pip install tensorflow opencv-python matplotlib

import tensorflow as tf
import os
import cv2
import numpy as np

# Загрузите и разархивируйте набор данных UCF101.
url = "https://www.crcv.ucf.edu/data/UCF101/UCF101.rar"
data_dir = tf.keras.utils.get_file('UCF101', origin=url, untar=True)

# Определить функции загрузки и предварительной обработки видео.
def load_video(path, max_frames=0, resize=(224, 224)):
    cap = cv2.VideoCapture(path)
    frames = []
    try:
        while True:
            ret, frame = cap.read()
            if not ret:
                break
            frame = cv2.resize(frame, resize)
            frame = frame[:, :, [2, 1, 0]]  # BGR to RGB
            frames.append(frame)
            if max_frames and len(frames) == max_frames:
                break
    finally:
        cap.release()
    return np.array(frames)

# Пример: загрузить видео
video_path = os.path.join(data_dir, 'ApplyEyeMakeup/v_ApplyEyeMakeup_g01_c01.avi')
video = load_video(video_path)
print(f"Video shape: {video.shape}")

from tensorflow.keras.applications import InceptionV3
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, GlobalAveragePooling2D, TimeDistributed, LSTM

# Загрузите предварительно обученную модель InceptionV3.
base_model = InceptionV3(weights='imagenet', include_top=False)

# Постройте модель распознавания действий
model = Sequential([
    TimeDistributed(base_model, input_shape=(None, 224, 224, 3)),
    TimeDistributed(GlobalAveragePooling2D()),
    LSTM(64),
    Dense(101, activation='softmax')  # UCF101 имеет 101 категорию действий.
])

# Скомпилировать модель
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

# Посмотреть структуру модели
model.summary()

from tensorflow.keras.utils import Sequence

class VideoDataGenerator(Sequence):
    def __init__(self, video_paths, labels, batch_size=8, max_frames=40, resize=(224, 224)):
        self.video_paths = video_paths
        self.labels = labels
        self.batch_size = batch_size
        self.max_frames = max_frames
        self.resize = resize

    def __len__(self):
        return len(self.video_paths) // self.batch_size

    def __getitem__(self, idx):
        batch_x = self.video_paths[idx * self.batch_size:(idx + 1) * self.batch_size]
        batch_y = self.labels[idx * self.batch_size:(idx + 1) * self.batch_size]
        return np.array([load_video(path, self.max_frames, self.resize) for path in batch_x]), np.array(batch_y)

# Пример: создание генератора данных
video_paths = [os.path.join(data_dir, 'ApplyEyeMakeup/v_ApplyEyeMakeup_g01_c01.avi')] * 10
labels = [0] * 10  # Предположим, что метки всех видео равны 0.
train_generator = VideoDataGenerator(video_paths, labels)

# Модель обучения
model.fit(train_generator, epochs=10)

# Пример: оценка модели
test_video_paths = [os.path.join(data_dir, 'ApplyEyeMakeup/v_ApplyEyeMakeup_g01_c01.avi')] * 2
test_labels = [0] * 2
test_generator = VideoDataGenerator(test_video_paths, test_labels)

loss, accuracy = model.evaluate(test_generator)
print(f"Test accuracy: {accuracy * 100:.2f}%")

import matplotlib.pyplot as plt

# Визуализируйте видеокадры
def visualize_video(video):
    fig, axes = plt.subplots(1, len(video), figsize=(20, 5))
    for i, frame in enumerate(video):
        axes[i].imshow(frame)
        axes[i].axis('off')
    plt.show()

# Пример：Визуализируйте видеокадры
visualize_video(video[:5])