc\u200c\u06a9\u0646\u06cc\u0645.<\/span><\/div>\n
<\/div>\n
# \u0627\u06cc\u062c\u0627\u062f \u0645\u062c\u0645\u0648\u0639\u0647\u200c\u0627\u06cc \u0627\u0632 \u0646\u0642\u0627\u0637 \u062f\u0627\u062f\u0647 \u062a\u0635\u0627\u062f\u0641\u06cc<\/span><\/div>\n
data = np.random.rand(50) * 20<\/span><\/div>\n
<\/div>\n
# \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627<\/span><\/div>\n
normalized = min_max_normalize(data)<\/span><\/div>\n
\u0645\u0642\u0627\u06cc\u0633\u0647\u200c\u06cc \u062a\u0648\u0632\u06cc\u0639 \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0642\u0628\u0644 \u0648 \u0628\u0639\u062f \u0627\u0632 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc<\/span><\/div>\n
<\/div>\n
\u062f\u0631 \u0646\u0645\u0648\u062f\u0627\u0631\u0647\u0627\u06cc \u0632\u06cc\u0631 \u0645\u06cc\u200c\u062a\u0648\u0627\u0646 \u0645\u0634\u0627\u0647\u062f\u0647 \u06a9\u0631\u062f \u06a9\u0647 \u0642\u0628\u0644 \u0627\u0632 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc\u060c \u0645\u0642\u0627\u062f\u06cc\u0631 \u062f\u0631 \u0628\u0627\u0632\u0647\u200c\u06cc \u06f0 \u062a\u0627 \u06f2\u06f0 \u0642\u0631\u0627\u0631 \u062f\u0627\u0634\u062a\u0646\u062f \u0648 \u0628\u06cc\u0634\u062a\u0631 \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0628\u06cc\u0646 \u06f5 \u062a\u0627 \u06f1\u06f0 \u0628\u0648\u062f\u0646\u062f. \u0627\u0645\u0627 \u0628\u0639\u062f \u0627\u0632 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc\u060c \u0645\u0642\u062f\u0627\u0631\u0647\u0627 \u0628\u0647 \u0628\u0627\u0632\u0647\u200c\u06cc \u06f0 \u062a\u0627 \u06f1 \u0645\u062d\u062f\u0648\u062f \u0634\u062f\u0647\u200c\u0627\u0646\u062f \u0648 \u0627\u06a9\u062b\u0631\u06cc\u062a \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0628\u06cc\u0646 \u06f0.\u06f2\u06f5 \u062a\u0627 \u06f0.\u06f5 \u0642\u0631\u0627\u0631 \u06af\u0631\u0641\u062a\u0647\u200c\u0627\u0646\u062f.<\/span><\/div>\n
<\/div>\n
\u0646\u06a9\u062a\u0647: \u0647\u0646\u06af\u0627\u0645 \u0627\u062c\u0631\u0627\u06cc \u0627\u06cc\u0646 \u06a9\u062f\u060c \u062a\u0648\u0632\u06cc\u0639 \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0645\u0645\u06a9\u0646 \u0627\u0633\u062a \u0645\u062a\u0641\u0627\u0648\u062a \u0628\u0627\u0634\u062f\u060c \u0632\u06cc\u0631\u0627 \u0645\u0642\u0627\u062f\u06cc\u0631 \u0628\u0647\u200c\u0635\u0648\u0631\u062a \u062a\u0635\u0627\u062f\u0641\u06cc \u062a\u0648\u0644\u06cc\u062f \u0645\u06cc\u200c\u0634\u0648\u0646\u062f.<\/span><\/div>\n
<\/div>\n
# \u0646\u0645\u0627\u06cc\u0634 \u062a\u0648\u0632\u06cc\u0639 \u062f\u0627\u062f\u0647\u200c\u0647\u0627<\/span><\/div>\n
figure, axes = plt.subplots(1, 2, sharey=True, dpi=100)<\/span><\/div>\n
sns.histplot(data, ax=axes[0])<\/span><\/div>\n
axes[0].set_title(‘\u0628\u062f\u0648\u0646 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc’)<\/span><\/div>\n
sns.histplot(normalized, ax=axes[1])<\/span><\/div>\n
axes[1].set_title(‘\u0646\u0631\u0645\u0627\u0644\u200c\u0634\u062f\u0647 \u0628\u0627 \u0645\u06cc\u0646-\u0645\u0627\u06a9\u0633’)<\/span><\/div>\n
<\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
\u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc Z-score<\/span><\/div>\n
<\/div>\n
\u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc Z-score\u060c \u06a9\u0647 \u0647\u0645\u0686\u0646\u06cc\u0646 \u0627\u0633\u062a\u0627\u0646\u062f\u0627\u0631\u062f\u0633\u0627\u0632\u06cc \u0646\u0627\u0645\u06cc\u062f\u0647 \u0645\u06cc\u200c\u0634\u0648\u062f\u060c \u0641\u0631\u0627\u06cc\u0646\u062f \u062a\u0646\u0638\u06cc\u0645 \u0645\u06cc\u0627\u0646\u06af\u06cc\u0646 \u0648 \u0627\u0646\u062d\u0631\u0627\u0641 \u0645\u0639\u06cc\u0627\u0631 \u06cc\u06a9 \u062a\u0648\u0632\u06cc\u0639 \u062f\u0627\u062f\u0647 \u0628\u0647 \u062a\u0631\u062a\u06cc\u0628 \u0628\u0631\u0627\u0628\u0631 \u0628\u0627 0 \u0648 1 \u0627\u0633\u062a. \u0645\u0639\u0627\u062f\u0644\u0647 \u0632\u06cc\u0631\u060c \u0645\u0639\u0627\u062f\u0644\u0647 \u0631\u06cc\u0627\u0636\u06cc \u0627\u0633\u062a \u06a9\u0647 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc Z-score \u0631\u0627 \u062a\u0639\u06cc\u06cc\u0646 \u0645\u06cc\u200c\u06a9\u0646\u062f. \u0627\u06cc\u0646 \u0634\u0627\u0645\u0644 \u06a9\u0645 \u06a9\u0631\u062f\u0646 \u0645\u06cc\u0627\u0646\u06af\u06cc\u0646 \u062a\u0648\u0632\u06cc\u0639 \u0627\u0632 \u0645\u0642\u062f\u0627\u0631 \u0645\u0648\u0631\u062f\u0646\u0638\u0631 \u0628\u0631\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0642\u0628\u0644 \u0627\u0632 \u062a\u0642\u0633\u06cc\u0645 \u0628\u0631 \u0627\u0646\u062d\u0631\u0627\u0641 \u0645\u0639\u06cc\u0627\u0631 \u062a\u0648\u0632\u06cc\u0639 \u0627\u0633\u062a.<\/span><\/div>\n
<\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
\u062a\u0627\u0628\u0639 \u062a\u0639\u0631\u06cc\u0641\u200c\u0634\u062f\u0647 \u062f\u0631 \u0632\u06cc\u0631\u060c \u0641\u0631\u0627\u06cc\u0646\u062f \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc Z-score \u0631\u0627 \u0634\u0628\u06cc\u0647\u200c\u0633\u0627\u0632\u06cc \u0645\u06cc\u200c\u06a9\u0646\u062f. \u0628\u0627 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0627\u0632 \u0627\u06cc\u0646 \u062a\u0627\u0628\u0639\u060c \u0645\u06cc\u200c\u062a\u0648\u0627\u0646\u06cc\u0645 \u0646\u06af\u0627\u0647\u06cc \u062f\u0642\u06cc\u0642\u200c\u062a\u0631 \u0628\u0647 \u0627\u06cc\u0646 \u0641\u0631\u0627\u06cc\u0646\u062f \u062f\u0627\u0634\u062a\u0647 \u0628\u0627\u0634\u06cc\u0645.<\/span><\/div>\n
def z_score_normalize(data_points: np.array):<\/span><\/div>\n
\u00a0<\/span> “””<\/span><\/div>\n
\u00a0 \u0627\u06cc\u0646 \u062a\u0627\u0628\u0639 \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0631\u0627 \u0628\u0627 \u0645\u062d\u0627\u0633\u0628\u0647 Z-score \u0622\u0646\u200c\u0647\u0627 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0645\u06cc\u200c\u06a9\u0646\u062f.<\/span><\/div>\n
\u00a0<\/span> “””<\/span><\/div>\n
\u00a0 # \u062a\u0628\u062f\u06cc\u0644 \u0644\u06cc\u0633\u062a \u0628\u0647 \u0622\u0631\u0627\u06cc\u0647\u200c\u06cc NumPy<\/span><\/div>\n
\u00a0 if type(data_points) == list:<\/span><\/div>\n
\u00a0 \u00a0 data_points = np.array(data_points)<\/span><\/div>\n
\u00a0 else:<\/span><\/div>\n
\u00a0 \u00a0 pass<\/span><\/div>\n
<\/div>\n
\u00a0 # \u0627\u06cc\u062c\u0627\u062f \u0644\u06cc\u0633\u062a\u06cc \u0628\u0631\u0627\u06cc \u0630\u062e\u06cc\u0631\u0647 \u062f\u0627\u062f\u0647\u200c\u0647\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0634\u062f\u0647<\/span><\/div>\n
\u00a0 normalized =<\/span> []<\/span><\/div>\n
<\/div>\n
\u00a0 # \u0645\u062d\u0627\u0633\u0628\u0647 \u0645\u06cc\u0627\u0646\u06af\u06cc\u0646 \u0648 \u0627\u0646\u062d\u0631\u0627\u0641 \u0645\u0639\u06cc\u0627\u0631<\/span><\/div>\n
\u00a0 mean = data_points.mean()<\/span><\/div>\n
\u00a0 std = data_points.std()<\/span><\/div>\n
<\/div>\n
\u00a0 # \u062a\u0628\u062f\u06cc\u0644 \u0628\u0647 \u0644\u06cc\u0633\u062a \u0628\u0631\u0627\u06cc \u0627\u0646\u062c\u0627\u0645 \u0639\u0645\u0644\u06cc\u0627\u062a \u067e\u06cc\u0645\u0627\u06cc\u0634<\/span><\/div>\n
\u00a0 data_points = list(data_points)<\/span><\/div>\n
<\/div>\n
\u00a0 # \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627<\/span><\/div>\n
\u00a0 for value in data_points:<\/span><\/div>\n
\u00a0 \u00a0 normalize =<\/span> (value<\/span> – mean) \/ std<\/span><\/div>\n
\u00a0 \u00a0 normalized.append(round(normalize, 2))<\/span><\/div>\n
<\/div>\n
\u00a0 return np.array(normalized)<\/span><\/div>\n
\u0628\u0627 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0627\u0632 \u062a\u0648\u0632\u06cc\u0639 \u062f\u0627\u062f\u0647\u200c\u0627\u06cc \u06a9\u0647 \u062f\u0631 \u0628\u062e\u0634 \u0642\u0628\u0644\u06cc \u062a\u0648\u0644\u06cc\u062f \u0634\u062f\u060c \u0628\u06cc\u0627\u06cc\u06cc\u062f \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0631\u0627 \u0628\u0627 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0627\u0632 \u062a\u0627\u0628\u0639 Z-score \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u06a9\u0646\u06cc\u0645.<\/span><\/div>\n
