我有两个不同的数据集可以馈入张量流中的模型。因此，我发现可馈送的迭代器是一个不错的选择。这是第一个数据集的形状：
fer +：
图像：类型：浮动/暗淡：64 x 64（这些是灰度图像）
标签：类型：浮动/暗淡：（？，8）
sewa：
图像：uint8 /暗（？，64、64）
名称：字符串（？，）
化合价：浮动/暗淡：（？，1）
arousal：float / dim：（？，1）
喜欢：浮动/暗淡：（？，1）
检测到：int（这将告诉我是否在一帧中检测到了脸部）。
代码如下：
def _parse_fer_example(example_proto):
features = {
'height': tf.FixedLenFeature([], tf.int64),
'width': tf.FixedLenFeature([], tf.int64),
'image_raw': tf.FixedLenFeature([], tf.string),
'labels': tf.FixedLenFeature([8], tf.float32)
}
parsed_features = tf.parse_single_example(example_proto, features)
# This is how we create one example, that is, extract one example from the database.
image = tf.decode_raw(parsed_features['image_raw'], tf.float32)
# The height and the weights are used to
height = tf.cast(parsed_features['height'], tf.int32)
width = tf.cast(parsed_features['width'], tf.int32)
image = tf.reshape(image, [64, 64])
labels = parsed_features['labels']
return image, labels
def _parse_sewa_example(example_proto):
# The annotation contains the following features: timestamp; arousal; valence; liking
features = {
'height': tf.FixedLenFeature([], tf.int64),
'width': tf.FixedLenFeature([], tf.int64),
'image_raw': tf.FixedLenFeature([], tf.string),
'name': tf.FixedLenFeature([], tf.string),
'frame_number': tf.FixedLenFeature([1], tf.int64),
'time': tf.FixedLenFeature([1], tf.float32),
'detected': tf.FixedLenFeature([1], tf.int64),
'arousal': tf.FixedLenFeature([1], tf.float32),
'valence': tf.FixedLenFeature([1], tf.float32),
'liking': tf.FixedLenFeature([1], tf.float32),
'istalking': tf.FixedLenFeature([1], tf.int64)
}
parsed_features = tf.parse_single_example(example_proto, features)
# This is how we create one example, that is, extract one example from the database.
image = tf.decode_raw(parsed_features['image_raw'], tf.uint8)
# The height and the weights are used to
height = tf.cast(parsed_features['height'], tf.int32)
width = tf.cast(parsed_features['width'], tf.int32)
# The image is reshaped since when stored as a binary format, it is flattened. Therefore, we need the
# height and the weight to restore the original image back.
# Tensor("Reshape:0", shape=(112, 112, 3), dtype=uint8)
image = tf.reshape(image, [112, 112, 3])
name = parsed_features['name']
istalking = parsed_features['istalking']
detected = parsed_features['detected']
arousal = parsed_features['arousal']
valence = parsed_features['valence']
liking = parsed_features['liking']
return name, detected, arousal, valence, liking, istalking, image
# This will load fer dataset.
def load_fer_data(file_name, batch_size):
dataset = tf.data.TFRecordDataset(file_name).map(_parse_fer_example).shuffle(15000).batch(batch_size)
iterator = dataset.make_initializable_iterator(shared_name='fer_iterator')
images, labels = iterator.get_next()
print(images, labels)
return images, labels, iterator
def load_sewa_data(file_name, batch_size):
dataset = tf.data.TFRecordDataset(file_name).map(_parse_sewa_example).batch(batch_size)
iterator = dataset.make_initializable_iterator(shared_name='sewa_iterator')
with tf.name_scope('sewa_next_batch'):
next_batch = iterator.get_next()
names, detected, arousal, valence, liking, istalkings, images = next_batch
print(names, detected, arousal, valence, liking, istalkings, images)
return names, detected, arousal, valence, liking, istalkings, images, iterator
if __name__ == '__main__':
tf_records = os.listdir(sewa_tfrecords_path)
sewa_train_files = [sewa_tfrecords_path + "/" + f for f in tf_records if 'Train' in f]
sewa_devel_files = [sewa_tfrecords_path + "/" + f for f in tf_records if 'Devel' in f]
print(sewa_train_files)
filenames = tf.placeholder(tf.string, shape=[None])
names, detected, arousal, valence, liking, istalkings, images_sewa, iterator_sewa = load_sewa_data(filenames, 34)
images_fer, labels, iterator_fer = load_fer_data(filenames, 64)
with tf.Session() as sess:
sess.run(iterator_sewa.initializer, feed_dict={filenames: sewa_train_files})
total = 0
# ....
打印后，这就是我得到的：
Tensor("sewa_next_batch/IteratorGetNext:0", shape=(?,), dtype=string) Tensor("sewa_next_batch/IteratorGetNext:1", shape=(?, 1), dtype=int64) Tensor("sewa_next_batch/IteratorGetNext:2", shape=(?, 1), dtype=float32) Tensor("sewa_next_batch/IteratorGetNext:3", shape=(?, 1), dtype=float32) Tensor("sewa_next_batch/IteratorGetNext:4", shape=(?, 1), dtype=float32) Tensor("sewa_next_batch/IteratorGetNext:5", shape=(?, 1), dtype=int64) Tensor("sewa_next_batch/IteratorGetNext:6", shape=(?, 112, 112, 3), dtype=uint8)
AND：
Tensor("IteratorGetNext:0", shape=(?, 64, 64), dtype=float32) Tensor("IteratorGetNext:1", shape=(?, 8), dtype=float32)
这样做的原因是：
我想在fer数据集上训练模型；然后删除分类器，为sewa数据库创建一个不同的分类器；然后再次训练整个模型。
因此，当在fer数据集上训练模型时，我将不得不在最后进行分类，那里我们有8个输出，而在sewa数据库上训练时，我将有2个输出。我需要一个不同的分类器。
因此，现在，我想使用类似于feedable iterator的方式来合并两个数据库。但是我不确定如何。
或者，如果我们看一下https://www.tensorflow.org/guide/datasets上的文档，则会提到：
# A feedable iterator is defined by a handle placeholder and its structure. We
# could use the `output_types` and `output_shapes` properties of either
# `training_dataset` or `validation_dataset` here, because they have
# identical structure.
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(
handle, training_dataset.output_types, training_dataset.output_shapes)
next_element = iterator.get_next()
因此，如果数据集具有不同的结构，该如何处理？