In [1]:
Copied!
!pip install tensorflow
!pip install tensorflow
Requirement already satisfied: tensorflow in /usr/local/lib/python3.10/dist-packages (2.17.1) Requirement already satisfied: absl-py>=1.0.0 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (1.4.0) Requirement already satisfied: astunparse>=1.6.0 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (1.6.3) Requirement already satisfied: flatbuffers>=24.3.25 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (24.3.25) Requirement already satisfied: gast!=0.5.0,!=0.5.1,!=0.5.2,>=0.2.1 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (0.6.0) Requirement already satisfied: google-pasta>=0.1.1 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (0.2.0) Requirement already satisfied: h5py>=3.10.0 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (3.12.1) Requirement already satisfied: libclang>=13.0.0 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (18.1.1) Requirement already satisfied: ml-dtypes<0.5.0,>=0.3.1 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (0.4.1) Requirement already satisfied: opt-einsum>=2.3.2 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (3.4.0) Requirement already satisfied: packaging in /usr/local/lib/python3.10/dist-packages (from tensorflow) (24.2) Requirement already satisfied: protobuf!=4.21.0,!=4.21.1,!=4.21.2,!=4.21.3,!=4.21.4,!=4.21.5,<5.0.0dev,>=3.20.3 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (4.25.5) Requirement already satisfied: requests<3,>=2.21.0 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (2.32.3) Requirement already satisfied: setuptools in /usr/local/lib/python3.10/dist-packages (from tensorflow) (75.1.0) Requirement already satisfied: six>=1.12.0 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (1.16.0) Requirement already satisfied: termcolor>=1.1.0 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (2.5.0) Requirement already satisfied: typing-extensions>=3.6.6 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (4.12.2) Requirement already satisfied: wrapt>=1.11.0 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (1.17.0) Requirement already satisfied: grpcio<2.0,>=1.24.3 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (1.68.1) Requirement already satisfied: tensorboard<2.18,>=2.17 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (2.17.1) Requirement already satisfied: keras>=3.2.0 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (3.5.0) Requirement already satisfied: tensorflow-io-gcs-filesystem>=0.23.1 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (0.37.1) Requirement already satisfied: numpy<2.0.0,>=1.23.5 in /usr/local/lib/python3.10/dist-packages (from tensorflow) (1.26.4) Requirement already satisfied: wheel<1.0,>=0.23.0 in /usr/local/lib/python3.10/dist-packages (from astunparse>=1.6.0->tensorflow) (0.45.1) Requirement already satisfied: rich in /usr/local/lib/python3.10/dist-packages (from keras>=3.2.0->tensorflow) (13.9.4) Requirement already satisfied: namex in /usr/local/lib/python3.10/dist-packages (from keras>=3.2.0->tensorflow) (0.0.8) Requirement already satisfied: optree in /usr/local/lib/python3.10/dist-packages (from keras>=3.2.0->tensorflow) (0.13.1) Requirement already satisfied: charset-normalizer<4,>=2 in /usr/local/lib/python3.10/dist-packages (from requests<3,>=2.21.0->tensorflow) (3.4.0) Requirement already satisfied: idna<4,>=2.5 in /usr/local/lib/python3.10/dist-packages (from requests<3,>=2.21.0->tensorflow) (3.10) Requirement already satisfied: urllib3<3,>=1.21.1 in /usr/local/lib/python3.10/dist-packages (from requests<3,>=2.21.0->tensorflow) (2.2.3) Requirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.10/dist-packages (from requests<3,>=2.21.0->tensorflow) (2024.8.30) Requirement already satisfied: markdown>=2.6.8 in /usr/local/lib/python3.10/dist-packages (from tensorboard<2.18,>=2.17->tensorflow) (3.7) Requirement already satisfied: tensorboard-data-server<0.8.0,>=0.7.0 in /usr/local/lib/python3.10/dist-packages (from tensorboard<2.18,>=2.17->tensorflow) (0.7.2) Requirement already satisfied: werkzeug>=1.0.1 in /usr/local/lib/python3.10/dist-packages (from tensorboard<2.18,>=2.17->tensorflow) (3.1.3) Requirement already satisfied: MarkupSafe>=2.1.1 in /usr/local/lib/python3.10/dist-packages (from werkzeug>=1.0.1->tensorboard<2.18,>=2.17->tensorflow) (3.0.2) Requirement already satisfied: markdown-it-py>=2.2.0 in /usr/local/lib/python3.10/dist-packages (from rich->keras>=3.2.0->tensorflow) (3.0.0) Requirement already satisfied: pygments<3.0.0,>=2.13.0 in /usr/local/lib/python3.10/dist-packages (from rich->keras>=3.2.0->tensorflow) (2.18.0) Requirement already satisfied: mdurl~=0.1 in /usr/local/lib/python3.10/dist-packages (from markdown-it-py>=2.2.0->rich->keras>=3.2.0->tensorflow) (0.1.2)
In [2]:
Copied!
import zipfile
import os
import shutil
import random
import gdown
from tqdm.auto import tqdm
from sklearn.linear_model import LogisticRegression
from tensorflow.keras.preprocessing.image import load_img, img_to_array
from torchvision import models, transforms
import torch
import cv2
from PIL import Image, ImageOps
from sklearn.svm import SVC
from sklearn.model_selection import train_test_split, GridSearchCV, RandomizedSearchCV
from sklearn.metrics import classification_report, confusion_matrix, accuracy_score
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
import zipfile
import os
import shutil
import random
import gdown
from tqdm.auto import tqdm
from sklearn.linear_model import LogisticRegression
from tensorflow.keras.preprocessing.image import load_img, img_to_array
from torchvision import models, transforms
import torch
import cv2
from PIL import Image, ImageOps
from sklearn.svm import SVC
from sklearn.model_selection import train_test_split, GridSearchCV, RandomizedSearchCV
from sklearn.metrics import classification_report, confusion_matrix, accuracy_score
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
In [3]:
Copied!
RANDOM_STATE = 42
random.seed(RANDOM_STATE)
RANDOM_STATE = 42
random.seed(RANDOM_STATE)
In [4]:
Copied!
file_id = "1FKZ9oHZ3zFMoFJX2f2aI34M2XZ2ikSb0"
gdown.download(
f"https://drive.google.com/uc?id={file_id}",
os.path.join(os.getcwd(), "dataset_32_classes.zip"),
quiet=False,
)
zip_name = "dataset_32_classes.zip"
file_id = "1FKZ9oHZ3zFMoFJX2f2aI34M2XZ2ikSb0"
gdown.download(
f"https://drive.google.com/uc?id={file_id}",
os.path.join(os.getcwd(), "dataset_32_classes.zip"),
quiet=False,
)
zip_name = "dataset_32_classes.zip"
Downloading... From (original): https://drive.google.com/uc?id=1FKZ9oHZ3zFMoFJX2f2aI34M2XZ2ikSb0 From (redirected): https://drive.google.com/uc?id=1FKZ9oHZ3zFMoFJX2f2aI34M2XZ2ikSb0&confirm=t&uuid=695eb786-1d3a-49a3-b0ad-33331af1edcc To: /content/dataset_32_classes.zip 100%|██████████| 641M/641M [00:13<00:00, 49.3MB/s]
In [5]:
Copied!
# Распаковка архива
with zipfile.ZipFile(zip_name, "r") as zip_ref:
zip_ref.extractall("./dataset")
# Распаковка архива
with zipfile.ZipFile(zip_name, "r") as zip_ref:
zip_ref.extractall("./dataset")
In [6]:
Copied!