# \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627<\/span><\/div>\n
z_normalized = z_score_normalize(data)<\/span><\/div>\n
<\/div>\n
# \u0628\u0631\u0631\u0633\u06cc \u0645\u0642\u062f\u0627\u0631 \u0645\u06cc\u0627\u0646\u06af\u06cc\u0646<\/span><\/div>\n
z_normalized.mean()<\/span><\/div>\n
>>>><\/span> -0.0006<\/span><\/div>\n
<\/div>\n
# \u0628\u0631\u0631\u0633\u06cc \u0645\u0642\u062f\u0627\u0631 \u0627\u0646\u062d\u0631\u0627\u0641 \u0645\u0639\u06cc\u0627\u0631<\/span><\/div>\n
z_normalized.std()<\/span><\/div>\n
>>>> 1.0000<\/span><\/div>\n
\u062f\u0648\u0628\u0627\u0631\u0647\u060c \u0628\u0627 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0627\u0632 \u0646\u0645\u0648\u062f\u0627\u0631\u0647\u0627\u06cc \u062a\u0648\u0632\u06cc\u0639\u060c \u0645\u06cc\u200c\u062a\u0648\u0627\u0646 \u0645\u0634\u0627\u0647\u062f\u0647 \u06a9\u0631\u062f \u06a9\u0647 \u0645\u0642\u0627\u062f\u06cc\u0631 \u062a\u0648\u0632\u06cc\u0639 \u0627\u0648\u0644\u06cc\u0647 \u062f\u0631 \u0628\u0627\u0632\u0647 0 \u062a\u0627 20 \u0642\u0631\u0627\u0631 \u062f\u0627\u0631\u0646\u062f\u060c \u062f\u0631 \u062d\u0627\u0644\u06cc \u06a9\u0647 \u0645\u0642\u0627\u062f\u06cc\u0631 \u0646\u0631\u0645\u0627\u0644\u200c\u0634\u062f\u0647 \u0628\u0627 Z-score \u0627\u06a9\u0646\u0648\u0646 \u062d\u0648\u0644 0 \u0645\u062a\u0645\u0631\u06a9\u0632 \u0634\u062f\u0647\u200c\u0627\u0646\u062f<\/span> (\u0645\u06cc\u0627\u0646\u06af\u06cc\u0646<\/span> \u0635\u0641\u0631) \u0648 \u062f\u0631 \u0628\u0627\u0632\u0647\u200c\u0627\u06cc \u062a\u0642\u0631\u06cc\u0628\u0627\u064b \u0628\u06cc\u0646<\/span> -1.5<\/span> \u062a\u0627 1.5 \u0642\u0631\u0627\u0631 \u06af\u0631\u0641\u062a\u0647\u200c\u0627\u0646\u062f \u06a9\u0647 \u06cc\u06a9 \u0628\u0627\u0632\u0647\u200c\u06cc \u0642\u0627\u0628\u0644 \u0645\u062f\u06cc\u0631\u06cc\u062a\u200c\u062a\u0631 \u0627\u0633\u062a.<\/span><\/div>\n
# \u0645\u0635\u0648\u0631\u0633\u0627\u0632\u06cc \u062a\u0648\u0632\u06cc\u0639\u200c\u0647\u0627<\/span><\/div>\n
figure, axes = plt.subplots(1, 2, sharey=True, dpi=100)<\/span><\/div>\n
sns.histplot(data, ax=axes[0])<\/span><\/div>\n
axes[0].set_title(‘unnormalized’)<\/span><\/div>\n
sns.histplot(z_normalized, ax=axes[1])<\/span><\/div>\n
axes[1].set_title(‘z-score normalized’)<\/span><\/div>\n
<\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
\u062f\u0644\u0627\u06cc\u0644 \u067e\u06cc\u0634\u200c\u067e\u0631\u062f\u0627\u0632\u0634 \u062f\u0627\u062f\u0647\u200c\u0647\u0627<\/span><\/div>\n
<\/div>\n
\u062f\u0631 \u06cc\u0627\u062f\u06af\u06cc\u0631\u06cc \u0645\u0627\u0634\u06cc\u0646\u060c \u062f\u0627\u062f\u0647\u200c\u0647\u0627\u06cc \u0648\u0631\u0648\u062f\u06cc \u0631\u0627 \u0628\u0647\u200c\u0639\u0646\u0648\u0627\u0646 \u0648\u06cc\u0698\u06af\u06cc\u200c\u0647\u0627\u06cc \u062c\u062f\u0627\u06af\u0627\u0646\u0647 \u062f\u0631 \u0646\u0638\u0631 \u0645\u06cc\u200c\u06af\u06cc\u0631\u06cc\u0645. \u0645\u0639\u0645\u0648\u0644\u0627\u064b \u0627\u06cc\u0646 \u0648\u06cc\u0698\u06af\u06cc\u200c\u0647\u0627 \u062f\u0631 \u06cc\u06a9 \u0645\u0642\u06cc\u0627\u0633 \u06cc\u06a9\u0633\u0627\u0646 \u0642\u0631\u0627\u0631 \u0646\u062f\u0627\u0631\u0646\u062f. \u0628\u0647\u200c\u0639\u0646\u0648\u0627\u0646 \u0645\u062b\u0627\u0644\u060c \u06cc\u06a9 \u062e\u0627\u0646\u0647 \u0631\u0627 \u062f\u0631 \u0646\u0638\u0631 \u0628\u06af\u06cc\u0631\u06cc\u062f \u06a9\u0647 \u062f\u0627\u0631\u0627\u06cc \u06f3 \u0627\u062a\u0627\u0642 \u062e\u0648\u0627\u0628 \u0648 \u06cc\u06a9 \u0633\u0627\u0644\u0646 \u0646\u0634\u06cc\u0645\u0646 \u0628\u0647 \u0645\u0633\u0627\u062d\u062a \u06f4\u06f0\u06f0 \u0641\u0648\u062a \u0645\u0631\u0628\u0639 \u0627\u0633\u062a. \u0627\u06cc\u0646 \u062f\u0648 \u0648\u06cc\u0698\u06af\u06cc \u062f\u0631 \u0645\u0642\u06cc\u0627\u0633\u200c\u0647\u0627\u06cc\u06cc \u06a9\u0627\u0645\u0644\u0627\u064b \u0645\u062a\u0641\u0627\u0648\u062a \u0642\u0631\u0627\u0631 \u062f\u0627\u0631\u0646\u062f. \u062d\u0627\u0644 \u0627\u06af\u0631 \u0627\u06cc\u0646 \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0631\u0627 \u0628\u0647 \u06cc\u06a9 \u0627\u0644\u06af\u0648\u0631\u06cc\u062a\u0645 \u06cc\u0627\u062f\u06af\u06cc\u0631\u06cc \u0645\u0627\u0634\u06cc\u0646 \u06a9\u0647 \u062a\u0648\u0633\u0637 \u06af\u0631\u0627\u062f\u06cc\u0627\u0646 \u0646\u0632\u0648\u0644\u06cc \u0628\u0647\u06cc\u0646\u0647\u200c\u0633\u0627\u0632\u06cc \u0645\u06cc\u200c\u0634\u0648\u062f \u0648\u0627\u0631\u062f \u06a9\u0646\u06cc\u0645\u060c \u0641\u0631\u0622\u06cc\u0646\u062f \u0628\u0647\u06cc\u0646\u0647\u200c\u0633\u0627\u0632\u06cc \u062f\u0634\u0648\u0627\u0631 \u062e\u0648\u0627\u0647\u062f \u0634\u062f\u061b \u0632\u06cc\u0631\u0627 \u0648\u06cc\u0698\u06af\u06cc \u0628\u0627 \u0645\u0642\u06cc\u0627\u0633 \u0628\u0632\u0631\u06af\u200c\u062a\u0631 \u062a\u0623\u062b\u06cc\u0631 \u0628\u06cc\u0634\u062a\u0631\u06cc \u0646\u0633\u0628\u062a \u0628\u0647 \u0633\u0627\u06cc\u0631 \u0648\u06cc\u0698\u06af\u06cc\u200c\u0647\u0627 \u062e\u0648\u0627\u0647\u062f \u062f\u0627\u0634\u062a. \u0628\u0646\u0627\u0628\u0631\u0627\u06cc\u0646\u060c \u0628\u0631\u0627\u06cc \u0633\u0627\u062f\u0647\u200c\u0633\u0627\u0632\u06cc \u0641\u0631\u0622\u06cc\u0646\u062f \u0628\u0647\u06cc\u0646\u0647\u200c\u0633\u0627\u0632\u06cc\u060c \u0628\u0647\u062a\u0631 \u0627\u0633\u062a \u0647\u0645\u0647 \u0648\u06cc\u0698\u06af\u06cc\u200c\u0647\u0627 \u0631\u0627 \u062f\u0631 \u06cc\u06a9 \u0645\u0642\u06cc\u0627\u0633 \u06cc\u06a9\u0633\u0627\u0646 \u0642\u0631\u0627\u0631 \u062f\u0647\u06cc\u0645.<\/span><\/div>\n
<\/div>\n
\u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0631 \u0644\u0627\u06cc\u0647\u200c\u0647\u0627\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc<\/span><\/div>\n
<\/div>\n
\u062f\u0631 \u06cc\u06a9 \u062a\u0635\u0648\u06cc\u0631\u060c \u062f\u0627\u062f\u0647\u200c\u0647\u0627\u06cc \u0648\u0631\u0648\u062f\u06cc \u0647\u0645\u0627\u0646 \u0645\u0642\u0627\u062f\u06cc\u0631 \u067e\u06cc\u06a9\u0633\u0644\u06cc \u0622\u0646 \u0647\u0633\u062a\u0646\u062f. \u0645\u0642\u062f\u0627\u0631 \u067e\u06cc\u06a9\u0633\u0644\u200c\u0647\u0627 \u0645\u0639\u0645\u0648\u0644\u0627\u064b \u0628\u06cc\u0646 \u06f0 \u062a\u0627 \u06f2\u06f5\u06f5 \u0645\u062a\u063a\u06cc\u0631 \u0627\u0633\u062a. \u0628\u0647 \u0647\u0645\u06cc\u0646 \u062f\u0644\u06cc\u0644\u060c \u0642\u0628\u0644 \u0627\u0632 \u0648\u0627\u0631\u062f \u06a9\u0631\u062f\u0646 \u062a\u0635\u0627\u0648\u06cc\u0631 \u0628\u0647 \u06cc\u06a9 \u0634\u0628\u06a9\u0647 \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc\u060c \u0628\u0647\u062a\u0631 \u0627\u0633\u062a \u0622\u0646\u200c\u0647\u0627 \u0631\u0627 \u0628\u0647 \u06af\u0648\u0646\u0647\u200c\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u06a9\u0646\u06cc\u0645 \u06a9\u0647 \u062a\u0645\u0627\u0645 \u067e\u06cc\u06a9\u0633\u0644\u200c\u0647\u0627 \u062f\u0631 \u06cc\u06a9 \u0628\u0627\u0632\u0647 \u0642\u0627\u0628\u0644 \u06a9\u0646\u062a\u0631\u0644 \u0642\u0631\u0627\u0631 \u0628\u06af\u06cc\u0631\u0646\u062f.<\/span><\/div>\n
<\/div>\n
