Tensorflow `from_tensor_slices` 卡死

• Z时代
• 2024-01-10
• 分类：技术分享

警告memory exceeded 10% of system memory
我load本地的tfrecord成np array

def CUB_load_data():
    ds = tfds.load('caltech_birds2011', download=False, data_dir='../../datasets/')
    train_data = ds['train']
    test_data = ds['test']
    train_x = []
    train_y = []
    test_x = []
    test_y = []
    for i in train_data.__iter__():
        resized = cv2.resize(i['image'].numpy(), dsize=(224,224))
        train_x.append(resized)
        train_y.append(i['label'])
    for i in test_data.__iter__():
        resized = cv2.resize(i['image'].numpy(), dsize=(224,224))
        test_x.append(resized)
        test_y.append(i['label'])
    return (train_x, train_y), (test_x, test_y)

这部分应该没问题

我用from_tensor_slices去创建tf.dataset，我尝试过修改batch，但一样没效果，卡死并且报超过系统10%容量的警告。
图片是CUB200_2011，所有图片也就1.G上下。用tfds.load生成的tfrecord

def load_data():
    (train_x, train_y), (test_x, test_y) =  CUB_load_data()
    SHUFFLE_BUFFER_SIZE = 500
    BATCH_SIZE = 2
    @tf.function
    def _parse_function(img, label):
        feature = {}
        img = tf.cast(img, dtype=tf.float32)
        img = img / 255.0
        feature["img"] = img
        feature["label"] = label
        return feature
    train_dataset_raw = tf.data.Dataset.from_tensor_slices(
        (train_x, train_y)).map(_parse_function)
    test_dataset_raw = tf.data.Dataset.from_tensor_slices(
        (test_x, test_y)).map(_parse_function)
    train_dataset = train_dataset_raw.shuffle(SHUFFLE_BUFFER_SIZE).batch(
        BATCH_SIZE)
    test_dataset = test_dataset_raw.shuffle(SHUFFLE_BUFFER_SIZE).batch(
        BATCH_SIZE)
    return train_dataset, test_dataset

以上是 Tensorflow `from_tensor_slices` 卡死 的全部内容， 来源链接： utcz.com/a/157388.html