DATASET_DIR = "./dataset"
TEMP_DIR = "./temp"
DATASET_DIR = "./dataset"
TEMP_DIR = "./temp"
In [7]:
Copied!
def set_image_size(img_path, save_path, size, color_background="white"):
img = Image.open(img_path)
if img.mode != "RGB":
img = img.convert("RGB")
ratio = img.width / img.height
# Широкое изображение
if ratio > 1:
new_width = size[0]
new_height = int(size[0] / ratio)
# Высокое изображение
else:
new_height = size[1]
new_width = int(size[1] * ratio)
img_resized = img.resize((new_width, new_height), Image.LANCZOS)
img_padded = ImageOps.pad(img_resized, size, color=color_background, centering=(0.5, 0.5))
img_padded.save(save_path)
def set_image_size(img_path, save_path, size, color_background="white"):
img = Image.open(img_path)
if img.mode != "RGB":
img = img.convert("RGB")
ratio = img.width / img.height
# Широкое изображение
if ratio > 1:
new_width = size[0]
new_height = int(size[0] / ratio)
# Высокое изображение
else:
new_height = size[1]
new_width = int(size[1] * ratio)
img_resized = img.resize((new_width, new_height), Image.LANCZOS)
img_padded = ImageOps.pad(img_resized, size, color=color_background, centering=(0.5, 0.5))
img_padded.save(save_path)
In [8]:
Copied!
def creat_temp_dataset(num_temp=None, use_sample=True):
if os.path.exists(TEMP_DIR):
shutil.rmtree(TEMP_DIR)
os.mkdir(TEMP_DIR)
for class_name in tqdm(os.listdir(DATASET_DIR)):
temp_class_path = os.path.join(TEMP_DIR, class_name)
if os.path.exists(temp_class_path) != True:
os.mkdir(temp_class_path)
basedir_class_path = os.path.join(DATASET_DIR, class_name)
image_names = os.listdir(basedir_class_path)
if use_sample and num_temp is not None:
# Если нужно использовать sample, и указано количество
reduce_image_names = random.sample(image_names, min(num_temp, len(image_names)))
else:
# Копируем все изображения
reduce_image_names = image_names
for image in reduce_image_names:
shutil.copy(os.path.join(basedir_class_path, image), os.path.join(temp_class_path, image))
def creat_temp_dataset(num_temp=None, use_sample=True):
if os.path.exists(TEMP_DIR):
shutil.rmtree(TEMP_DIR)
os.mkdir(TEMP_DIR)
for class_name in tqdm(os.listdir(DATASET_DIR)):
temp_class_path = os.path.join(TEMP_DIR, class_name)
if os.path.exists(temp_class_path) != True:
os.mkdir(temp_class_path)
basedir_class_path = os.path.join(DATASET_DIR, class_name)
image_names = os.listdir(basedir_class_path)
if use_sample and num_temp is not None:
# Если нужно использовать sample, и указано количество
reduce_image_names = random.sample(image_names, min(num_temp, len(image_names)))
else:
# Копируем все изображения
reduce_image_names = image_names
for image in reduce_image_names:
shutil.copy(os.path.join(basedir_class_path, image), os.path.join(temp_class_path, image))
In [9]:
Copied!
def resize_temp_dataset(size, color_background):
# Проверка, существует ли временный датасет
if not os.path.exists(TEMP_DIR):
print("Временный датасет TEMP_DIR не найден.")
return
# Проходим по всем классам (папкам) в TEMP_DIR
for class_name in tqdm(os.listdir(TEMP_DIR)):
class_path = os.path.join(TEMP_DIR, class_name)
# Проверяем, является ли это папкой
if os.path.isdir(class_path):
for image_name in os.listdir(class_path):
image_path = os.path.join(class_path, image_name)
# Проверяем, является ли это изображением
if image_name.lower().endswith((".png", ".jpg", ".jpeg")):
save_path = image_path # Сохраняем под тем же именем
set_image_size(image_path, save_path, size, color_background)
def resize_temp_dataset(size, color_background):
# Проверка, существует ли временный датасет
if not os.path.exists(TEMP_DIR):
print("Временный датасет TEMP_DIR не найден.")
return
# Проходим по всем классам (папкам) в TEMP_DIR
for class_name in tqdm(os.listdir(TEMP_DIR)):
class_path = os.path.join(TEMP_DIR, class_name)
# Проверяем, является ли это папкой
if os.path.isdir(class_path):
for image_name in os.listdir(class_path):
image_path = os.path.join(class_path, image_name)
# Проверяем, является ли это изображением
if image_name.lower().endswith((".png", ".jpg", ".jpeg")):
save_path = image_path # Сохраняем под тем же именем
set_image_size(image_path, save_path, size, color_background)
In [10]:
Copied!
# Функция для преобразования изображений перед передачей в модель ResNet18
def get_transform(size=(224, 224)):
return transforms.Compose(
[
transforms.ToPILImage(), # Преобразование в PIL Image
transforms.Resize(size), # Изменение размера
transforms.ToTensor(), # Преобразование в тензор
transforms.Normalize( # Нормализация в диапазон модели
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
),
]
)
# Функция для преобразования изображений перед передачей в модель ResNet18
def get_transform(size=(224, 224)):
return transforms.Compose(
[
transforms.ToPILImage(), # Преобразование в PIL Image
transforms.Resize(size), # Изменение размера
transforms.ToTensor(), # Преобразование в тензор
transforms.Normalize( # Нормализация в диапазон модели
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
),
]
)
In [11]:
Copied!
# Функция для получения признаков и меток классов из изображений
def get_X_and_y(size=(224, 224), color="color"):
X = []
y = []
for class_name in os.listdir(TEMP_DIR):
path_class = os.path.join(TEMP_DIR, class_name)
for img in os.listdir(path_class):
img_path = os.path.join(path_class, img)
# Загружаем изображение
if color == "grey":
img = load_img(img_path, color_mode="grayscale") # Загрузка в градациях серого
img = img_to_array(img)
img = np.repeat(img, 3, axis=-1)
else:
img = load_img(img_path) # Загрузка цветного изображения
img = img_to_array(img)
# Добавляем размерность для модели
transform = get_transform(size)
img = transform(img) # Применяем трансформации
img = img.unsqueeze(0) # Добавляем размерность батча
with torch.no_grad():
features = resn18(img) # Извлекаем признаки
X.append(features.squeeze().numpy()) # Плоский массив
y.append(class_name)
return np.array(X), np.array(y)
# Функция для получения признаков и меток классов из изображений
def get_X_and_y(size=(224, 224), color="color"):
X = []
y = []
for class_name in os.listdir(TEMP_DIR):
path_class = os.path.join(TEMP_DIR, class_name)
for img in os.listdir(path_class):
img_path = os.path.join(path_class, img)
# Загружаем изображение
if color == "grey":
img = load_img(img_path, color_mode="grayscale") # Загрузка в градациях серого
img = img_to_array(img)
img = np.repeat(img, 3, axis=-1)
else:
img = load_img(img_path) # Загрузка цветного изображения
img = img_to_array(img)
# Добавляем размерность для модели
transform = get_transform(size)
img = transform(img) # Применяем трансформации
img = img.unsqueeze(0) # Добавляем размерность батча
with torch.no_grad():
features = resn18(img) # Извлекаем признаки
X.append(features.squeeze().numpy()) # Плоский массив
y.append(class_name)
return np.array(X), np.array(y)
In [12]:
Copied!
# Загрузка модели ResNet18
resn18 = models.resnet18(pretrained=True)
resn18 = torch.nn.Sequential(*(list(resn18.children())[:-1])) # Удаление последнего слоя (выходного классификатора)
for param in resn18.parameters():
param.requires_grad = False
resn18.eval() # Установка модели в режим оценки (eval), нам нужно использовать resnet18 только для извлечения признаков, модель будет обучаться с помощью SVC и LogisticRegression
# Загрузка модели ResNet18
resn18 = models.resnet18(pretrained=True)
resn18 = torch.nn.Sequential(*(list(resn18.children())[:-1])) # Удаление последнего слоя (выходного классификатора)
for param in resn18.parameters():
param.requires_grad = False
resn18.eval() # Установка модели в режим оценки (eval), нам нужно использовать resnet18 только для извлечения признаков, модель будет обучаться с помощью SVC и LogisticRegression
/usr/local/lib/python3.10/dist-packages/torchvision/models/_utils.py:208: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and may be removed in the future, please use 'weights' instead. warnings.warn( /usr/local/lib/python3.10/dist-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and may be removed in the future. The current behavior is equivalent to passing `weights=ResNet18_Weights.IMAGENET1K_V1`. You can also use `weights=ResNet18_Weights.DEFAULT` to get the most up-to-date weights. warnings.warn(msg) Downloading: "https://download.pytorch.org/models/resnet18-f37072fd.pth" to /root/.cache/torch/hub/checkpoints/resnet18-f37072fd.pth 100%|██████████| 44.7M/44.7M [00:00<00:00, 109MB/s]
Out[12]:
Sequential(
(0): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
(1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace=True)
(3): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
(4): Sequential(
(0): BasicBlock(
(conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(1): BasicBlock(
(conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(5): Sequential(
(0): BasicBlock(
(conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(downsample): Sequential(
(0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(1): BasicBlock(
(conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(6): Sequential(
(0): BasicBlock(
(conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(downsample): Sequential(
(0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(1): BasicBlock(
(conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(7): Sequential(
(0): BasicBlock(
(conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(downsample): Sequential(
(0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(1): BasicBlock(
(conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(8): AdaptiveAvgPool2d(output_size=(1, 1))
)
Обучение модели на выборочном датасете из фотографий размером 224x224 px¶
Создаем выборочный датасет из 500 изображений в каждом классе
In [13]:
Copied!
creat_temp_dataset(num_temp=500, use_sample=True)
creat_temp_dataset(num_temp=500, use_sample=True)
0%| | 0/32 [00:00<?, ?it/s]
Приводим все изображения в выборочном датасете к размеру 224x224 px (рекомендованный размер фотографий для использования ResNet18)
In [14]:
Copied!
resize_temp_dataset((224, 224), color_background="white")
resize_temp_dataset((224, 224), color_background="white")
0%| | 0/32 [00:00<?, ?it/s]
In [15]:
Copied!