\u062d\u062a\u06cc \u0628\u0627 \u0627\u0646\u062c\u0627\u0645 \u0627\u06cc\u0646 \u06a9\u0627\u0631\u060c \u0647\u0646\u06af\u0627\u0645 \u0622\u0645\u0648\u0632\u0634 \u06cc\u06a9 \u0634\u0628\u06a9\u0647 \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc\u060c \u0645\u0642\u0627\u062f\u06cc\u0631 \u0648\u0632\u0646\u200c\u0647\u0627<\/span> (\u0639\u0646\u0627\u0635\u0631<\/span> \u0645\u0648\u062c\u0648\u062f \u062f\u0631 \u0641\u06cc\u0644\u062a\u0631\u0647\u0627) \u0645\u0645\u06a9\u0646 \u0627\u0633\u062a \u0628\u06cc\u0634\u200c\u0627\u0632\u062d\u062f \u0628\u0632\u0631\u06af \u0634\u0648\u0646\u062f \u0648 \u062f\u0631 \u0646\u062a\u06cc\u062c\u0647\u060c \u0646\u0642\u0634\u0647\u200c\u0647\u0627\u06cc \u0648\u06cc\u0698\u06af\u06cc\u200c\u0627\u06cc<\/span> (Feature<\/span> Maps) \u062a\u0648\u0644\u06cc\u062f \u06a9\u0646\u0646\u062f \u06a9\u0647 \u0645\u0642\u0627\u062f\u06cc\u0631 \u067e\u06cc\u06a9\u0633\u0644\u06cc \u0622\u0646\u200c\u0647\u0627 \u062f\u0631 \u06cc\u06a9 \u0628\u0627\u0632\u0647 \u0648\u0633\u06cc\u0639 \u067e\u0631\u0627\u06a9\u0646\u062f\u0647 \u0628\u0627\u0634\u062f. \u0627\u06cc\u0646 \u0627\u0645\u0631 \u0628\u0627\u0639\u062b \u0645\u06cc\u200c\u0634\u0648\u062f \u06a9\u0647 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0627\u0646\u062c\u0627\u0645\u200c\u0634\u062f\u0647 \u062f\u0631 \u0645\u0631\u062d\u0644\u0647 \u067e\u06cc\u0634\u200c\u067e\u0631\u062f\u0627\u0632\u0634 \u0628\u06cc\u200c\u0627\u062b\u0631 \u0634\u0648\u062f. \u0639\u0644\u0627\u0648\u0647 \u0628\u0631 \u0627\u06cc\u0646\u060c \u0686\u0646\u06cc\u0646 \u0648\u0636\u0639\u06cc\u062a\u06cc \u0645\u06cc\u200c\u062a\u0648\u0627\u0646\u062f \u0641\u0631\u0622\u06cc\u0646\u062f \u0628\u0647\u06cc\u0646\u0647\u200c\u0633\u0627\u0632\u06cc \u0631\u0627 \u06a9\u0646\u062f \u06a9\u0631\u062f\u0647 \u06cc\u0627 \u062f\u0631 \u0645\u0648\u0627\u0631\u062f \u0634\u062f\u06cc\u062f\u060c \u0645\u0646\u062c\u0631 \u0628\u0647 \u0645\u0634\u06a9\u0644<\/span> \u00ab\u06af\u0631\u0627\u062f\u06cc\u0627\u0646\u200c\u0647\u0627\u06cc<\/span> \u0646\u0627\u067e\u0627\u06cc\u062f\u0627\u0631\u00bb<\/span> (Unstable<\/span> Gradients) \u0634\u0648\u062f\u060c \u06a9\u0647 \u0645\u0645\u06a9\u0646 \u0627\u0633\u062a \u0645\u0627\u0646\u0639 \u0627\u0632 \u0628\u0647\u06cc\u0646\u0647\u200c\u0633\u0627\u0632\u06cc \u0628\u06cc\u0634\u062a\u0631 \u0648\u0632\u0646\u200c\u0647\u0627\u06cc \u0634\u0628\u06a9\u0647 \u0634\u0648\u062f.<\/span><\/div>\n
<\/div>\n
\u0641\u0631\u0622\u06cc\u0646\u062f \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc<\/span> (Batch<\/span> Normalization)<\/span><\/div>\n
<\/div>\n
\u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc<\/span> (Batch<\/span> Normalization) \u0627\u0633\u0627\u0633\u0627\u064b \u0645\u0642\u0627\u062f\u06cc\u0631 \u067e\u06cc\u06a9\u0633\u0644\u06cc \u062a\u0645\u0627\u0645 \u0646\u0642\u0634\u0647\u200c\u0647\u0627\u06cc \u0648\u06cc\u0698\u06af\u06cc \u062f\u0631 \u06cc\u06a9 \u0644\u0627\u06cc\u0647 \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u0631\u0627 \u0628\u0647 \u06cc\u06a9 \u0645\u06cc\u0627\u0646\u06af\u06cc\u0646 \u0648 \u0627\u0646\u062d\u0631\u0627\u0641 \u0645\u0639\u06cc\u0627\u0631 \u062c\u062f\u06cc\u062f \u062a\u0646\u0638\u06cc\u0645 \u0645\u06cc\u200c\u06a9\u0646\u062f. \u0628\u0647\u200c\u0637\u0648\u0631 \u0645\u0639\u0645\u0648\u0644\u060c \u0627\u06cc\u0646 \u0641\u0631\u0622\u06cc\u0646\u062f \u0628\u0627 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc Z-score \u0647\u0645\u0647 \u067e\u06cc\u06a9\u0633\u0644\u200c\u0647\u0627 \u0622\u063a\u0627\u0632 \u0645\u06cc\u200c\u0634\u0648\u062f \u0648 \u0633\u067e\u0633 \u0645\u0642\u0627\u062f\u06cc\u0631 \u0646\u0631\u0645\u0627\u0644\u200c\u0634\u062f\u0647 \u062f\u0631 \u06cc\u06a9 \u067e\u0627\u0631\u0627\u0645\u062a\u0631 \u062f\u0644\u062e\u0648\u0627\u0647 \u0622\u0644\u0641\u0627<\/span> (\u0645\u0642\u06cc\u0627\u0633)<\/span> \u0636\u0631\u0628 \u0634\u062f\u0647 \u0648 \u062f\u0631 \u0646\u0647\u0627\u06cc\u062a\u060c \u06cc\u06a9 \u067e\u0627\u0631\u0627\u0645\u062a\u0631 \u062f\u0644\u062e\u0648\u0627\u0647 \u0628\u062a\u0627<\/span> (\u0627\u0641\u0633\u062a)<\/span> \u0628\u0647 \u0622\u0646 \u0627\u0636\u0627\u0641\u0647 \u0645\u06cc\u200c\u0634\u0648\u062f.<\/span><\/div>\n
<\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
\u0627\u06cc\u0646 \u062f\u0648 \u067e\u0627\u0631\u0627\u0645\u062a\u0631\u060c \u0622\u0644\u0641\u0627 \u0648 \u0628\u062a\u0627\u060c \u067e\u0627\u0631\u0627\u0645\u062a\u0631\u0647\u0627\u06cc \u0642\u0627\u0628\u0644 \u06cc\u0627\u062f\u06af\u06cc\u0631\u06cc \u0647\u0633\u062a\u0646\u062f \u06a9\u0647 \u0634\u0628\u06a9\u0647 \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc<\/span> (ConvNet)<\/span> \u0627\u0632 \u0622\u0646\u200c\u0647\u0627 \u0628\u0631\u0627\u06cc \u0627\u0637\u0645\u06cc\u0646\u0627\u0646 \u0627\u0632 \u0642\u0631\u0627\u0631 \u06af\u0631\u0641\u062a\u0646 \u0645\u0642\u0627\u062f\u06cc\u0631 \u067e\u06cc\u06a9\u0633\u0644\u200c\u0647\u0627 \u062f\u0631 \u0646\u0642\u0634\u0647\u200c\u0647\u0627\u06cc \u0648\u06cc\u0698\u06af\u06cc \u062f\u0631 \u06cc\u06a9 \u0628\u0627\u0632\u0647 \u0642\u0627\u0628\u0644 \u06a9\u0646\u062a\u0631\u0644 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0645\u06cc\u200c\u06a9\u0646\u062f. \u0627\u06cc\u0646 \u06a9\u0627\u0631\u060c \u0645\u0634\u06a9\u0644 \u06af\u0631\u0627\u062f\u06cc\u0627\u0646\u200c\u0647\u0627\u06cc \u0646\u0627\u067e\u0627\u06cc\u062f\u0627\u0631 \u0631\u0627 \u06a9\u0627\u0647\u0634 \u0645\u06cc\u200c\u062f\u0647\u062f.<\/span><\/div>\n
<\/div>\n
\u0627\u062c\u0631\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc<\/span> (Batch<\/span> Normalization in Action)<\/span><\/div>\n
<\/div>\n
\u0628\u0631\u0627\u06cc \u0627\u0631\u0632\u06cc\u0627\u0628\u06cc \u0648\u0627\u0642\u0639\u06cc \u062a\u0623\u062b\u06cc\u0631 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u062f\u0631 \u0644\u0627\u06cc\u0647\u200c\u0647\u0627\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc\u060c \u0628\u0627\u06cc\u062f \u062f\u0648 \u0634\u0628\u06a9\u0647 \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u0631\u0627 \u0645\u0642\u0627\u06cc\u0633\u0647 \u06a9\u0646\u06cc\u0645: \u06cc\u06a9\u06cc \u0628\u062f\u0648\u0646 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0648 \u062f\u06cc\u06af\u0631\u06cc \u0628\u0627 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc. \u0628\u0631\u0627\u06cc \u0627\u06cc\u0646 \u0645\u0646\u0638\u0648\u0631\u060c \u0627\u0632 \u0645\u0639\u0645\u0627\u0631\u06cc LeNet-5 \u0648 \u0645\u062c\u0645\u0648\u0639\u0647 \u062f\u0627\u062f\u0647 MNIST \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u062e\u0648\u0627\u0647\u06cc\u0645 \u06a9\u0631\u062f.<\/span><\/div>\n
<\/div>\n
\u0645\u062c\u0645\u0648\u0639\u0647 \u062f\u0627\u062f\u0647 \u0648 \u06a9\u0644\u0627\u0633 \u0634\u0628\u06a9\u0647 \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc<\/span><\/div>\n
<\/div>\n
\u0647\u0645\u0627\u0646\u200c\u0637\u0648\u0631 \u06a9\u0647 \u0642\u0628\u0644\u0627\u064b \u0627\u0634\u0627\u0631\u0647 \u0634\u062f\u060c \u062f\u0631 \u0627\u06cc\u0646 \u0645\u0642\u0627\u0644\u0647 \u0627\u0632 \u0645\u062c\u0645\u0648\u0639\u0647 \u062f\u0627\u062f\u0647 MNIST \u0628\u0631\u0627\u06cc \u0627\u0646\u062c\u0627\u0645 \u0622\u0632\u0645\u0627\u06cc\u0634\u200c\u0647\u0627 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u062e\u0648\u0627\u0647\u062f \u0634\u062f. \u0627\u06cc\u0646 \u0645\u062c\u0645\u0648\u0639\u0647 \u062f\u0627\u062f\u0647 \u0634\u0627\u0645\u0644 \u062a\u0635\u0627\u0648\u06cc\u0631\u06cc \u0628\u0627 \u0627\u0628\u0639\u0627\u062f \u06f2\u06f8 \u00d7 \u06f2\u06f8 \u067e\u06cc\u06a9\u0633\u0644 \u0627\u0632 \u0627\u0631\u0642\u0627\u0645 \u062f\u0633\u062a\u200c\u0646\u0648\u06cc\u0633 \u0627\u0633\u062a \u06a9\u0647 \u0645\u0642\u0627\u062f\u06cc\u0631 \u0622\u0646\u200c\u0647\u0627 \u0627\u0632 \u06f0 \u062a\u0627 \u06f9 \u0645\u062a\u063a\u06cc\u0631 \u0628\u0648\u062f\u0647 \u0648 \u0628\u0647\u200c\u0637\u0648\u0631 \u0645\u0646\u0627\u0633\u0628 \u0628\u0631\u0686\u0633\u0628\u200c\u06af\u0630\u0627\u0631\u06cc \u0634\u062f\u0647\u200c\u0627\u0646\u062f.<\/span><\/div>\n
<\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
\u0627\u06cc\u0646 \u06a9\u062f \u0634\u0627\u0645\u0644 \u062f\u0648 \u0628\u062e\u0634 \u0627\u0635\u0644\u06cc \u0627\u0633\u062a:<\/span><\/div>\n
