AestheticPredictor/tensorflow/data.py

import keras
import numpy as np
import os
import tensorflow as tf
import tensorflow_hub as hub

from PIL import Image


class EfficientNetFeatureExtractor:
    def __init__(self):
        self.model = hub.KerasLayer("https://tfhub.dev/google/imagenet/efficientnet_v2_imagenet1k_m/feature_vector/2", trainable=False)
        self.preprocess = tf.keras.applications.efficientnet.preprocess_input
        self.feature_dims = 1280

    def extract(self, image_path):
        image = Image.open(image_path).convert('RGB')
        image = image.resize((480, 480))  # Resize the image to match EfficientNet's input size
        image = tf.keras.preprocessing.image.img_to_array(image)
        image = self.preprocess(image)
        image = tf.expand_dims(image, axis=0)
        embedding = self.model(image)[0]
        return embedding


class LAIONAestheticsDataGenerator(keras.utils.Sequence):
    def __init__(self, img_files, scores, img_dir, feature_extractor: EfficientNetFeatureExtractor, batch_size=32, shuffle=True):
        self.feature_extractor = feature_extractor
        self.img_path = img_dir
        self.img_files = img_files
        self.scores = scores
        self.embeddings = {}
        self.batch_size = batch_size
        self.shuffle = shuffle
        self.on_epoch_end()

    def __len__(self):
        return len(self.img_files)

    def __getitem__(self, idx):
        if tf.is_tensor(idx):
            idx = idx.numpy()

        start = idx * self.batch_size
        stop = start + self.batch_size
        batch_idxs = self.indexes[start:stop]

        embeddings, scores = self.__data_generation(batch_idxs)

        return embeddings, scores
    
    def __data_generation(self, idxs):
        embeddings = np.empty((self.batch_size, self.feature_extractor.feature_dims))
        scores = np.empty((self.batch_size))

        # Generate data
        for i, idx in enumerate(idxs):
            # Store sample
            embedding = self.embeddings.get(idx, None)
            if embedding is None:
                img_path = os.path.join(self.img_path, self.img_files[idx])
                embedding = self.feature_extractor.extract(img_path)
                self.embeddings[idx] = embedding

            embeddings[i,] = embedding
            scores[i] = self.scores[idx]

        return embeddings, scores
    
    def on_epoch_end(self):
        'Updates indexes after each epoch'
        self.indexes = np.arange(len(self.img_files))
        if self.shuffle == True:
            np.random.shuffle(self.indexes)


def train_valid_split(data_dir, train_percent=0.8, limit=None, batch_size=32):
    annotations_file = os.path.join(data_dir, 'labels.tsv')
    data = []

    with open(annotations_file) as f:
        for i, row in enumerate(f.readlines()):
            if limit is not None and i >= limit:
                break
            img_name, _, aesthetic_score = row.split('\t')[:3]
            data.append((img_name, tf.constant(float(aesthetic_score))))

    np.random.shuffle(data)

    feature_extractor = EfficientNetFeatureExtractor()

    train_size = int(train_percent * len(data))
    train_data = data[:train_size]
    valid_data = data[train_size:]

    train_imgs, train_scores = zip(*train_data)
    valid_imgs, valid_scores = zip(*valid_data)

    train_generator = LAIONAestheticsDataGenerator(train_imgs, 
                                                   train_scores, 
                                                   data_dir, 
                                                   feature_extractor, 
                                                   batch_size, 
                                                   shuffle=True)
    valid_generator = LAIONAestheticsDataGenerator(valid_imgs, 
                                                   valid_scores, 
                                                   data_dir, 
                                                   feature_extractor, 
                                                   batch_size, 
                                                   shuffle=False)

    return train_generator, valid_generator