X, y = get_X_and_y(color="color")
X, y = get_X_and_y(color="color")
In [16]:
Copied!
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
Делаем подбор гиперпараметров с помощью GridSearchCV
In [17]:
Copied!
param_grid = {
"C": [1, 5, 10, 50],
"kernel": ["rbf", "linear", "poly", "sigmoid"],
"gamma": ["scale"],
}
svс = GridSearchCV(SVC(), param_grid)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
param_grid = {
"C": [1, 5, 10, 50],
"kernel": ["rbf", "linear", "poly", "sigmoid"],
"gamma": ["scale"],
}
svс = GridSearchCV(SVC(), param_grid)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
In [18]:
Copied!
svс.best_params_, svс.best_score_
svс.best_params_, svс.best_score_
Out[18]:
({'C': 50, 'gamma': 'scale', 'kernel': 'rbf'}, 0.961328125)
In [19]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.88 0.89 0.89 94
Avocado 0.96 0.86 0.91 103
Banana 0.88 0.96 0.91 90
Bean 0.97 1.00 0.99 110
Bitter_Gourd 1.00 1.00 1.00 124
Bottle_Gourd 1.00 0.99 1.00 108
Brinjal 0.99 0.99 0.99 103
Broccoli 0.99 0.99 0.99 100
Cabbage 0.99 1.00 0.99 83
Capsicum 0.98 0.99 0.98 97
Carrot 1.00 1.00 1.00 98
Cauliflower 0.99 0.99 0.99 93
Cherry 0.96 0.91 0.94 117
Cucumber 1.00 0.99 0.99 97
Grape 0.99 1.00 1.00 111
Kiwi 0.90 0.95 0.92 94
Mango 0.87 0.89 0.88 87
Nut 1.00 1.00 1.00 85
Onion 1.00 1.00 1.00 110
Orange 0.94 0.85 0.89 103
Papaya 1.00 1.00 1.00 97
Peach 0.99 1.00 0.99 84
Pear 1.00 1.00 1.00 92
Pepper 1.00 1.00 1.00 99
Pinenapple 0.92 0.93 0.93 90
Plum 1.00 1.00 1.00 102
Potato 1.00 1.00 1.00 94
Pumpkin 0.99 1.00 1.00 111
Radish 1.00 0.99 0.99 100
Strawberry 0.92 0.97 0.94 94
Tomato 0.99 1.00 1.00 115
Watermelon 0.94 0.91 0.93 115
accuracy 0.97 3200
macro avg 0.97 0.97 0.97 3200
weighted avg 0.97 0.97 0.97 3200
Попытаемся получше подобрать параметр C, тем самым улучшив метрики
In [20]:
Copied!
param_grid = {"C": [40, 45, 50, 60], "kernel": ["rbf"], "gamma": ["scale"]}
svс = GridSearchCV(SVC(), param_grid, cv=2)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
param_grid = {"C": [40, 45, 50, 60], "kernel": ["rbf"], "gamma": ["scale"]}
svс = GridSearchCV(SVC(), param_grid, cv=2)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
In [21]:
Copied!
svс.best_params_, svс.best_score_
svс.best_params_, svс.best_score_
Out[21]:
({'C': 40, 'gamma': 'scale', 'kernel': 'rbf'}, 0.9451562499999999)
In [22]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.88 0.89 0.89 94
Avocado 0.96 0.86 0.91 103
Banana 0.88 0.96 0.91 90
Bean 0.97 1.00 0.99 110
Bitter_Gourd 1.00 1.00 1.00 124
Bottle_Gourd 1.00 0.99 1.00 108
Brinjal 0.99 0.99 0.99 103
Broccoli 0.99 0.99 0.99 100
Cabbage 0.99 1.00 0.99 83
Capsicum 0.98 0.99 0.98 97
Carrot 1.00 1.00 1.00 98
Cauliflower 0.99 0.99 0.99 93
Cherry 0.96 0.91 0.94 117
Cucumber 1.00 0.99 0.99 97
Grape 0.99 1.00 1.00 111
Kiwi 0.90 0.95 0.92 94
Mango 0.87 0.89 0.88 87
Nut 1.00 1.00 1.00 85
Onion 1.00 1.00 1.00 110
Orange 0.94 0.85 0.89 103
Papaya 1.00 1.00 1.00 97
Peach 0.99 1.00 0.99 84
Pear 1.00 1.00 1.00 92
Pepper 1.00 1.00 1.00 99
Pinenapple 0.92 0.93 0.93 90
Plum 1.00 1.00 1.00 102
Potato 1.00 1.00 1.00 94
Pumpkin 0.99 1.00 1.00 111
Radish 1.00 0.99 0.99 100
Strawberry 0.92 0.97 0.94 94
Tomato 0.99 1.00 1.00 115
Watermelon 0.94 0.91 0.93 115
accuracy 0.97 3200
macro avg 0.97 0.97 0.97 3200
weighted avg 0.97 0.97 0.97 3200
In [23]:
Copied!
y_name = os.listdir(DATASET_DIR)
y_name = os.listdir(DATASET_DIR)
In [24]:
Copied!
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_test, y_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_test, y_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
In [25]:
Copied!
param_grid = {"C": np.arange(12, 42, 2), "kernel": ["rbf"], "gamma": ["scale"]}
svс = GridSearchCV(SVC(), param_grid)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
param_grid = {"C": np.arange(12, 42, 2), "kernel": ["rbf"], "gamma": ["scale"]}
svс = GridSearchCV(SVC(), param_grid)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
/usr/local/lib/python3.10/dist-packages/numpy/ma/core.py:2820: RuntimeWarning: invalid value encountered in cast _data = np.array(data, dtype=dtype, copy=copy,
In [26]:
Copied!
svс.best_params_, svс.best_score_
svс.best_params_, svс.best_score_
Out[26]:
({'C': 12, 'gamma': 'scale', 'kernel': 'rbf'}, 0.961328125)
In [27]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.88 0.89 0.89 94
Avocado 0.96 0.86 0.91 103
Banana 0.88 0.96 0.91 90
Bean 0.97 1.00 0.99 110
Bitter_Gourd 1.00 1.00 1.00 124
Bottle_Gourd 1.00 0.99 1.00 108
Brinjal 0.99 0.99 0.99 103
Broccoli 0.99 0.99 0.99 100
Cabbage 0.99 1.00 0.99 83
Capsicum 0.98 0.99 0.98 97
Carrot 1.00 1.00 1.00 98
Cauliflower 0.99 0.99 0.99 93
Cherry 0.96 0.91 0.94 117
Cucumber 1.00 0.99 0.99 97
Grape 0.99 1.00 1.00 111
Kiwi 0.90 0.95 0.92 94
Mango 0.87 0.89 0.88 87
Nut 1.00 1.00 1.00 85
Onion 1.00 1.00 1.00 110
Orange 0.94 0.85 0.89 103
Papaya 1.00 1.00 1.00 97
Peach 0.99 1.00 0.99 84
Pear 1.00 1.00 1.00 92
Pepper 1.00 1.00 1.00 99
Pinenapple 0.92 0.93 0.93 90
Plum 1.00 1.00 1.00 102
Potato 1.00 1.00 1.00 94
Pumpkin 0.99 1.00 1.00 111
Radish 1.00 0.99 0.99 100
Strawberry 0.92 0.97 0.94 94
Tomato 0.99 1.00 1.00 115
Watermelon 0.94 0.91 0.93 115
accuracy 0.97 3200
macro avg 0.97 0.97 0.97 3200
weighted avg 0.97 0.97 0.97 3200
In [28]:
Copied!