\u06f1. \u0628\u0627\u0631\u06af\u06cc\u0631\u06cc \u0645\u062c\u0645\u0648\u0639\u0647 \u062f\u0627\u062f\u0647 MNIST<\/span><\/div>\n
\u06f2. \u062a\u0639\u0631\u06cc\u0641 \u06a9\u0644\u0627\u0633 \u0634\u0628\u06a9\u0647 \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc<\/span> (Convolutional<\/span> Neural Network – CNN) \u0648 \u0622\u0645\u0648\u0632\u0634 \u0622\u0646<\/span><\/div>\n
<\/div>\n
\u06f1. \u0628\u0627\u0631\u06af\u06cc\u0631\u06cc \u0645\u062c\u0645\u0648\u0639\u0647 \u062f\u0627\u062f\u0647 MNIST \u062f\u0631 PyTorch<\/span><\/div>\n
<\/div>\n
\u0645\u062c\u0645\u0648\u0639\u0647 \u062f\u0627\u062f\u0647 MNIST \u0634\u0627\u0645\u0644 \u062a\u0635\u0627\u0648\u06cc\u0631 \u06f2\u06f8\u00d7\u06f2\u06f8 \u067e\u06cc\u06a9\u0633\u0644\u06cc \u0627\u0632 \u0627\u0639\u062f\u0627\u062f \u062f\u0633\u062a\u200c\u0646\u0648\u06cc\u0633 \u06f0 \u062a\u0627 \u06f9 \u0627\u0633\u062a. \u0627\u0645\u0627 \u0628\u0631\u0627\u06cc \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0627\u0632 LeNet-5\u060c \u062a\u0635\u0627\u0648\u06cc\u0631 \u0628\u0627\u06cc\u062f \u0628\u0647 \u06f3\u06f2\u00d7\u06f3\u06f2 \u067e\u06cc\u06a9\u0633\u0644 \u062a\u063a\u06cc\u06cc\u0631 \u0627\u0646\u062f\u0627\u0632\u0647 \u062f\u0627\u062f\u0647 \u0634\u0648\u0646\u062f.<\/span><\/div>\n
<\/div>\n
\u06a9\u062f \u0645\u0631\u0628\u0648\u0637 \u0628\u0647 \u0628\u0627\u0631\u06af\u06cc\u0631\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627:<\/span><\/div>\n
# \u0628\u0627\u0631\u06af\u06cc\u0631\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627\u06cc \u0622\u0645\u0648\u0632\u0634\u06cc<\/span><\/div>\n
training_set = Datasets.MNIST(root=’.\/’, download=True,<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 transform=transforms.Compose([transforms.ToTensor(),<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 transforms.Resize((32, 32))]))<\/span><\/div>\n
<\/div>\n
# \u0628\u0627\u0631\u06af\u06cc\u0631\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627\u06cc \u0627\u0631\u0632\u06cc\u0627\u0628\u06cc<\/span><\/div>\n
validation_set = Datasets.MNIST(root=’.\/’, download=True, train=False,<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 transform=transforms.Compose([transforms.ToTensor(),<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 transforms.Resize((32, 32))]))<\/span><\/div>\n
\u2022 \u0645\u062c\u0645\u0648\u0639\u0647 \u0622\u0645\u0648\u0632\u0634\u06cc \u06f6\u06f0,\u06f0\u06f0\u06f0 \u062a\u0635\u0648\u06cc\u0631 \u062f\u0627\u0631\u062f.<\/span><\/div>\n
\u2022 \u0645\u062c\u0645\u0648\u0639\u0647 \u0627\u0639\u062a\u0628\u0627\u0631\u0633\u0646\u062c\u06cc \u06f1\u06f0,\u06f0\u06f0\u06f0 \u062a\u0635\u0648\u06cc\u0631 \u062f\u0627\u0631\u062f.<\/span><\/div>\n
\u2022 \u062a\u0635\u0627\u0648\u06cc\u0631 \u0628\u0647 \u062a\u0646\u0633\u0648\u0631<\/span> (Tensor)<\/span> \u062a\u0628\u062f\u06cc\u0644 \u0645\u06cc\u200c\u0634\u0648\u0646\u062f \u0648 \u0628\u0647 \u06f3\u06f2\u00d7\u06f3\u06f2 \u062a\u063a\u06cc\u06cc\u0631 \u0627\u0646\u062f\u0627\u0632\u0647 \u062f\u0627\u062f\u0647 \u0645\u06cc\u200c\u0634\u0648\u0646\u062f.<\/span><\/div>\n
<\/div>\n
<\/div>\n
<\/div>\n
\u06f2. \u062a\u0639\u0631\u06cc\u0641 \u06a9\u0644\u0627\u0633 \u0634\u0628\u06a9\u0647 \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc<\/span> (CNN)<\/span><\/div>\n
<\/div>\n
\u062f\u0631 \u0627\u06cc\u0646 \u0628\u062e\u0634\u060c \u06a9\u0644\u0627\u0633 ConvolutionalNeuralNet \u062a\u0639\u0631\u06cc\u0641 \u0634\u062f\u0647 \u0627\u0633\u062a \u06a9\u0647 \u0634\u0627\u0645\u0644 \u0645\u062a\u062f\u0647\u0627\u06cc\u06cc \u0628\u0631\u0627\u06cc \u0622\u0645\u0648\u0632\u0634 \u0645\u062f\u0644\u060c \u0645\u062d\u0627\u0633\u0628\u0647 \u062f\u0642\u062a\u060c \u0648 \u0627\u0646\u062c\u0627\u0645 \u067e\u06cc\u0634\u200c\u0628\u06cc\u0646\u06cc \u0627\u0633\u062a.<\/span><\/div>\n
<\/div>\n
\u0648\u06cc\u0698\u06af\u06cc\u200c\u0647\u0627\u06cc \u06a9\u0644\u0627\u0633:<\/span><\/div>\n
\u2022 __init__(): \u0645\u0642\u062f\u0627\u0631\u062f\u0647\u06cc \u0627\u0648\u0644\u06cc\u0647 \u0645\u062f\u0644 \u0648 \u062a\u0646\u0638\u06cc\u0645 \u0628\u0647\u06cc\u0646\u0647\u200c\u0633\u0627\u0632 Adam.<\/span><\/div>\n
\u2022 train(): \u0634\u0627\u0645\u0644 \u0645\u0631\u0627\u062d\u0644 \u0632\u06cc\u0631 \u0627\u0633\u062a:<\/span><\/div>\n
\u2022 \u0645\u0642\u062f\u0627\u0631\u062f\u0647\u06cc \u0627\u0648\u0644\u06cc\u0647 \u0648\u0632\u0646\u200c\u0647\u0627 \u0628\u0627 Xavier Initialization<\/span><\/div>\n
\u2022 \u0627\u06cc\u062c\u0627\u062f DataLoader \u0628\u0631\u0627\u06cc \u0645\u062c\u0645\u0648\u0639\u0647\u200c\u0647\u0627\u06cc \u0622\u0645\u0648\u0632\u0634\u06cc \u0648 \u0627\u0631\u0632\u06cc\u0627\u0628\u06cc<\/span><\/div>\n
\u2022 \u0622\u0645\u0648\u0632\u0634 \u0645\u062f\u0644 \u0628\u0627 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0627\u0632 \u062a\u0627\u0628\u0639 \u0647\u0632\u06cc\u0646\u0647 \u0648 \u0628\u0647\u200c\u0631\u0648\u0632\u0631\u0633\u0627\u0646\u06cc \u0648\u0632\u0646\u200c\u0647\u0627<\/span><\/div>\n
\u2022 \u0645\u062d\u0627\u0633\u0628\u0647 \u062f\u0642\u062a \u0645\u062f\u0644 \u0631\u0648\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627\u06cc \u0622\u0645\u0648\u0632\u0634\u06cc \u0648 \u0627\u0631\u0632\u06cc\u0627\u0628\u06cc<\/span><\/div>\n
\u2022 predict(): \u062f\u0631\u06cc\u0627\u0641\u062a \u0648\u0631\u0648\u062f\u06cc \u0648 \u062e\u0631\u0648\u062c\u06cc \u0634\u0628\u06a9\u0647 \u0631\u0627 \u0628\u0631\u0645\u06cc\u200c\u06af\u0631\u062f\u0627\u0646\u062f.<\/span><\/div>\n
<\/div>\n
\u06a9\u062f \u0645\u0631\u0628\u0648\u0637 \u0628\u0647 \u0645\u0642\u062f\u0627\u0631\u062f\u0647\u06cc \u0627\u0648\u0644\u06cc\u0647 \u0645\u062f\u0644:<\/span><\/div>\n
class ConvolutionalNeuralNet():<\/span><\/div>\n
\u00a0 \u00a0 def __init__(self, network):<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.network = network.to(device)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.optimizer = torch.optim.Adam(self.network.parameters(), lr=1e-3)<\/span><\/div>\n
\u2022 \u0645\u062f\u0644 \u0631\u0648\u06cc GPU<\/span> (\u0627\u06af\u0631<\/span> \u0645\u0648\u062c\u0648\u062f \u0628\u0627\u0634\u062f) \u06cc\u0627 CPU \u0642\u0631\u0627\u0631 \u0645\u06cc\u200c\u06af\u06cc\u0631\u062f.<\/span><\/div>\n
\u2022 \u0628\u0647\u06cc\u0646\u0647\u200c\u0633\u0627\u0632 Adam \u0628\u0627 \u0646\u0631\u062e \u06cc\u0627\u062f\u06af\u06cc\u0631\u06cc \u06f1e-3 \u062a\u0646\u0638\u06cc\u0645 \u0645\u06cc\u200c\u0634\u0648\u062f.<\/span><\/div>\n
<\/div>\n
<\/div>\n
<\/div>\n
\u06f3. \u062a\u0627\u0628\u0639 \u0622\u0645\u0648\u0632\u0634<\/span> (train())<\/span><\/div>\n
<\/div>\n
\u0627\u06cc\u0646 \u062a\u0627\u0628\u0639 \u0634\u0628\u06a9\u0647 \u0639\u0635\u0628\u06cc \u0631\u0627 \u0622\u0645\u0648\u0632\u0634 \u062f\u0627\u062f\u0647 \u0648 \u062f\u0642\u062a \u0631\u0627 \u0645\u062d\u0627\u0633\u0628\u0647 \u0645\u06cc\u200c\u06a9\u0646\u062f.<\/span><\/div>\n
<\/div>\n
\u0645\u0631\u0627\u062d\u0644:<\/span><\/div>\n
1. \u0627\u06cc\u062c\u0627\u062f \u062a\u0627\u0628\u0639 \u0645\u0642\u062f\u0627\u0631\u062f\u0647\u06cc \u0627\u0648\u0644\u06cc\u0647 \u0648\u0632\u0646\u200c\u0647\u0627:<\/span><\/div>\n
def init_weights(module):<\/span><\/div>\n
\u00a0 \u00a0 if isinstance(module, nn.Conv2d):<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 torch.nn.init.xavier_uniform_(module.weight)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 module.bias.data.fill_(0.01)<\/span><\/div>\n
\u00a0 \u00a0 elif isinstance(module, nn.Linear):<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 torch.nn.init.xavier_uniform_(module.weight)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 module.bias.data.fill_(0.01)<\/span><\/div>\n
