超大型数据集可能不适合分配给进程的内存。在前面的步骤中,我们设置了一个流水线,用于将整个数据集引入内存、准备数据,并将工作集传递给训练函数。而是会提供一个替代训练函数 (fit_generator
),用于批量拉取数据。这样,我们就可以仅对数据中的一小部分(batch_size
的倍数)应用数据流水线中的转换。在实验期间,我们针对 DBPedia、Amazon 评价、Ag News 和 Yelp 评价等数据集使用了批处理(GitHub 中的代码)。
以下代码说明了如何生成数据批次并将其提供给 fit_generator
。
def _data_generator(x, y, num_features, batch_size): """Generates batches of vectorized texts for training/validation. # Arguments x: np.matrix, feature matrix. y: np.ndarray, labels. num_features: int, number of features. batch_size: int, number of samples per batch. # Returns Yields feature and label data in batches. """ num_samples = x.shape[0] num_batches = num_samples // batch_size if num_samples % batch_size: num_batches += 1 while 1: for i in range(num_batches): start_idx = i * batch_size end_idx = (i + 1) * batch_size if end_idx > num_samples: end_idx = num_samples x_batch = x[start_idx:end_idx] y_batch = y[start_idx:end_idx] yield x_batch, y_batch # Create training and validation generators. training_generator = _data_generator( x_train, train_labels, num_features, batch_size) validation_generator = _data_generator( x_val, val_labels, num_features, batch_size) # Get number of training steps. This indicated the number of steps it takes # to cover all samples in one epoch. steps_per_epoch = x_train.shape[0] // batch_size if x_train.shape[0] % batch_size: steps_per_epoch += 1 # Get number of validation steps. validation_steps = x_val.shape[0] // batch_size if x_val.shape[0] % batch_size: validation_steps += 1 # Train and validate model. history = model.fit_generator( generator=training_generator, steps_per_epoch=steps_per_epoch, validation_data=validation_generator, validation_steps=validation_steps, callbacks=callbacks, epochs=epochs, verbose=2) # Logs once per epoch.