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_test, y_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_test, y_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
Проверим как модель обучается с помощью логистической регрессии
In [29]:
Copied!
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(max_iter=200)), # Увеличиваем max_iter
]
)
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(max_iter=200)), # Увеличиваем max_iter
]
)
In [30]:
Copied!
param_grid = {"logreg__C": [0.01, 0.1, 1, 10]}
logreg = GridSearchCV(pipeline, param_grid)
logreg.fit(X_train, y_train)
y_pred = logreg.predict(X_test)
param_grid = {"logreg__C": [0.01, 0.1, 1, 10]}
logreg = GridSearchCV(pipeline, param_grid)
logreg.fit(X_train, y_train)
y_pred = logreg.predict(X_test)
In [31]:
Copied!
logreg.best_params_, logreg.best_score_
logreg.best_params_, logreg.best_score_
Out[31]:
({'logreg__C': 0.1}, 0.9353906249999999)
In [32]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.75 0.74 0.75 90
Avocado 0.80 0.84 0.82 110
Banana 0.91 0.94 0.92 124
Bean 0.94 0.99 0.96 90
Bitter_Gourd 1.00 1.00 1.00 111
Bottle_Gourd 0.97 0.98 0.98 110
Brinjal 0.97 0.97 0.97 115
Broccoli 0.98 0.99 0.98 92
Cabbage 0.98 0.98 0.98 84
Capsicum 0.99 1.00 1.00 103
Carrot 0.99 1.00 0.99 99
Cauliflower 0.95 0.97 0.96 87
Cherry 0.79 0.86 0.83 94
Cucumber 0.99 0.92 0.95 103
Grape 0.98 0.99 0.99 115
Kiwi 0.91 0.89 0.90 97
Mango 0.92 0.74 0.82 117
Nut 1.00 1.00 1.00 102
Onion 0.98 1.00 0.99 103
Orange 0.74 0.77 0.76 111
Papaya 0.93 0.99 0.96 97
Peach 1.00 1.00 1.00 100
Pear 0.96 1.00 0.98 100
Pepper 0.97 1.00 0.98 83
Pinenapple 0.92 0.89 0.91 94
Plum 1.00 0.99 0.99 94
Potato 0.99 0.98 0.98 98
Pumpkin 0.99 0.95 0.97 94
Radish 0.98 0.98 0.98 93
Strawberry 0.88 0.87 0.87 97
Tomato 0.92 0.89 0.90 85
Watermelon 0.91 0.89 0.90 108
accuracy 0.94 3200
macro avg 0.94 0.94 0.94 3200
weighted avg 0.94 0.94 0.94 3200
SVC:модель показала лучшие данные на цветных фотографиях (accuracy на трейне = 0.96, accuracy на тесте = 0.97) при С = 12, kernel = "rbf". LogisticRegression: при с = 0.1 accuracy на трейне = 0.94, accuracy на тесте = 0.94
Обучим модель на черно-белых фотографиях и посмотрим значения метрик
In [33]:
Copied!
X_grey, y_grey = get_X_and_y("grey")
X_grey, y_grey = get_X_and_y("grey")
In [34]:
Copied!
X_grey_train, X_grey_test, y_grey_train, y_grey_test = train_test_split(X_grey, y_grey, test_size=0.2, random_state=42)
X_grey_train, X_grey_test, y_grey_train, y_grey_test = train_test_split(X_grey, y_grey, test_size=0.2, random_state=42)
In [35]:
Copied!
param_grid = {"C": [5, 7, 10, 12, 14], "kernel": ["rbf"], "gamma": ["scale"]}
svс_grey = GridSearchCV(SVC(), param_grid)
svс_grey.fit(X_train, y_train)
y_grey_pred = svс_grey.predict(X_grey_test)
param_grid = {"C": [5, 7, 10, 12, 14], "kernel": ["rbf"], "gamma": ["scale"]}
svс_grey = GridSearchCV(SVC(), param_grid)
svс_grey.fit(X_train, y_train)
y_grey_pred = svс_grey.predict(X_grey_test)
In [36]:
Copied!
svс_grey.best_params_, svс_grey.best_score_
svс_grey.best_params_, svс_grey.best_score_
Out[36]:
({'C': 12, 'gamma': 'scale', 'kernel': 'rbf'}, 0.961328125)
In [37]:
Copied!
print(classification_report(y_grey_test, y_grey_pred))
print(classification_report(y_grey_test, y_grey_pred))
precision recall f1-score support
Apple 0.39 0.64 0.49 94
Avocado 0.42 0.86 0.57 103
Banana 0.70 0.83 0.76 90
Bean 0.92 0.92 0.92 110
Bitter_Gourd 1.00 0.12 0.22 124
Bottle_Gourd 0.97 0.77 0.86 108
Brinjal 0.72 0.93 0.81 103
Broccoli 1.00 0.93 0.96 100
Cabbage 1.00 0.83 0.91 83
Capsicum 1.00 0.11 0.20 97
Carrot 0.86 0.63 0.73 98
Cauliflower 0.93 0.98 0.95 93
Cherry 0.67 0.51 0.58 117
Cucumber 0.85 0.63 0.72 97
Grape 0.77 0.97 0.86 111
Kiwi 0.47 0.82 0.60 94
Mango 0.60 0.53 0.56 87
Nut 0.81 0.93 0.87 85
Onion 1.00 0.86 0.93 110
Orange 0.68 0.27 0.39 103
Papaya 0.35 0.06 0.11 97
Peach 0.96 0.55 0.70 84
Pear 0.96 0.47 0.63 92
Pepper 0.94 0.80 0.86 99
Pinenapple 0.84 0.90 0.87 90
Plum 0.54 0.92 0.68 102
Potato 0.41 0.80 0.54 94
Pumpkin 0.65 0.95 0.77 111
Radish 0.74 0.92 0.82 100
Strawberry 0.97 0.33 0.49 94
Tomato 0.82 0.69 0.75 115
Watermelon 0.53 0.94 0.68 115
accuracy 0.70 3200
macro avg 0.76 0.70 0.68 3200
weighted avg 0.76 0.70 0.68 3200
In [38]:
Copied!
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_grey_test, y_grey_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_grey_test, y_grey_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
При с = 12, kernel = 'rbf' accuracy на трейне = 0.96, accuracy на тесте = 0.7, что может говорить о переобучении. Так как значения метрик сильно лучше на цветных фотографиях, то дальнейшее обучение будет проходить на датасете цветных фотографий
Обучение модели на выборочном датасете из фотографий размером 64x64 px¶
In [39]:
Copied!
creat_temp_dataset(num_temp=500, use_sample=True)
creat_temp_dataset(num_temp=500, use_sample=True)
0%| | 0/32 [00:00<?, ?it/s]
In [40]:
Copied!
resize_temp_dataset((64, 64), color_background="white")
resize_temp_dataset((64, 64), color_background="white")
0%| | 0/32 [00:00<?, ?it/s]
In [41]:
Copied!
X, y = get_X_and_y(size=(64, 64), color="color")
X, y = get_X_and_y(size=(64, 64), color="color")
In [42]:
Copied!
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
In [43]:
Copied!
param_grid = {"C": [8, 10, 12, 14], "kernel": ["rbf"], "gamma": ["scale"]}
svс = GridSearchCV(SVC(), param_grid)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
param_grid = {"C": [8, 10, 12, 14], "kernel": ["rbf"], "gamma": ["scale"]}
svс = GridSearchCV(SVC(), param_grid)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
In [44]:
Copied!
svс.best_params_, svс.best_score_
svс.best_params_, svс.best_score_
Out[44]:
({'C': 14, 'gamma': 'scale', 'kernel': 'rbf'}, 0.8943749999999999)
In [45]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.70 0.66 0.68 94
Avocado 0.82 0.77 0.79 103
Banana 0.70 0.84 0.76 90
Bean 0.96 0.97 0.96 110
Bitter_Gourd 0.98 0.95 0.97 124
Bottle_Gourd 0.95 1.00 0.97 108
Brinjal 0.90 0.94 0.92 103
Broccoli 0.94 0.95 0.95 100
Cabbage 0.92 0.95 0.93 83
Capsicum 0.90 0.98 0.94 97
Carrot 0.99 1.00 0.99 98
Cauliflower 0.96 0.98 0.97 93
Cherry 0.73 0.79 0.76 117
Cucumber 0.98 0.93 0.95 97
Grape 1.00 1.00 1.00 111
Kiwi 0.77 0.68 0.72 94
Mango 0.78 0.82 0.80 87
Nut 0.97 1.00 0.98 85
Onion 1.00 1.00 1.00 110
Orange 0.84 0.76 0.80 103
Papaya 0.94 0.94 0.94 97
Peach 0.99 1.00 0.99 84
Pear 0.97 0.98 0.97 92
Pepper 1.00 1.00 1.00 99
Pinenapple 0.75 0.81 0.78 90
Plum 1.00 1.00 1.00 102
Potato 1.00 0.98 0.99 94
Pumpkin 0.95 0.94 0.94 111
Radish 0.99 0.98 0.98 100
Strawberry 0.75 0.77 0.76 94
Tomato 0.95 0.88 0.91 115
Watermelon 0.84 0.69 0.76 115
accuracy 0.90 3200
macro avg 0.90 0.90 0.90 3200
weighted avg 0.90 0.90 0.90 3200
In [46]:
Copied!