\u2022 \u0648\u0632\u0646\u200c\u0647\u0627\u06cc \u0644\u0627\u06cc\u0647\u200c\u0647\u0627\u06cc Conv2D \u0648 Linear \u0628\u0627 Xavier Initialization \u0645\u0642\u062f\u0627\u0631\u062f\u0647\u06cc \u0645\u06cc\u200c\u0634\u0648\u0646\u062f.<\/span><\/div>\n
<\/div>\n
2. \u0627\u06cc\u062c\u0627\u062f \u062a\u0627\u0628\u0639 \u0645\u062d\u0627\u0633\u0628\u0647 \u062f\u0642\u062a \u0645\u062f\u0644:<\/span><\/div>\n
def accuracy(network, dataloader):<\/span><\/div>\n
\u00a0 \u00a0 network.eval()<\/span><\/div>\n
\u00a0 \u00a0 total_correct = 0<\/span><\/div>\n
\u00a0 \u00a0 total_instances = 0<\/span><\/div>\n
\u00a0 \u00a0 for images, labels in tqdm(dataloader):<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 images, labels = images.to(device), labels.to(device)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 predictions = torch.argmax(network(images), dim=1)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 correct_predictions = sum(predictions == labels).item()<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 total_correct += correct_predictions<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 total_instances += len(images)<\/span><\/div>\n
\u00a0 \u00a0 return round(total_correct \/ total_instances, 3)<\/span><\/div>\n
\u2022 \u0645\u062f\u0644 \u0628\u0647 \u062d\u0627\u0644\u062a \u0627\u0631\u0632\u06cc\u0627\u0628\u06cc \u062a\u063a\u06cc\u06cc\u0631 \u0645\u06cc\u200c\u06a9\u0646\u062f.<\/span><\/div>\n
\u2022 \u062a\u0639\u062f\u0627\u062f \u067e\u06cc\u0634\u200c\u0628\u06cc\u0646\u06cc\u200c\u0647\u0627\u06cc \u0635\u062d\u06cc\u062d \u0645\u062d\u0627\u0633\u0628\u0647 \u0648 \u0630\u062e\u06cc\u0631\u0647 \u0645\u06cc\u200c\u0634\u0648\u062f.<\/span><\/div>\n
<\/div>\n
3. \u0622\u0645\u0648\u0632\u0634 \u0645\u062f\u0644 \u0628\u0631\u0627\u06cc epochs \u0645\u0634\u062e\u0635 \u0634\u062f\u0647:<\/span><\/div>\n
for epoch in range(epochs):<\/span><\/div>\n
\u00a0 \u00a0 print(f’Epoch {epoch+1}\/{epochs}’)<\/span><\/div>\n
\u00a0 \u00a0 train_losses =<\/span> []<\/span><\/div>\n
\u00a0 \u00a0\u00a0<\/span><\/div>\n
\u00a0 \u00a0 # \u0645\u0631\u062d\u0644\u0647 \u0622\u0645\u0648\u0632\u0634<\/span><\/div>\n
\u00a0 \u00a0 print(‘training…’)<\/span><\/div>\n
\u00a0 \u00a0 for images, labels in tqdm(train_loader):<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 images, labels = images.to(device), labels.to(device)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.optimizer.zero_grad()<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 predictions = self.network(images)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 loss = loss_function(predictions, labels)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 loss.backward()<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.optimizer.step()<\/span><\/div>\n
\u2022 \u0645\u062f\u0644 \u0631\u0648\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627\u06cc \u0622\u0645\u0648\u0632\u0634\u06cc \u0627\u062c\u0631\u0627 \u0645\u06cc\u200c\u0634\u0648\u062f.<\/span><\/div>\n
\u2022 \u062a\u0627\u0628\u0639 \u0647\u0632\u06cc\u0646\u0647 \u0645\u062d\u0627\u0633\u0628\u0647 \u0648 \u0645\u0642\u062f\u0627\u0631 \u0622\u0646 \u0630\u062e\u06cc\u0631\u0647 \u0645\u06cc\u200c\u0634\u0648\u062f.<\/span><\/div>\n
\u2022 \u06af\u0631\u0627\u062f\u06cc\u0627\u0646 \u0645\u062d\u0627\u0633\u0628\u0647 \u0648 \u0648\u0632\u0646\u200c\u0647\u0627 \u0628\u0647\u200c\u0631\u0648\u0632\u0631\u0633\u0627\u0646\u06cc \u0645\u06cc\u200c\u0634\u0648\u0646\u062f.<\/span><\/div>\n
<\/div>\n
4. \u0645\u062d\u0627\u0633\u0628\u0647 \u062f\u0642\u062a \u0645\u062f\u0644 \u0631\u0648\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627\u06cc \u0627\u0631\u0632\u06cc\u0627\u0628\u06cc:<\/span><\/div>\n
with torch.no_grad():<\/span><\/div>\n
\u00a0 \u00a0 print(‘deriving validation accuracy…’)<\/span><\/div>\n
\u00a0 \u00a0 val_accuracy = accuracy(self.network, val_loader)<\/span><\/div>\n
\u00a0 \u00a0 log_dict[‘validation_accuracy_per_epoch’].append(val_accuracy)<\/span><\/div>\n
\u2022 \u0645\u062f\u0644 \u0631\u0648\u06cc \u062f\u0627\u062f\u0647\u200c\u0647\u0627\u06cc Validation \u062a\u0633\u062a \u0645\u06cc\u200c\u0634\u0648\u062f \u0648 \u062f\u0642\u062a \u0622\u0646 \u0645\u062d\u0627\u0633\u0628\u0647 \u0645\u06cc\u200c\u0634\u0648\u062f.<\/span><\/div>\n
<\/div>\n
5. \u0646\u0645\u0627\u06cc\u0634 \u062e\u0631\u0648\u062c\u06cc \u0647\u0631 \u062f\u0648\u0631\u0647<\/span> (epoch):<\/span><\/div>\n
print(f’training_loss: {round(train_losses, 4)} \u00a0training_accuracy:<\/span> ‘+<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 f'{train_accuracy} \u00a0validation_loss: {round(val_losses, 4)}<\/span> ‘+<\/span> \u00a0<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 f’validation_accuracy: {val_accuracy}\\n’)<\/span><\/div>\n
<\/div>\n
\u06f4. \u067e\u06cc\u0634\u200c\u0628\u06cc\u0646\u06cc \u0628\u0627 \u0645\u062f\u0644<\/span> (predict())<\/span><\/div>\n
def predict(self, x):<\/span><\/div>\n
\u00a0 \u00a0 return self.network(x)<\/span><\/div>\n
\u2022 \u0628\u0631\u0627\u06cc \u067e\u06cc\u0634\u200c\u0628\u06cc\u0646\u06cc \u0639\u062f\u062f \u062f\u0633\u062a\u200c\u0646\u0648\u06cc\u0633 \u0627\u0632 \u0645\u062f\u0644 \u0622\u0645\u0648\u0632\u0634\u200c\u062f\u06cc\u062f\u0647 \u0634\u062f\u0647 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0645\u06cc\u200c\u0634\u0648\u062f.<\/span><\/div>\n
<\/div>\n
\u0646\u062a\u06cc\u062c\u0647\u200c\u06af\u06cc\u0631\u06cc<\/span><\/div>\n
<\/div>\n
\u0627\u06cc\u0646 \u06a9\u062f \u0628\u0631\u0627\u06cc \u0622\u0645\u0648\u0632\u0634 \u0645\u062f\u0644 LeNet-5 \u0631\u0648\u06cc \u0645\u062c\u0645\u0648\u0639\u0647 \u062f\u0627\u062f\u0647 MNIST \u0637\u0631\u0627\u062d\u06cc \u0634\u062f\u0647 \u0627\u0633\u062a.<\/span><\/div>\n
\u2022 \u0627\u0632 CNN \u0628\u0631\u0627\u06cc \u062a\u0634\u062e\u06cc\u0635 \u0627\u0639\u062f\u0627\u062f \u062f\u0633\u062a\u200c\u0646\u0648\u06cc\u0633 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0645\u06cc\u200c\u0634\u0648\u062f.<\/span><\/div>\n
\u2022 \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0628\u0647 \u06f3\u06f2\u00d7\u06f3\u06f2 \u062a\u063a\u06cc\u06cc\u0631 \u0627\u0646\u062f\u0627\u0632\u0647 \u062f\u0627\u062f\u0647 \u0634\u062f\u0647\u200c\u0627\u0646\u062f.<\/span><\/div>\n
\u2022 \u0627\u0632 \u0628\u0647\u06cc\u0646\u0647\u200c\u0633\u0627\u0632 Adam \u0628\u0631\u0627\u06cc \u0628\u0647\u200c\u0631\u0648\u0632\u0631\u0633\u0627\u0646\u06cc \u0648\u0632\u0646\u200c\u0647\u0627 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0634\u062f\u0647 \u0627\u0633\u062a.<\/span><\/div>\n
\u2022 \u0645\u062f\u0644 \u0645\u06cc\u200c\u062a\u0648\u0627\u0646\u062f \u0627\u0639\u062f\u0627\u062f \u062f\u0633\u062a\u200c\u0646\u0648\u06cc\u0633 \u062c\u062f\u06cc\u062f \u0631\u0627 \u067e\u06cc\u0634\u200c\u0628\u06cc\u0646\u06cc \u06a9\u0646\u062f.<\/span><\/div>\n
<\/div>\n
LeNet-5<\/span><\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
LeNet-5<\/span> (Y.<\/span> Lecun et al) \u06cc\u06a9\u06cc \u0627\u0632 \u0627\u0648\u0644\u06cc\u0646 \u0634\u0628\u06a9\u0647\u200c\u0647\u0627\u06cc \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u0627\u0633\u062a \u06a9\u0647 \u0628\u0647\u200c\u0637\u0648\u0631 \u062e\u0627\u0635 \u0628\u0631\u0627\u06cc \u062a\u0634\u062e\u06cc\u0635 \u0648 \u0637\u0628\u0642\u0647\u200c\u0628\u0646\u062f\u06cc \u062a\u0635\u0627\u0648\u06cc\u0631 \u0627\u0631\u0642\u0627\u0645 \u062f\u0633\u062a\u200c\u0646\u0648\u06cc\u0633 \u0637\u0631\u0627\u062d\u06cc \u0634\u062f\u0647 \u0627\u0633\u062a. \u0645\u0639\u0645\u0627\u0631\u06cc \u0622\u0646 \u062f\u0631 \u062a\u0635\u0648\u06cc\u0631 \u0628\u0627\u0644\u0627 \u0646\u0645\u0627\u06cc\u0634 \u062f\u0627\u062f\u0647 \u0634\u062f\u0647 \u0648 \u067e\u06cc\u0627\u062f\u0647\u200c\u0633\u0627\u0632\u06cc \u0622\u0646 \u062f\u0631 PyTorch \u062f\u0631 \u06a9\u062f \u0632\u06cc\u0631 \u0627\u0631\u0627\u0626\u0647 \u0634\u062f\u0647 \u0627\u0633\u062a.<\/span><\/div>\n
