如何使用 tf.data.Dataset api 训练 keras.concatenate 模型？

Question

这是我的 keras 模型”

input1 = keras.layers.Input(shape=(100,))
...
net1 = keras.layers.Dense(n1, activation='relu')()


input2 = keras.layers.Input(shape=(50,))
...
net2 = keras.layers.Dense(n2, activation='relu')()


merge = keras.layers.concatenate([net1,net2])
output = keras.layers.Dense(num_classes, activation='softmax')(merge)
model = keras.Model(inputs=[input1, input1], outputs=[output])

和tfrecord数据（大数据）：

feature_description = {
    'f1': tf.io.FixedLenFeature([100], tf.int64),
    'f2': tf.io.FixedLenFeature([50], tf.int64),
    'label': tf.io.FixedLenFeature([], tf.int64),
}

def parser(example_proto):
  return tf.io.parse_single_example(example_proto, feature_description)

ilename_queue = tf.data.Dataset.list_files(filename_queue, shuffle=True)
dataset = tf.data.TFRecordDataset(filename_queue).map(parser)

如何将数据拟合到模型中？

我知道你注定要这样做：

model.fit([f1,f2], y, epochs=epochs, batch_size=batch_size,validation_split=0.3)

而 f1,f2 是 df.dataframe。

Answer 1

这里是我写的 sn-p，用于创建虚拟数据集并将其写入 tfrecord 文件并构建模型。

import tensorflow as tf

num_classes = 10
n_samples = 10000

f1 = tf.random.uniform(shape=[n_samples, 100], maxval=500, dtype=tf.int32).numpy()
f2 = tf.random.uniform(shape=[n_samples, 50], maxval=500, dtype=tf.int32).numpy()
labels = tf.random.uniform(shape=[n_samples], maxval=num_classes, dtype=tf.int32).numpy()



def make_example(f1, f2, label):
    feature = {
        'f1': tf.train.Feature(int64_list=tf.train.Int64List(value=f1)),
        'f2': tf.train.Feature(int64_list=tf.train.Int64List(value=f2)),
        'label': tf.train.Feature(int64_list=tf.train.Int64List(value=[label])),
    }
    return tf.train.Example(features=tf.train.Features(feature=feature))

def write_tfrecord(f1, f2, labels, tfrecord_path):
    with tf.io.TFRecordWriter(tfrecord_path) as writer:
        for i in range(len(f1)):
            example = make_example(f1[i], f2[i], labels[i])
            writer.write(example.SerializeToString())

write_tfrecord(f1, f2, labels, 'test.tfrecord')

n1 = 16
n2 = 16
input1 = tf.keras.layers.Input(shape=(100,))
input2 = tf.keras.layers.Input(shape=(50,))

net1 = tf.keras.layers.Dense(n1, activation='relu')(input1)
net2 = tf.keras.layers.Dense(n2, activation='relu')(input2)

merge = tf.keras.layers.concatenate([net1, net2])
output = tf.keras.layers.Dense(num_classes, activation='softmax')(merge)
model = tf.keras.Model(inputs=[input1, input2], outputs=[output])

现在解析您的 tfrecord 文件并创建一个 tf.data.Dataset 对象应该很简单。 由于您的模型有两个输入和一个输出，因此您的 tf.data.Dataset 应该具有匹配的结构。所以我就是这样做的

feature_description = {
    'f1': tf.io.FixedLenFeature([100], tf.int64),
    'f2': tf.io.FixedLenFeature([50], tf.int64),
    'label': tf.io.FixedLenFeature([], tf.int64),
}

def parser(example_proto):
    parsed_example = tf.io.parse_single_example(example_proto, feature_description)
    f1 = parsed_example['f1']
    f2 = parsed_example['f2']
    label = parsed_example['label']
    return (f1, f2), label

dataset = tf.data.TFRecordDataset('test.tfrecord')
dataset = dataset.map(parser, num_parallel_calls=tf.data.experimental.AUTOTUNE)
dataset = dataset.batch(4)
dataset = dataset.shuffle(16)
dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)

在打印dataset 的结构时，您应该看到以下输出

<PrefetchDataset shapes: (((None, 100), (None, 50)), (None,)), types: ((tf.int64, tf.int64), tf.int64)>

确保所有这些工作正常并且训练循环运行没有任何错误

model.summary()
model.compile(loss='sparse_categorical_crossentropy', optimizer='sgd')
model.fit(dataset, steps_per_epoch=10, epochs=2)

这是最终输出

Model: "model_2"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_7 (InputLayer)            [(None, 100)]        0                                            
__________________________________________________________________________________________________
input_8 (InputLayer)            [(None, 50)]         0                                            
__________________________________________________________________________________________________
dense_6 (Dense)                 (None, 16)           1616        input_7[0][0]                    
__________________________________________________________________________________________________
dense_7 (Dense)                 (None, 16)           816         input_8[0][0]                    
__________________________________________________________________________________________________
concatenate_2 (Concatenate)     (None, 32)           0           dense_6[0][0]                    
                                                                 dense_7[0][0]                    
__________________________________________________________________________________________________
dense_8 (Dense)                 (None, 10)           330         concatenate_2[0][0]              
==================================================================================================
Total params: 2,762
Trainable params: 2,762
Non-trainable params: 0
__________________________________________________________________________________________________
Train for 10 steps
Epoch 1/2
10/10 [==============================] - 0s 12ms/step - loss: 4735.8942
Epoch 2/2
10/10 [==============================] - 0s 1ms/step - loss: 2.7339