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_test, y_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_test, y_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
In [47]:
Copied!
param_grid = {"C": np.arange(10, 20, 2), "kernel": ["rbf"], "gamma": ["scale"]}
svс = GridSearchCV(SVC(), param_grid)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
param_grid = {"C": np.arange(10, 20, 2), "kernel": ["rbf"], "gamma": ["scale"]}
svс = GridSearchCV(SVC(), param_grid)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
In [48]:
Copied!
svс.best_params_, svс.best_score_
svс.best_params_, svс.best_score_
Out[48]:
({'C': 18, 'gamma': 'scale', 'kernel': 'rbf'}, 0.8946093749999999)
In [49]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.70 0.67 0.68 94
Avocado 0.82 0.77 0.79 103
Banana 0.70 0.84 0.76 90
Bean 0.96 0.97 0.96 110
Bitter_Gourd 0.98 0.95 0.97 124
Bottle_Gourd 0.95 1.00 0.97 108
Brinjal 0.90 0.94 0.92 103
Broccoli 0.94 0.95 0.95 100
Cabbage 0.91 0.95 0.93 83
Capsicum 0.90 0.98 0.94 97
Carrot 0.99 1.00 0.99 98
Cauliflower 0.96 0.98 0.97 93
Cherry 0.73 0.79 0.76 117
Cucumber 0.98 0.93 0.95 97
Grape 1.00 1.00 1.00 111
Kiwi 0.78 0.68 0.73 94
Mango 0.77 0.79 0.78 87
Nut 0.97 1.00 0.98 85
Onion 1.00 1.00 1.00 110
Orange 0.80 0.75 0.77 103
Papaya 0.94 0.94 0.94 97
Peach 0.99 1.00 0.99 84
Pear 0.97 0.98 0.97 92
Pepper 1.00 1.00 1.00 99
Pinenapple 0.76 0.80 0.78 90
Plum 1.00 1.00 1.00 102
Potato 1.00 0.98 0.99 94
Pumpkin 0.95 0.94 0.94 111
Radish 0.99 0.98 0.98 100
Strawberry 0.78 0.78 0.78 94
Tomato 0.95 0.88 0.91 115
Watermelon 0.83 0.69 0.75 115
accuracy 0.90 3200
macro avg 0.90 0.90 0.90 3200
weighted avg 0.90 0.90 0.90 3200
Проверим как модель обучается с помощью логистической регрессии на фотографиях размера 64x64
In [50]:
Copied!
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(max_iter=200)), # Увеличиваем max_iter
]
)
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(max_iter=200)), # Увеличиваем max_iter
]
)
In [51]:
Copied!
param_grid = {"logreg__C": [0.01, 0.1, 1, 10]}
logreg = GridSearchCV(pipeline, param_grid)
logreg.fit(X_train, y_train)
y_pred = logreg.predict(X_test)
param_grid = {"logreg__C": [0.01, 0.1, 1, 10]}
logreg = GridSearchCV(pipeline, param_grid)
logreg.fit(X_train, y_train)
y_pred = logreg.predict(X_test)
In [52]:
Copied!
logreg.best_params_, logreg.best_score_
logreg.best_params_, logreg.best_score_
Out[52]:
({'logreg__C': 0.1}, 0.848046875)
In [53]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.64 0.57 0.60 90
Avocado 0.69 0.71 0.70 110
Banana 0.84 0.74 0.79 124
Bean 0.91 0.90 0.91 90
Bitter_Gourd 0.91 0.94 0.92 111
Bottle_Gourd 0.95 0.95 0.95 110
Brinjal 0.87 0.86 0.86 115
Broccoli 0.96 0.95 0.95 92
Cabbage 0.93 0.89 0.91 84
Capsicum 0.88 0.94 0.91 103
Carrot 0.97 0.96 0.96 99
Cauliflower 0.97 0.97 0.97 87
Cherry 0.67 0.78 0.72 94
Cucumber 0.88 0.86 0.87 103
Grape 0.99 1.00 1.00 115
Kiwi 0.68 0.76 0.72 97
Mango 0.67 0.61 0.64 117
Nut 0.97 0.99 0.98 102
Onion 1.00 1.00 1.00 103
Orange 0.73 0.60 0.66 111
Papaya 0.90 0.90 0.90 97
Peach 0.98 0.98 0.98 100
Pear 0.93 0.97 0.95 100
Pepper 0.93 0.99 0.96 83
Pinenapple 0.68 0.71 0.69 94
Plum 1.00 1.00 1.00 94
Potato 0.95 0.95 0.95 98
Pumpkin 0.91 0.87 0.89 94
Radish 0.98 0.95 0.96 93
Strawberry 0.70 0.78 0.74 97
Tomato 0.83 0.85 0.84 85
Watermelon 0.70 0.70 0.70 108
accuracy 0.86 3200
macro avg 0.86 0.86 0.86 3200
weighted avg 0.86 0.86 0.86 3200
SVC: accuracy на трейне = 0.89, accuracy на тесте = 0.90 при С = 18, kernel = "rbf". LogisticRegression: при с = 0.1 accuracy на трейне = 0.85, accuracy на тесте = 0.86
Обучение модели на выборочном датасете из фотографий размером 128x128 px¶
In [54]:
Copied!
creat_temp_dataset(num_temp=500, use_sample=True)
creat_temp_dataset(num_temp=500, use_sample=True)
0%| | 0/32 [00:00<?, ?it/s]
In [55]:
Copied!
resize_temp_dataset((128, 128), color_background="white")
resize_temp_dataset((128, 128), color_background="white")
0%| | 0/32 [00:00<?, ?it/s]
In [56]:
Copied!
X, y = get_X_and_y(size=(128, 128), color="color")
X, y = get_X_and_y(size=(128, 128), color="color")
In [57]:
Copied!
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
In [58]:
Copied!
param_grid = {"C": [4, 8, 10, 12, 16], "kernel": ["rbf"], "gamma": ["scale"]}
svс = GridSearchCV(SVC(), param_grid)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
param_grid = {"C": [4, 8, 10, 12, 16], "kernel": ["rbf"], "gamma": ["scale"]}
svс = GridSearchCV(SVC(), param_grid)
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
In [59]:
Copied!
svс.best_params_, svс.best_score_
svс.best_params_, svс.best_score_
Out[59]:
({'C': 10, 'gamma': 'scale', 'kernel': 'rbf'}, 0.9364062499999999)
In [60]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.86 0.83 0.84 94
Avocado 0.86 0.86 0.86 103
Banana 0.86 0.86 0.86 90
Bean 0.97 1.00 0.99 110
Bitter_Gourd 0.99 1.00 1.00 124
Bottle_Gourd 0.99 0.99 0.99 108
Brinjal 0.97 0.97 0.97 103
Broccoli 1.00 0.98 0.99 100
Cabbage 1.00 1.00 1.00 83
Capsicum 0.94 0.97 0.95 97
Carrot 1.00 0.99 0.99 98
Cauliflower 1.00 1.00 1.00 93
Cherry 0.90 0.81 0.85 117
Cucumber 0.97 0.95 0.96 97
Grape 1.00 1.00 1.00 111
Kiwi 0.87 0.86 0.87 94
Mango 0.78 0.79 0.78 87
Nut 1.00 1.00 1.00 85
Onion 0.99 1.00 1.00 110
Orange 0.85 0.77 0.81 103
Papaya 0.96 1.00 0.98 97
Peach 0.99 1.00 0.99 84
Pear 1.00 1.00 1.00 92
Pepper 1.00 1.00 1.00 99
Pinenapple 0.89 0.97 0.93 90
Plum 1.00 1.00 1.00 102
Potato 0.98 0.98 0.98 94
Pumpkin 0.96 0.98 0.97 111
Radish 1.00 0.99 0.99 100
Strawberry 0.85 0.82 0.83 94
Tomato 0.97 0.98 0.97 115
Watermelon 0.89 0.94 0.91 115
accuracy 0.95 3200
macro avg 0.95 0.95 0.95 3200
weighted avg 0.95 0.95 0.95 3200
In [61]:
Copied!