<\/div>\n
class LeNet5(nn.Module):<\/span><\/div>\n
\u00a0 def __init__(self):<\/span><\/div>\n
\u00a0 \u00a0 super().__init__()<\/span><\/div>\n
\u00a0 \u00a0 self.conv1 = nn.Conv2d(1, 6, 5)<\/span><\/div>\n
\u00a0 \u00a0 self.pool1 = nn.AvgPool2d(2)<\/span><\/div>\n
\u00a0 \u00a0 self.conv2 = nn.Conv2d(6, 16, 5)<\/span><\/div>\n
\u00a0 \u00a0 self.pool2 = nn.AvgPool2d(2)<\/span><\/div>\n
\u00a0 \u00a0 self.linear1 = nn.Linear(5*5*16, 120)<\/span><\/div>\n
\u00a0 \u00a0 self.linear2 = nn.Linear(120, 84)<\/span><\/div>\n
\u00a0 \u00a0 self.linear3 = nn.Linear(84, 10)<\/span><\/div>\n
<\/div>\n
\u00a0 def forward(self, x):<\/span><\/div>\n
\u00a0 \u00a0 x = x.view(-1, 1, 32, 32)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 # \u0644\u0627\u06cc\u0647 \u06f1<\/span><\/div>\n
\u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 output_1 = self.conv1(x)<\/span><\/div>\n
\u00a0 \u00a0 output_1 = torch.tanh(output_1)<\/span><\/div>\n
\u00a0 \u00a0 output_1 = self.pool1(output_1)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 # \u0644\u0627\u06cc\u0647 \u06f2<\/span><\/div>\n
\u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 output_2 = self.conv2(output_1)<\/span><\/div>\n
\u00a0 \u00a0 output_2 = torch.tanh(output_2)<\/span><\/div>\n
\u00a0 \u00a0 output_2 = self.pool2(output_2)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 # \u0645\u0633\u0637\u062d\u200c\u0633\u0627\u0632\u06cc<\/span><\/div>\n
\u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 output_2 = output_2.view(-1, 5*5*16)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 # \u0644\u0627\u06cc\u0647 \u06f3<\/span><\/div>\n
\u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 output_3 = self.linear1(output_2)<\/span><\/div>\n
\u00a0 \u00a0 output_3 = torch.tanh(output_3)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 # \u0644\u0627\u06cc\u0647 \u06f4<\/span><\/div>\n
\u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 output_4 = self.linear2(output_3)<\/span><\/div>\n
\u00a0 \u00a0 output_4 = torch.tanh(output_4)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 #————-<\/span><\/div>\n
\u00a0 \u00a0 # \u0644\u0627\u06cc\u0647 \u062e\u0631\u0648\u062c\u06cc<\/span><\/div>\n
\u00a0 \u00a0 #————-<\/span><\/div>\n
\u00a0 \u00a0 output_5 = self.linear3(output_4)<\/span><\/div>\n
\u00a0 \u00a0 return(F.softmax(output_5, dim=1))<\/span><\/div>\n
\u0628\u0627 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0627\u0632 \u0645\u0639\u0645\u0627\u0631\u06cc LeNet-5 \u062a\u0639\u0631\u06cc\u0641\u200c\u0634\u062f\u0647 \u062f\u0631 \u0628\u0627\u0644\u0627\u060c \u0645\u062f\u0644 model_1 \u0631\u0627 \u06a9\u0647 \u0639\u0636\u0648\u06cc \u0627\u0632 \u06a9\u0644\u0627\u0633 ConvolutionalNeuralNet \u0627\u0633\u062a\u060c \u0628\u0627 \u067e\u0627\u0631\u0627\u0645\u062a\u0631\u0647\u0627\u06cc \u0645\u0634\u062e\u0635\u200c\u0634\u062f\u0647 \u062f\u0631 \u06a9\u062f \u0632\u06cc\u0631 \u0645\u0642\u062f\u0627\u0631\u062f\u0647\u06cc \u0645\u06cc\u200c\u06a9\u0646\u06cc\u0645. \u0627\u06cc\u0646 \u0645\u062f\u0644 \u0628\u0647\u200c\u0639\u0646\u0648\u0627\u0646 \u06cc\u06a9 \u062e\u0637 \u067e\u0627\u06cc\u0647 \u0628\u0631\u0627\u06cc \u0627\u0647\u062f\u0627\u0641 \u0627\u0631\u0632\u06cc\u0627\u0628\u06cc \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u062e\u0648\u0627\u0647\u062f \u0634\u062f.<\/span><\/div>\n
# \u0622\u0645\u0648\u0632\u0634 \u0645\u062f\u0644 \u06f1<\/span><\/div>\n
model_1 = ConvolutionalNeuralNet(LeNet5())<\/span><\/div>\n
<\/div>\n
log_dict_1 = model_1.train(nn.CrossEntropyLoss(), epochs=10, batch_size=64,\u00a0<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0training_set=training_set, validation_set=validation_set)<\/span><\/div>\n
\u067e\u0633 \u0627\u0632 \u0622\u0645\u0648\u0632\u0634 \u0628\u0647 \u0645\u062f\u062a \u06f1\u06f0 \u062f\u0648\u0631\u0647 \u0648 \u0645\u0634\u0627\u0647\u062f\u0647 \u062f\u0642\u062a\u200c\u0647\u0627 \u0627\u0632 \u0637\u0631\u06cc\u0642 \u0644\u0627\u06af \u0645\u062a\u0631\u06cc\u06a9\u200c\u0647\u0627\u060c \u0645\u06cc\u200c\u0628\u06cc\u0646\u06cc\u0645 \u06a9\u0647 \u0647\u0631 \u062f\u0648 \u062f\u0642\u062a \u0622\u0645\u0648\u0632\u0634 \u0648 \u0627\u0639\u062a\u0628\u0627\u0631\u0633\u0646\u062c\u06cc \u062f\u0631 \u0637\u0648\u0644 \u0641\u0631\u0622\u06cc\u0646\u062f \u0622\u0645\u0648\u0632\u0634 \u0627\u0641\u0632\u0627\u06cc\u0634 \u06cc\u0627\u0641\u062a\u0647\u200c\u0627\u0646\u062f. \u062f\u0631 \u0627\u06cc\u0646 \u0622\u0632\u0645\u0627\u06cc\u0634\u060c \u062f\u0642\u062a \u0627\u0639\u062a\u0628\u0627\u0631\u0633\u0646\u062c\u06cc \u062f\u0631 \u0627\u0628\u062a\u062f\u0627 \u062a\u0642\u0631\u06cc\u0628\u0627\u064b \u06f9\u06f3\u066a \u062f\u0631 \u0627\u0648\u0644\u06cc\u0646 \u062f\u0648\u0631\u0647 \u0628\u0648\u062f \u0648 \u0633\u067e\u0633 \u0628\u0647\u200c\u0637\u0648\u0631 \u067e\u06cc\u0648\u0633\u062a\u0647 \u0637\u06cc \u06f9 \u062a\u06a9\u0631\u0627\u0631 \u0628\u0639\u062f\u06cc \u0627\u0641\u0632\u0627\u06cc\u0634 \u06cc\u0627\u0641\u062a \u0648 \u062f\u0631 \u0646\u0647\u0627\u06cc\u062a \u062f\u0631 \u062f\u0648\u0631\u0647 \u062f\u0647\u0645 \u0628\u0647 \u0628\u06cc\u0634 \u0627\u0632 \u06f9\u06f8\u066a \u0631\u0633\u06cc\u062f.<\/span><\/div>\n
sns.lineplot(y=log_dict_1[‘training_accuracy_per_epoch’], x=range(len(log_dict_1[‘training_accuracy_per_epoch’])), label=’training’)<\/span><\/div>\n
<\/div>\n
sns.lineplot(y=log_dict_1[‘validation_accuracy_per_epoch’], x=range(len(log_dict_1[‘validation_accuracy_per_epoch’])), label=’validation’)<\/span><\/div>\n
<\/div>\n
plt.xlabel(‘epoch’)<\/span><\/div>\n
plt.ylabel(‘accuracy’)<\/span><\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
LeNet-5 \u0628\u0627 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc<\/span> (Batch<\/span> Normalized LeNet-5)<\/span><\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
\u0627\u0632 \u0622\u0646\u062c\u0627\u06cc\u06cc \u06a9\u0647 \u0645\u0648\u0636\u0648\u0639 \u0627\u06cc\u0646 \u0645\u0642\u0627\u0644\u0647 \u062d\u0648\u0644 \u0645\u062d\u0648\u0631 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u062f\u0631 \u0644\u0627\u06cc\u0647\u200c\u0647\u0627\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u0627\u0633\u062a\u060c \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0641\u0642\u0637 \u062f\u0631 \u062f\u0648 \u0644\u0627\u06cc\u0647\u200c\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u0645\u0648\u062c\u0648\u062f \u062f\u0631 \u0627\u06cc\u0646 \u0645\u0639\u0645\u0627\u0631\u06cc \u0627\u0639\u0645\u0627\u0644 \u0634\u062f\u0647 \u0627\u0633\u062a\u060c \u0647\u0645\u0627\u0646\u200c\u0637\u0648\u0631 \u06a9\u0647 \u062f\u0631 \u062a\u0635\u0648\u06cc\u0631 \u0628\u0627\u0644\u0627 \u0646\u0634\u0627\u0646 \u062f\u0627\u062f\u0647 \u0634\u062f\u0647 \u0627\u0633\u062a.<\/span><\/div>\n
class LeNet5_BatchNorm(nn.Module):<\/span><\/div>\n
\u00a0 \u00a0 def __init__(self):<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 super().__init__()<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.conv1 = nn.Conv2d(1, 6, 5)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.batchnorm1 = nn.BatchNorm2d(6)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.pool1 = nn.AvgPool2d(2)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.conv2 = nn.Conv2d(6, 16, 5)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.batchnorm2 = nn.BatchNorm2d(16)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.pool2 = nn.AvgPool2d(2)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.linear1 = nn.Linear(5*5*16, 120)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.linear2 = nn.Linear(120, 84)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.linear3 = nn.Linear(84, 10)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 def forward(self, x):<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 x = x.view(-1, 1, 32, 32)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 # LAYER 1<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_1 = self.conv1(x)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_1 = torch.tanh(output_1)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_1 = self.batchnorm1(output_1)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_1 = self.pool1(output_1)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 # LAYER 2<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_2 = self.conv2(output_1)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_2 = torch.tanh(output_2)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_2 = self.batchnorm2(output_2)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_2 = self.pool2(output_2)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 # FLATTEN<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_2 = output_2.view(-1, 5*5*16)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 # LAYER 3<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_3 = self.linear1(output_2)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_3 = torch.tanh(output_3)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 # LAYER 4<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #———-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_4 = self.linear2(output_3)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_4 = torch.tanh(output_4)<\/span><\/div>\n
<\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #————-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 # OUTPUT LAYER<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 #————-<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 output_5 = self.linear3(output_4)<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 return F.softmax(output_5, dim=1)<\/span><\/div>\n
\u0622\u0645\u0648\u0632\u0634 \u0645\u062f\u0644 \u062f\u0627\u0631\u0627\u06cc Batch Normalization<\/span><\/div>\n
<\/div>\n
\u0628\u0627 \u0627\u0633\u062a\u0641\u0627\u062f\u0647 \u0627\u0632 \u0642\u0637\u0639\u0647 \u06a9\u062f \u0632\u06cc\u0631\u060c \u0645\u062f\u0644 model_2 \u0631\u0627 \u06a9\u0647 \u0634\u0627\u0645\u0644 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0627\u0633\u062a \u0645\u0642\u062f\u0627\u0631\u062f\u0647\u06cc \u0627\u0648\u0644\u06cc\u0647 \u0645\u06cc\u200c\u06a9\u0646\u06cc\u0645 \u0648 \u0622\u0645\u0648\u0632\u0634 \u0622\u0646 \u0631\u0627 \u0628\u0627 \u0647\u0645\u0627\u0646 \u067e\u0627\u0631\u0627\u0645\u062a\u0631\u0647\u0627\u06cc model_1 \u0622\u063a\u0627\u0632 \u0645\u06cc\u200c\u06a9\u0646\u06cc\u0645. \u0633\u067e\u0633\u060c \u0627\u0645\u062a\u06cc\u0627\u0632\u0627\u062a \u062f\u0642\u062a \u0631\u0627 \u0627\u0631\u0632\u06cc\u0627\u0628\u06cc \u0645\u06cc\u200c\u06a9\u0646\u06cc\u0645.<\/span><\/div>\n
# \u00a0\u0622\u0645\u0648\u0632\u0634 \u0645\u062f\u0644 \u06f2<\/span><\/div>\n
model_2 = ConvolutionalNeuralNet(LeNet5_BatchNorm())<\/span><\/div>\n
<\/div>\n
log_dict_2 = model_2.train(nn.CrossEntropyLoss(), epochs=10, batch_size=64,\u00a0<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0training_set=training_set, validation_set=validation_set)<\/span><\/div>\n
\u0646\u062a\u0627\u06cc\u062c \u0648 \u062a\u062d\u0644\u06cc\u0644<\/span><\/div>\n
<\/div>\n
\u0628\u0627 \u0645\u0634\u0627\u0647\u062f\u0647\u200c\u06cc \u0646\u0645\u0648\u062f\u0627\u0631\u060c \u0645\u0634\u062e\u0635 \u0627\u0633\u062a \u06a9\u0647 \u0647\u0631 \u062f\u0648 \u062f\u0642\u062a \u0622\u0645\u0648\u0632\u0634\u06cc \u0648 \u0627\u0639\u062a\u0628\u0627\u0631\u0633\u0646\u062c\u06cc \u062f\u0631 \u0637\u0648\u0644 \u0622\u0645\u0648\u0632\u0634 \u0645\u0634\u0627\u0628\u0647 \u0645\u062f\u0644 \u0628\u062f\u0648\u0646 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0627\u0641\u0632\u0627\u06cc\u0634 \u06cc\u0627\u0641\u062a\u0647\u200c\u0627\u0646\u062f.<\/span><\/div>\n
\u062f\u0642\u062a \u0627\u0639\u062a\u0628\u0627\u0631\u0633\u0646\u062c\u06cc \u067e\u0633 \u0627\u0632 \u0627\u0648\u0644\u06cc\u0646 \u0627\u06cc\u067e\u0627\u06a9 \u0628\u06cc\u0634 \u0627\u0632 95\u066a \u0628\u0648\u062f\u060c \u06cc\u0639\u0646\u06cc \u06f3\u066a \u0628\u06cc\u0634\u062a\u0631 \u0627\u0632 model_1 \u062f\u0631 \u0647\u0645\u0627\u0646 \u0646\u0642\u0637\u0647. \u0633\u067e\u0633\u060c \u0627\u06cc\u0646 \u0645\u0642\u062f\u0627\u0631 \u0628\u0647 \u062a\u062f\u0631\u06cc\u062c \u0627\u0641\u0632\u0627\u06cc\u0634 \u06cc\u0627\u0641\u062a \u0648 \u062f\u0631 \u0646\u0647\u0627\u06cc\u062a \u062d\u062f\u0648\u062f 98.5\u066a \u0634\u062f \u06a9\u0647 0.5\u066a \u0628\u0627\u0644\u0627\u062a\u0631 \u0627\u0632 model_1 \u0627\u0633\u062a.<\/span><\/div>\n
sns.lineplot(y=log_dict_2[‘training_accuracy_per_epoch’],\u00a0<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0x=range(len(log_dict_2[‘training_accuracy_per_epoch’])), label=’training’)<\/span><\/div>\n
<\/div>\n
sns.lineplot(y=log_dict_2[‘validation_accuracy_per_epoch’],\u00a0<\/span><\/div>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0x=range(len(log_dict_2[‘validation_accuracy_per_epoch’])), label=’validation’)<\/span><\/div>\n
<\/div>\n
plt.xlabel(‘epoch’)<\/span><\/div>\n
plt.ylabel(‘accuracy’)<\/span><\/div>\n
<\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
\u0645\u0642\u0627\u06cc\u0633\u0647 \u0645\u062f\u0644\u200c\u0647\u0627<\/span><\/div>\n
<\/div>\n
\u0628\u0627 \u0645\u0642\u0627\u06cc\u0633\u0647\u200c\u06cc \u0647\u0631 \u062f\u0648 \u0645\u062f\u0644\u060c \u0645\u0634\u062e\u0635 \u0627\u0633\u062a \u06a9\u0647 \u0645\u062f\u0644 LeNet-5 \u0628\u0627 \u0644\u0627\u06cc\u0647\u200c\u0647\u0627\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u062f\u0627\u0631\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0639\u0645\u0644\u06a9\u0631\u062f \u0628\u0647\u062a\u0631\u06cc \u0646\u0633\u0628\u062a \u0628\u0647 \u0645\u062f\u0644 \u0645\u0639\u0645\u0648\u0644\u06cc \u0628\u062f\u0648\u0646 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u062f\u0627\u0634\u062a\u0647 \u0627\u0633\u062a. \u0628\u0646\u0627\u0628\u0631\u0627\u06cc\u0646\u060c \u0645\u06cc\u200c\u062a\u0648\u0627\u0646 \u0646\u062a\u06cc\u062c\u0647 \u06af\u0631\u0641\u062a \u06a9\u0647 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u062f\u0631 \u0627\u06cc\u0646 \u0645\u0648\u0631\u062f \u0628\u0647 \u0628\u0647\u0628\u0648\u062f \u0639\u0645\u0644\u06a9\u0631\u062f \u06a9\u0645\u06a9 \u06a9\u0631\u062f\u0647 \u0627\u0633\u062a.<\/span><\/div>\n
<\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
\u0645\u0642\u0627\u06cc\u0633\u0647\u200c\u06cc \u0636\u0631\u0627\u06cc\u0628 \u062e\u0637\u0627<\/span> (Loss)<\/span> \u062f\u0631 \u0622\u0645\u0648\u0632\u0634 \u0648 \u0627\u0639\u062a\u0628\u0627\u0631\u0633\u0646\u062c\u06cc \u0628\u06cc\u0646 \u0645\u062f\u0644 LeNet-5 \u0645\u0639\u0645\u0648\u0644\u06cc \u0648 \u0645\u062f\u0644 \u062f\u0627\u0631\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0646\u0634\u0627\u0646 \u0645\u06cc\u200c\u062f\u0647\u062f \u06a9\u0647 \u0645\u062f\u0644 \u062f\u0627\u0631\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0633\u0631\u06cc\u0639\u200c\u062a\u0631 \u0627\u0632 \u0645\u062f\u0644 \u0645\u0639\u0645\u0648\u0644\u06cc \u0628\u0647 \u0645\u0642\u0627\u062f\u06cc\u0631 \u062e\u0637\u0627\u06cc \u067e\u0627\u06cc\u06cc\u0646\u200c\u062a\u0631 \u062f\u0633\u062a \u067e\u06cc\u062f\u0627 \u0645\u06cc\u200c\u06a9\u0646\u062f. \u0627\u06cc\u0646 \u0645\u0648\u0636\u0648\u0639 \u0646\u0634\u0627\u0646\u200c\u062f\u0647\u0646\u062f\u0647\u200c\u06cc \u0627\u06cc\u0646 \u0627\u0633\u062a \u06a9\u0647 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0633\u0631\u0639\u062a \u0628\u0647\u06cc\u0646\u0647\u200c\u0633\u0627\u0632\u06cc \u0648\u0632\u0646\u200c\u0647\u0627\u06cc \u0645\u062f\u0644 \u0631\u0627 \u062f\u0631 \u062c\u0647\u062a \u0635\u062d\u06cc\u062d \u0627\u0641\u0632\u0627\u06cc\u0634 \u0645\u06cc\u200c\u062f\u0647\u062f\u060c \u06cc\u0627 \u0628\u0647 \u0639\u0628\u0627\u0631\u062a \u062f\u06cc\u06af\u0631\u060c \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0633\u0631\u0639\u062a \u06cc\u0627\u062f\u06af\u06cc\u0631\u06cc \u0634\u0628\u06a9\u0647\u200c\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u0631\u0627 \u0627\u0641\u0632\u0627\u06cc\u0634 \u0645\u06cc\u200c\u062f\u0647\u062f.<\/span><\/div>\n