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_test, y_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_test, y_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
Проверим как модель обучается с помощью логистической регрессии на фотографиях размера 128x128
In [62]:
Copied!
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(max_iter=200)), # Увеличиваем max_iter
]
)
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(max_iter=200)), # Увеличиваем max_iter
]
)
In [63]:
Copied!
param_grid = {"logreg__C": [0.01, 0.1, 1, 10]}
logreg = GridSearchCV(pipeline, param_grid)
logreg.fit(X_train, y_train)
y_pred = logreg.predict(X_test)
param_grid = {"logreg__C": [0.01, 0.1, 1, 10]}
logreg = GridSearchCV(pipeline, param_grid)
logreg.fit(X_train, y_train)
y_pred = logreg.predict(X_test)
In [64]:
Copied!
logreg.best_params_, logreg.best_score_
logreg.best_params_, logreg.best_score_
Out[64]:
({'logreg__C': 0.1}, 0.899453125)
In [65]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.69 0.70 0.70 90
Avocado 0.76 0.81 0.78 110
Banana 0.82 0.75 0.78 124
Bean 0.94 0.97 0.95 90
Bitter_Gourd 0.99 0.97 0.98 111
Bottle_Gourd 0.96 0.96 0.96 110
Brinjal 1.00 0.90 0.95 115
Broccoli 0.99 1.00 0.99 92
Cabbage 0.99 1.00 0.99 84
Capsicum 0.93 0.94 0.94 103
Carrot 0.98 0.97 0.97 99
Cauliflower 0.98 0.99 0.98 87
Cherry 0.69 0.77 0.73 94
Cucumber 0.92 0.95 0.93 103
Grape 1.00 1.00 1.00 115
Kiwi 0.82 0.73 0.77 97
Mango 0.84 0.79 0.81 117
Nut 0.99 1.00 1.00 102
Onion 0.98 1.00 0.99 103
Orange 0.81 0.79 0.80 111
Papaya 0.91 0.96 0.93 97
Peach 0.99 1.00 1.00 100
Pear 0.98 0.96 0.97 100
Pepper 0.97 1.00 0.98 83
Pinenapple 0.81 0.85 0.83 94
Plum 1.00 1.00 1.00 94
Potato 0.96 0.98 0.97 98
Pumpkin 0.97 0.95 0.96 94
Radish 0.94 0.98 0.96 93
Strawberry 0.80 0.76 0.78 97
Tomato 0.93 0.91 0.92 85
Watermelon 0.75 0.77 0.76 108
accuracy 0.91 3200
macro avg 0.91 0.91 0.91 3200
weighted avg 0.91 0.91 0.91 3200
SVC: accuracy на трейне = 0.94, accuracy на тесте = 0.95 при С = 10, kernel = "rbf". LogisticRegression: при с = 0.1 accuracy на трейне = 0.90, accuracy на тесте = 0.91
Обучение модели на полном датасете цветных фотографий размера 224x224 px¶
In [66]:
Copied!
creat_temp_dataset(use_sample=False)
creat_temp_dataset(use_sample=False)
0%| | 0/32 [00:00<?, ?it/s]
In [67]:
Copied!
resize_temp_dataset((224, 224), color_background="white")
resize_temp_dataset((224, 224), color_background="white")
0%| | 0/32 [00:00<?, ?it/s]
In [68]:
Copied!
X, y = get_X_and_y(color="color")
X, y = get_X_and_y(color="color")
In [69]:
Copied!
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
In [70]:
Copied!
svс = SVC(C=12, kernel="rbf")
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
svс = SVC(C=12, kernel="rbf")
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
In [71]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.95 0.93 0.94 290
Avocado 0.93 0.94 0.94 286
Banana 0.90 0.95 0.92 253
Bean 1.00 1.00 1.00 271
Bitter_Gourd 1.00 1.00 1.00 259
Bottle_Gourd 1.00 1.00 1.00 259
Brinjal 0.99 0.99 0.99 303
Broccoli 1.00 0.99 1.00 294
Cabbage 0.99 1.00 1.00 296
Capsicum 0.99 1.00 0.99 273
Carrot 1.00 1.00 1.00 286
Cauliflower 1.00 0.99 0.99 299
Cherry 0.94 0.94 0.94 273
Cucumber 1.00 1.00 1.00 271
Grape 1.00 1.00 1.00 273
Kiwi 0.96 0.94 0.95 252
Mango 0.90 0.91 0.90 296
Nut 1.00 1.00 1.00 287
Onion 1.00 1.00 1.00 289
Orange 0.95 0.90 0.92 292
Papaya 0.99 1.00 0.99 297
Peach 1.00 1.00 1.00 302
Pear 1.00 1.00 1.00 295
Pepper 1.00 1.00 1.00 261
Pinenapple 0.96 0.95 0.95 297
Plum 1.00 1.00 1.00 266
Potato 1.00 1.00 1.00 277
Pumpkin 0.99 1.00 0.99 266
Radish 1.00 1.00 1.00 266
Strawberry 0.91 0.97 0.94 277
Tomato 0.99 1.00 1.00 270
Watermelon 0.96 0.90 0.93 284
accuracy 0.98 8960
macro avg 0.98 0.98 0.98 8960
weighted avg 0.98 0.98 0.98 8960
In [72]:
Copied!
y_pred_tr = svс.predict(X_train)
y_pred_tr = svс.predict(X_train)
In [73]:
Copied!
print(classification_report(y_train, y_pred_tr))
print(classification_report(y_train, y_pred_tr))
precision recall f1-score support
Apple 1.00 1.00 1.00 1110
Avocado 1.00 1.00 1.00 1114
Banana 1.00 1.00 1.00 1147
Bean 1.00 1.00 1.00 1129
Bitter_Gourd 1.00 1.00 1.00 1141
Bottle_Gourd 1.00 1.00 1.00 1141
Brinjal 1.00 1.00 1.00 1097
Broccoli 1.00 1.00 1.00 1106
Cabbage 1.00 1.00 1.00 1104
Capsicum 1.00 1.00 1.00 1127
Carrot 1.00 1.00 1.00 1114
Cauliflower 1.00 1.00 1.00 1101
Cherry 1.00 1.00 1.00 1127
Cucumber 1.00 1.00 1.00 1129
Grape 1.00 1.00 1.00 1127
Kiwi 1.00 1.00 1.00 1148
Mango 1.00 1.00 1.00 1104
Nut 1.00 1.00 1.00 1113
Onion 1.00 1.00 1.00 1111
Orange 1.00 1.00 1.00 1108
Papaya 1.00 1.00 1.00 1103
Peach 1.00 1.00 1.00 1098
Pear 1.00 1.00 1.00 1105
Pepper 1.00 1.00 1.00 1139
Pinenapple 1.00 1.00 1.00 1103
Plum 1.00 1.00 1.00 1134
Potato 1.00 1.00 1.00 1123
Pumpkin 1.00 1.00 1.00 1134
Radish 1.00 1.00 1.00 1134
Strawberry 1.00 1.00 1.00 1123
Tomato 1.00 1.00 1.00 1130
Watermelon 1.00 1.00 1.00 1116
accuracy 1.00 35840
macro avg 1.00 1.00 1.00 35840
weighted avg 1.00 1.00 1.00 35840
In [74]:
Copied!
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_test, y_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
plt.figure(figsize=(8, 8))
mat = confusion_matrix(y_test, y_pred)
sns.heatmap(mat.T, square=True, annot=True, fmt="d", cbar=False, xticklabels=y_name, yticklabels=y_name)
plt.xlabel("true label")
plt.ylabel("predicted label")
plt.show()
LogisticRegression
In [75]:
Copied!
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(C=0.1, max_iter=200)), # Увеличиваем max_iter
]
)
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(C=0.1, max_iter=200)), # Увеличиваем max_iter
]
)
In [76]:
Copied!
pipeline.fit(X_train, y_train)
y_pred = pipeline.predict(X_test)
pipeline.fit(X_train, y_train)
y_pred = pipeline.predict(X_test)
In [77]:
Copied!
y_pred_tr = pipeline.predict(X_train)
accuracy_score(y_train, y_pred_tr)
y_pred_tr = pipeline.predict(X_train)
accuracy_score(y_train, y_pred_tr)
Out[77]:
0.9899274553571429
In [78]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.86 0.82 0.84 284
Avocado 0.90 0.90 0.90 271
Banana 0.94 0.90 0.92 294
Bean 0.99 1.00 0.99 302
Bitter_Gourd 1.00 1.00 1.00 292
Bottle_Gourd 1.00 1.00 1.00 303
Brinjal 0.98 0.99 0.98 273
Broccoli 0.99 0.99 0.99 277
Cabbage 0.99 0.99 0.99 286
Capsicum 0.96 0.99 0.97 271
Carrot 1.00 1.00 1.00 297
Cauliflower 1.00 0.99 0.99 253
Cherry 0.86 0.86 0.86 273
Cucumber 0.99 0.98 0.99 266
Grape 1.00 1.00 1.00 273
Kiwi 0.91 0.93 0.92 252
Mango 0.83 0.83 0.83 266
Nut 1.00 1.00 1.00 290
Onion 0.99 1.00 1.00 296
Orange 0.88 0.87 0.88 287
Papaya 0.98 0.97 0.98 289
Peach 0.99 1.00 1.00 299
Pear 0.99 1.00 0.99 259
Pepper 0.98 1.00 0.99 296
Pinenapple 0.93 0.96 0.94 270
Plum 1.00 1.00 1.00 295
Potato 0.98 0.99 0.99 261
Pumpkin 0.98 0.98 0.98 286
Radish 0.99 0.99 0.99 297
Strawberry 0.91 0.91 0.91 266
Tomato 0.99 0.96 0.97 277
Watermelon 0.87 0.90 0.88 259
accuracy 0.96 8960
macro avg 0.96 0.96 0.96 8960
weighted avg 0.96 0.96 0.96 8960
SVC: accuracy на трейне = 1.0, accuracy на тесте = 0.98 при С = 12, kernel = "rbf". LogisticRegression: при с = 0.1 accuracy на трейне = 0.99, accuracy на тесте = 0.96
Обучение модели на полном датасете цветных фотографий размером 128x128 px¶
In [79]:
Copied!
creat_temp_dataset(use_sample=False)
creat_temp_dataset(use_sample=False)
0%| | 0/32 [00:00<?, ?it/s]
In [80]:
Copied!
resize_temp_dataset((128, 128), color_background="white")
resize_temp_dataset((128, 128), color_background="white")
0%| | 0/32 [00:00<?, ?it/s]
In [81]:
Copied!
X, y = get_X_and_y(size=(128, 128), color="color")
X, y = get_X_and_y(size=(128, 128), color="color")
In [82]:
Copied!
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
In [83]:
Copied!
svс = SVC(C=10, kernel="rbf")
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
svс = SVC(C=10, kernel="rbf")
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
In [84]:
Copied!
y_pred_tr = svс.predict(X_train)
y_pred_tr = svс.predict(X_train)
In [85]:
Copied!
accuracy_score(y_train, y_pred_tr)
accuracy_score(y_train, y_pred_tr)
Out[85]:
0.9996651785714286
In [86]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.89 0.91 0.90 290
Avocado 0.89 0.91 0.90 286
Banana 0.87 0.92 0.89 253
Bean 0.99 1.00 0.99 271
Bitter_Gourd 1.00 1.00 1.00 259
Bottle_Gourd 1.00 1.00 1.00 259
Brinjal 0.97 0.98 0.98 303
Broccoli 1.00 0.99 0.99 294
Cabbage 0.99 1.00 0.99 296
Capsicum 0.97 1.00 0.98 273
Carrot 0.99 0.99 0.99 286
Cauliflower 0.99 0.99 0.99 299
Cherry 0.87 0.92 0.89 273
Cucumber 1.00 1.00 1.00 271
Grape 1.00 1.00 1.00 273
Kiwi 0.93 0.90 0.92 252
Mango 0.89 0.83 0.86 296
Nut 1.00 1.00 1.00 287
Onion 1.00 1.00 1.00 289
Orange 0.92 0.89 0.91 292
Papaya 0.99 0.99 0.99 297
Peach 1.00 1.00 1.00 302
Pear 1.00 1.00 1.00 295
Pepper 1.00 1.00 1.00 261
Pinenapple 0.95 0.92 0.93 297
Plum 1.00 1.00 1.00 266
Potato 0.99 1.00 0.99 277
Pumpkin 0.99 0.99 0.99 266
Radish 0.99 1.00 1.00 266
Strawberry 0.91 0.90 0.91 277
Tomato 1.00 0.99 0.99 270
Watermelon 0.91 0.87 0.89 284
accuracy 0.96 8960
macro avg 0.96 0.97 0.96 8960
weighted avg 0.97 0.96 0.96 8960
LogisticRegression
In [87]:
Copied!
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(C=0.1, max_iter=200)), # Увеличиваем max_iter
]
)
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(C=0.1, max_iter=200)), # Увеличиваем max_iter
]
)
In [88]:
Copied!
pipeline.fit(X_train, y_train)
y_pred = pipeline.predict(X_test)
pipeline.fit(X_train, y_train)
y_pred = pipeline.predict(X_test)
In [89]:
Copied!
y_pred_tr = pipeline.predict(X_train)
accuracy_score(y_train, y_pred_tr)
y_pred_tr = pipeline.predict(X_train)
accuracy_score(y_train, y_pred_tr)
Out[89]:
0.9740513392857143
In [90]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.82 0.77 0.79 284
Avocado 0.80 0.81 0.81 271
Banana 0.84 0.84 0.84 294
Bean 0.98 0.99 0.98 302
Bitter_Gourd 0.99 0.99 0.99 292
Bottle_Gourd 0.97 0.98 0.98 303
Brinjal 0.96 0.96 0.96 273
Broccoli 0.98 0.99 0.98 277
Cabbage 0.99 0.97 0.98 286
Capsicum 0.96 0.98 0.97 271
Carrot 0.99 0.99 0.99 297
Cauliflower 0.99 0.99 0.99 253
Cherry 0.83 0.80 0.81 273
Cucumber 0.97 0.95 0.96 266
Grape 0.99 1.00 0.99 273
Kiwi 0.84 0.87 0.85 252
Mango 0.77 0.74 0.76 266
Nut 1.00 1.00 1.00 290
Onion 0.99 1.00 0.99 296
Orange 0.81 0.81 0.81 287
Papaya 0.95 0.95 0.95 289
Peach 0.99 1.00 0.99 299
Pear 0.98 1.00 0.99 259
Pepper 1.00 1.00 1.00 296
Pinenapple 0.87 0.90 0.89 270
Plum 0.99 1.00 0.99 295
Potato 0.98 0.97 0.97 261
Pumpkin 0.98 0.97 0.97 286
Radish 0.99 0.99 0.99 297
Strawberry 0.79 0.80 0.79 266
Tomato 0.94 0.95 0.95 277
Watermelon 0.79 0.80 0.80 259
accuracy 0.93 8960
macro avg 0.93 0.93 0.93 8960
weighted avg 0.93 0.93 0.93 8960
SVC: accuracy на трейне = 1.0, accuracy на тесте = 0.96 при С = 10, kernel = "rbf". LogisticRegression: при с = 0.1 accuracy на трейне = 0.97, accuracy на тесте = 0.93
Обучение модели на полном датасете цветных фотографий размером 64x64 px¶
In [91]:
Copied!
creat_temp_dataset(use_sample=False)
creat_temp_dataset(use_sample=False)
0%| | 0/32 [00:00<?, ?it/s]
In [92]:
Copied!
resize_temp_dataset((64, 64), color_background="white")
resize_temp_dataset((64, 64), color_background="white")
0%| | 0/32 [00:00<?, ?it/s]
In [93]:
Copied!
X, y = get_X_and_y(size=(64, 64), color="color")
X, y = get_X_and_y(size=(64, 64), color="color")
In [94]:
Copied!