<\/div>\n
\n
<\/figcaption><\/figure>\n<\/div>\n
\u0636\u0631\u0627\u06cc\u0628 \u062e\u0637\u0627 \u062f\u0631 \u0622\u0645\u0648\u0632\u0634 \u0648 \u0627\u0639\u062a\u0628\u0627\u0631\u0633\u0646\u062c\u06cc<\/b><\/span><\/span><\/div>\n
<\/div>\n
\u0646\u062a\u06cc\u062c\u0647\u200c\u06af\u06cc\u0631\u06cc \u0646\u0647\u0627\u06cc\u06cc<\/span><\/div>\n
<\/div>\n
\u062f\u0631 \u0627\u06cc\u0646 \u0645\u0642\u0627\u0644\u0647\u060c \u0645\u0641\u0647\u0648\u0645 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0631\u0627 \u062f\u0631 \u0632\u0645\u06cc\u0646\u0647\u200c\u06cc \u06cc\u0627\u062f\u06af\u06cc\u0631\u06cc \u0645\u0627\u0634\u06cc\u0646 \u0648 \u06cc\u0627\u062f\u06af\u06cc\u0631\u06cc \u0639\u0645\u06cc\u0642 \u0628\u0631\u0631\u0633\u06cc \u06a9\u0631\u062f\u06cc\u0645. \u0647\u0645\u0686\u0646\u06cc\u0646\u060c \u0641\u0631\u0622\u06cc\u0646\u062f\u0647\u0627\u06cc \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0631\u0627 \u0628\u0647\u200c\u0639\u0646\u0648\u0627\u0646 \u06af\u0627\u0645\u200c\u0647\u0627\u06cc\u06cc \u062f\u0631 \u067e\u06cc\u0634\u200c\u067e\u0631\u062f\u0627\u0632\u0634 \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0645\u0648\u0631\u062f \u0645\u0637\u0627\u0644\u0639\u0647 \u0642\u0631\u0627\u0631 \u062f\u0627\u062f\u06cc\u0645 \u0648 \u062f\u06cc\u062f\u06cc\u0645 \u06a9\u0647 \u0686\u06af\u0648\u0646\u0647 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u0631\u0627 \u0645\u06cc\u200c\u062a\u0648\u0627\u0646 \u0641\u0631\u0627\u062a\u0631 \u0627\u0632 \u067e\u06cc\u0634\u200c\u067e\u0631\u062f\u0627\u0632\u0634 \u062f\u0627\u062f\u0647\u200c\u0647\u0627 \u0648 \u062f\u0631\u0648\u0646 \u0644\u0627\u06cc\u0647\u200c\u0647\u0627\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u0627\u0632 \u0637\u0631\u06cc\u0642 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc<\/span> (Batch<\/span> Normalization) \u0628\u0647\u200c\u06a9\u0627\u0631 \u06af\u0631\u0641\u062a.<\/span><\/div>\n
<\/div>\n
\u0633\u067e\u0633\u060c \u0641\u0631\u0622\u06cc\u0646\u062f \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0631\u0627 \u0628\u0631\u0631\u0633\u06cc \u06a9\u0631\u062f\u0647 \u0648 \u062a\u0623\u062b\u06cc\u0631 \u0622\u0646 \u0631\u0627 \u0628\u0627 \u0645\u0642\u0627\u06cc\u0633\u0647\u200c\u06cc \u062f\u0648 \u0646\u0633\u062e\u0647\u200c\u06cc \u0645\u062e\u062a\u0644\u0641 \u0627\u0632 \u0645\u062f\u0644 LeNet-5<\/span> (\u06cc\u06a9\u06cc<\/span> \u0628\u062f\u0648\u0646 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0648 \u062f\u06cc\u06af\u0631\u06cc \u0628\u0627 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc) \u0628\u0631 \u0631\u0648\u06cc \u0645\u062c\u0645\u0648\u0639\u0647 \u062f\u0627\u062f\u0647\u200c\u06cc MNIST \u0627\u0631\u0632\u06cc\u0627\u0628\u06cc \u06a9\u0631\u062f\u06cc\u0645. \u0646\u062a\u0627\u06cc\u062c \u0646\u0634\u0627\u0646 \u062f\u0627\u062f \u06a9\u0647 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0628\u0627\u0639\u062b \u0627\u0641\u0632\u0627\u06cc\u0634 \u0639\u0645\u0644\u06a9\u0631\u062f \u0645\u062f\u0644 \u0648 \u0633\u0631\u0639\u062a \u0628\u0647\u06cc\u0646\u0647\u200c\u0633\u0627\u0632\u06cc \u0648\u0632\u0646\u200c\u0647\u0627 \u0645\u06cc\u200c\u0634\u0648\u062f.<\/span><\/div>\n
<\/div>\n
\u0647\u0645\u0686\u0646\u06cc\u0646\u060c \u0628\u0631\u062e\u06cc \u067e\u06cc\u0634\u0646\u0647\u0627\u062f \u06a9\u0631\u062f\u0647\u200c\u0627\u0646\u062f \u06a9\u0647 \u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0627\u0632 \u062a\u063a\u06cc\u06cc\u0631\u0627\u062a \u062f\u0627\u062e\u0644\u06cc \u062a\u0648\u0632\u06cc\u0639 \u0648\u06cc\u0698\u06af\u06cc\u200c\u0647\u0627<\/span> (Internal<\/span> Covariate Shift) \u062c\u0644\u0648\u06af\u06cc\u0631\u06cc \u0645\u06cc\u200c\u06a9\u0646\u062f\u060c \u0627\u0645\u0627 \u0647\u0646\u0648\u0632 \u0627\u062c\u0645\u0627\u0639 \u06a9\u0627\u0645\u0644\u06cc \u062f\u0631 \u0627\u06cc\u0646 \u0632\u0645\u06cc\u0646\u0647 \u0648\u062c\u0648\u062f \u0646\u062f\u0627\u0631\u062f.<\/span><\/div>\n
<\/div>","protected":false},"excerpt":{"rendered":"

\u0646\u0631\u0645\u0627\u0644\u200c\u0633\u0627\u0632\u06cc \u062f\u0633\u062a\u0647\u200c\u0627\u06cc \u0627\u0635\u0637\u0644\u0627\u062d\u06cc \u0627\u0633\u062a \u06a9\u0647 \u0645\u0639\u0645\u0648\u0644\u0627\u064b \u062f\u0631 \u0632\u0645\u06cc\u0646\u0647 \u0634\u0628\u06a9\u0647\u200c\u0647\u0627\u06cc \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u0630\u06a9\u0631 \u0645\u06cc\u200c\u0634\u0648\u062f. \u062f\u0631 \u0627\u06cc\u0646 \u0645\u0642\u0627\u0644\u0647\u060c \u0642\u0635\u062f \u062f\u0627\u0631\u06cc\u0645 \u0628\u0631\u0631\u0633\u06cc \u06a9\u0646\u06cc\u0645 \u06a9\u0647 \u0627\u06cc\u0646 \u0645\u0641\u0647\u0648\u0645 \u062f\u0642\u06cc\u0642\u0627\u064b \u0634\u0627\u0645\u0644 \u0686\u0647 \u0645\u0648\u0627\u0631\u062f\u06cc \u0627\u0633\u062a \u0648 \u0686\u0647 \u062a\u0623\u062b\u06cc\u0631\u06cc (\u062f\u0631 \u0635\u0648\u0631\u062a \u0648\u062c\u0648\u062f) \u0628\u0631 \u0639\u0645\u0644\u06a9\u0631\u062f \u06cc\u0627 \u0631\u0641\u062a\u0627\u0631 \u06a9\u0644\u06cc \u0634\u0628\u06a9\u0647\u200c\u0647\u0627\u06cc \u0639\u0635\u0628\u06cc \u06a9\u0627\u0646\u0648\u0644\u0648\u0634\u0646\u06cc \u062f\u0627\u0631\u062f. \u067e\u06cc\u0634\u200c\u0646\u06cc\u0627\u0632\u0647\u0627 \u2022 \u067e\u0627\u06cc\u062a\u0648\u0646: \u0628\u0631\u0627\u06cc \u0627\u062c\u0631\u0627\u06cc \u06a9\u062f\u0647\u0627\u06cc \u0627\u06cc\u0646 \u0645\u0642\u0627\u0644\u0647\u060c \u0633\u06cc\u0633\u062a\u0645 \u0634\u0645\u0627 \u0628\u0627\u06cc\u062f \u067e\u0627\u06cc\u062a\u0648\u0646 […]<\/p>\n","protected":false},"author":1,"featured_media":5352,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"yasr_overall_rating":0,"yasr_post_is_review":"","yasr_auto_insert_disabled":"","yasr_review_type":"","footnotes":""},"categories":[278],"tags":[393,392,353],"class_list":["post-5350","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-278","tag-393","tag-392","tag-353"],"aioseo_notices":[],"yasr_visitor_votes":{"stars_attributes":{"read_only":false,"span_bottom":false},"number_of_votes":0,"sum_votes":0},"_links":{"self":[{"href":"https:\/\/parmin.cloud\/blog\/wp-json\/wp\/v2\/posts\/5350","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/parmin.cloud\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/parmin.cloud\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/parmin.cloud\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/parmin.cloud\/blog\/wp-json\/wp\/v2\/comments?post=5350"}],"version-history":[{"count":1,"href":"https:\/\/parmin.cloud\/blog\/wp-json\/wp\/v2\/posts\/5350\/revisions"}],"predecessor-version":[{"id":5351,"href":"https:\/\/parmin.cloud\/blog\/wp-json\/wp\/v2\/posts\/5350\/revisions\/5351"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/parmin.cloud\/blog\/wp-json\/wp\/v2\/media\/5352"}],"wp:attachment":[{"href":"https:\/\/parmin.cloud\/blog\/wp-json\/wp\/v2\/media?parent=5350"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/parmin.cloud\/blog\/wp-json\/wp\/v2\/categories?post=5350"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/parmin.cloud\/blog\/wp-json\/wp\/v2\/tags?post=5350"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}