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
In [95]:
Copied!
svс = SVC(C=18, kernel="rbf")
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
svс = SVC(C=18, kernel="rbf")
svс.fit(X_train, y_train)
y_pred = svс.predict(X_test)
In [96]:
Copied!
y_pred_tr = svс.predict(X_train)
accuracy_score(y_train, y_pred_tr)
y_pred_tr = svс.predict(X_train)
accuracy_score(y_train, y_pred_tr)
Out[96]:
0.9974609375
In [97]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.86 0.83 0.84 284
Avocado 0.81 0.84 0.82 271
Banana 0.85 0.89 0.87 294
Bean 0.97 0.99 0.98 302
Bitter_Gourd 0.98 0.98 0.98 292
Bottle_Gourd 0.99 1.00 0.99 303
Brinjal 0.95 0.97 0.96 273
Broccoli 0.97 0.97 0.97 277
Cabbage 0.97 0.94 0.96 286
Capsicum 0.94 0.98 0.96 271
Carrot 1.00 0.98 0.99 297
Cauliflower 0.98 0.97 0.98 253
Cherry 0.84 0.83 0.84 273
Cucumber 0.99 0.99 0.99 266
Grape 1.00 1.00 1.00 273
Kiwi 0.85 0.86 0.86 252
Mango 0.84 0.81 0.82 266
Nut 1.00 1.00 1.00 290
Onion 1.00 1.00 1.00 296
Orange 0.85 0.82 0.83 287
Papaya 0.99 0.97 0.98 289
Peach 1.00 1.00 1.00 299
Pear 1.00 1.00 1.00 259
Pepper 1.00 1.00 1.00 296
Pinenapple 0.86 0.88 0.87 270
Plum 1.00 1.00 1.00 295
Potato 0.99 1.00 0.99 261
Pumpkin 0.95 0.97 0.96 286
Radish 0.99 1.00 0.99 297
Strawberry 0.85 0.86 0.85 266
Tomato 0.96 0.96 0.96 277
Watermelon 0.87 0.80 0.83 259
accuracy 0.94 8960
macro avg 0.94 0.94 0.94 8960
weighted avg 0.94 0.94 0.94 8960
LogisticRegression
In [98]:
Copied!
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(C=0.1, max_iter=200)), # Увеличиваем max_iter
]
)
pipeline = Pipeline(
[
("scaler", StandardScaler()), # Масштабируем данные
("logreg", LogisticRegression(C=0.1, max_iter=200)), # Увеличиваем max_iter
]
)
In [99]:
Copied!
pipeline.fit(X_train, y_train)
y_pred = pipeline.predict(X_test)
pipeline.fit(X_train, y_train)
y_pred = pipeline.predict(X_test)
In [100]:
Copied!
y_pred_tr = pipeline.predict(X_train)
accuracy_score(y_train, y_pred_tr)
y_pred_tr = pipeline.predict(X_train)
accuracy_score(y_train, y_pred_tr)
Out[100]:
0.9469308035714286
In [101]:
Copied!
print(classification_report(y_test, y_pred))
print(classification_report(y_test, y_pred))
precision recall f1-score support
Apple 0.71 0.64 0.67 284
Avocado 0.76 0.72 0.74 271
Banana 0.76 0.79 0.78 294
Bean 0.94 0.95 0.94 302
Bitter_Gourd 0.94 0.97 0.95 292
Bottle_Gourd 0.96 0.98 0.97 303
Brinjal 0.90 0.90 0.90 273
Broccoli 0.96 0.94 0.95 277
Cabbage 0.91 0.89 0.90 286
Capsicum 0.90 0.94 0.92 271
Carrot 0.98 0.96 0.97 297
Cauliflower 0.95 0.96 0.95 253
Cherry 0.75 0.76 0.75 273
Cucumber 0.93 0.91 0.92 266
Grape 0.99 0.99 0.99 273
Kiwi 0.78 0.77 0.77 252
Mango 0.68 0.67 0.68 266
Nut 0.98 1.00 0.99 290
Onion 1.00 1.00 1.00 296
Orange 0.74 0.72 0.73 287
Papaya 0.90 0.91 0.91 289
Peach 0.99 0.99 0.99 299
Pear 0.95 0.98 0.96 259
Pepper 0.98 0.99 0.99 296
Pinenapple 0.73 0.82 0.77 270
Plum 1.00 1.00 1.00 295
Potato 0.95 0.93 0.94 261
Pumpkin 0.94 0.87 0.90 286
Radish 0.97 0.98 0.97 297
Strawberry 0.70 0.73 0.72 266
Tomato 0.90 0.91 0.91 277
Watermelon 0.77 0.73 0.75 259
accuracy 0.89 8960
macro avg 0.88 0.88 0.88 8960
weighted avg 0.89 0.89 0.89 8960
Итоги¶
Для нахождения лучших гиперпараметров использовался GridSearchCV. Для оценки производительности модели применялась метрика accuracy, так как даёт хорошие результаты на сбалансированных классах.**
Результаты обучения с помощью ResNet18 + SVC/LogisticRegression для выборки из 500 изображений каждого класса
| Модель | Гиперпараметры | Размер изображения | Цветное | accuracy на трейне | accuracy на test |
|---|---|---|---|---|---|
| SVC | C=50, kernel='rbf' | 224px | да | 0.96 | 0.97 |
| SVC | C=40, kernel='rbf' | 224px | да | 0.95 | 0.97 |
| SVC | C=12, kernel='rbf' | 224px | да | 0.96 | 0.97 |
| SVC | C=12, kernel='rbf' | 224px | нет | 0.96 | 0.70 |
| SVC | C=10, kernel='rbf' | 128px | да | 0.94 | 0.95 |
| SVC | C=14, kernel='rbf' | 64px | да | 0.89 | 0.90 |
| SVC | C=18, kernel='rbf' | 64px | да | 0.89 | 0.90 |
| LogisticRegression | C=0.1 | 224px | да | 0.94 | 0.94 |
| LogisticRegression | C=0.1 | 128px | да | 0.90 | 0.91 |
| LogisticRegression | C=0.1 | 64px | да | 0.85 | 0.86 |
Выводы:
- Как мы видим, при разных значениях гиперпараметра С модель выдает схожие результаты метрик. Это может быть связано с тем, что признаки, извлеченные с помощью ResNet18, могут быть очень информативными и хорошо разделяющими классы, что делает модель SVC менее чувствительной к параметрам C. Если данные хорошо разделены, модель будет давать схожие результаты при различных значениях параметров.
- Модель, обученная на черно-белых изображениях, показала accuracy на трейне, сравнимое со значением модели, обученной на цветных изображениях, но на test-выборке показатели сильно упали, что может говорить о переобучении. Поэтому модель будет обучаться на полном датасете с цветными фотографиями.
Результаты обучения c помощью ResNet18 + SVC на полном датасете:
| Модель | Гиперпараметры | Размер изображения | Цветное | accuracy на трейне | accuracy на test | Время извлечения признаков датасета | Время обучения модели |
|---|---|---|---|---|---|---|---|
| SVC | C=12, kernel='rbf' | 224px | да | 1.0 | 0.98 | ≈1 ч | ≈2 мин |
| SVC | C=10, kernel='rbf' | 128px | да | 1.0 | 0.96 | ≈0.5 ч | ≈3 мин |
| SVC | C=18, kernel='rbf' | 64px | да | 1.0 | 0.94 | ≈0.25 ч | ≈5 мин |
| LogisticRegression | C=0.1 | 224px | да | 0.99 | 0.96 | ≈1 ч | ≈0.5 мин |
| LogisticRegression | C=0.1 | 128px | да | 0.97 | 0.93 | ≈0.5 ч | ≈0.67 мин |
| LogisticRegression | C=0.1 | 64px | да | 0.95 | 0.89 | ≈0.25 ч | ≈1 мин |
Итог: модели, основанные на SVC, продемонстрировали наивысшую точность на тестовой выборке, достигая 0.98 при размере изображения 224px. Logistic Regression немного уступила по точности, показав максимальный результат 0.96 на тесте при тех же параметрах. С уменьшением размера изображения точность моделей падает, особенно заметно на тестовой выборке. Для SVC: точность снизилась с 0.98 (224px) до 0.94 (64px). Для Logistic Regression: точность снизилась с 0.96 (224px) до 0.89 (64px). Большие изображения (224px) лучше подходят для извлечения информативных признаков, обеспечивая более высокую точность. Увеличение разрешения изображений существенно повышает время, необходимое для извлечения признаков и обучения модели. Например, обработка изображений размером 224px требует около часа, тогда как для 64px — лишь 15 минут. Таким образом, выбор оптимального размера изображения зависит от доступных ресурсов и требований к скорости обработки.