在第一次尝试解决问题时，我将代码更改为首先组合通道以构建特征列，然后制作特征列的序列。这将输入的形状从 [batch_size, channel_num, sequence_length] 更改为 [batch_size, sequence_length, channel_num] 并按照模型的预期为标签添加了维度。这是新代码：
import tensorflow as tf
import numpy as np
BATCH_SIZE = 3
SEQ_LENGTH = 4
NUM_CLASSES = 2
LSTM_UNITS = 64
NUM_SHARDS = 4
NUM_CHANNELS = 2
tf.enable_eager_execution()
def parse_values(f1, f2, label):
features = [f1,f2]
return features, label
def keras_model():
inputs = tf.keras.layers.Input(shape=(SEQ_LENGTH,NUM_CHANNELS))
x = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(LSTM_UNITS, return_sequences=True))(inputs)
outputs = tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(NUM_CLASSES, activation='relu'))(x)
return tf.keras.Model(inputs, outputs)
dataset = tf.data.experimental.CsvDataset(filenames='../../input/aFile.csv', header=True,record_defaults=[tf.int64] * 3, select_cols=[0,1,2])
dataset= dataset.map(parse_values).window(size=SEQ_LENGTH, shift=1, drop_remainder=True).flat_map(lambda features, label:
tf.data.Dataset.zip((features.batch(SEQ_LENGTH), label.batch(SEQ_LENGTH))))
dataset = dataset.batch(BATCH_SIZE, drop_remainder=True)
train_iterator = dataset.make_one_shot_iterator()
train_features, train_labels = train_iterator.get_next()
print(train_features)
#train_labels = train_labels[:,SEQ_LENGTH-1] # output => [0 1 0]
#print(train_labels)
train_labels = np.expand_dims(train_labels, axis=2)
print(train_labels)
model = keras_model()
model.summary()
model.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])
model.fit(x=train_features,y=train_labels, batch_size=BATCH_SIZE,epochs=1, steps_per_epoch=10)
下面是输出：
...tf.Tensor(
[[[ 0 100]
[ 1 101]
[ 2 102]
[ 3 103]]
[[ 1 101]
[ 2 102]
[ 3 103]
[ 4 104]]
[[ 2 102]
[ 3 103]
[ 4 104]
[ 5 105]]], shape=(3, 4, 2), dtype=int64)
[[[0]
[0]
[0]
[0]]
[[0]
[0]
[0]
[1]]
[[0]
[0]
[1]
[0]]]
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, 4, 2) 0
_________________________________________________________________
bidirectional (Bidirectional (None, 4, 128) 34304
_________________________________________________________________
time_distributed (TimeDistri (None, 4, 2) 258
=================================================================
Total params: 34,562
Trainable params: 34,562
Non-trainable params: 0
_________________________________________________________________
1/10 [==>...........................] - ETA: 8s - loss: 13.3860 - acc: 0.1667
10/10 [==============================] - 1s 101ms/step - loss: 12.9909 - acc: 0.1667
Process finished with exit code 0
对我来说，每个序列只有一个标签来确定序列是属于类别 0 还是类别 1 会更有意义（在我的情况下，每个批次有 3 个值，因为批次大小为 3）。我试图通过添加一行代码（如下所示）来做到这一点，因为它导致异常“不兼容的形状：[3] 与 [3,4]”，因此我必须稍后将其注释掉
train_labels = train_labels[:,SEQ_LENGTH-1] # 输出 => [0 1 0]
我无法弄清楚如何修复该错误，因此正如您在输出中看到的那样，我将序列中包含的所有行的标签提供给模型。后来我想出了一个技巧，可以为序列中的所有项目使用相同的标签。我决定将序列中的所有标签设置为序列的最后一个标签。例如，[0 0 0 1] 将更改为 [1 1 1 1]，而 [0 0 1 0] 将更改为 [0 0 0 0]。我还将损失函数更改为“binary_crossentropy”，因为这里的问题是二进制分类。下面是代码：
import tensorflow as tf
import numpy as np
BATCH_SIZE = 3
SEQ_LENGTH = 4
NUM_CLASSES = 1
LSTM_UNITS = 64
NUM_SHARDS = 4
NUM_CHANNELS = 2
tf.enable_eager_execution()
def parse_values(f1, f2, label):
features = [f1,f2]
return features, label
def map_label(features, label):
sequence_label1 = tf.fill([SEQ_LENGTH],label[SEQ_LENGTH-1])
return features, sequence_label1
def keras_model():
inputs = tf.keras.layers.Input(shape=(SEQ_LENGTH,NUM_CHANNELS),batch_size=BATCH_SIZE)
x = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(LSTM_UNITS, return_sequences=True))(inputs)
outputs = tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(NUM_CLASSES, activation='sigmoid'))(x)
return tf.keras.Model(inputs, outputs)
dataset = tf.data.experimental.CsvDataset(filenames='../../input/aFile.csv', header=True,record_defaults=[tf.int64] * 3, select_cols=[0,1,2])
dataset= dataset.map(parse_values).window(size=SEQ_LENGTH, shift=1, drop_remainder=True).flat_map(lambda features, label:
tf.data.Dataset.zip((features.batch(SEQ_LENGTH), label.batch(SEQ_LENGTH)))).map(map_label)
dataset = dataset.batch(BATCH_SIZE, drop_remainder=True)
train_iterator = dataset.make_one_shot_iterator()
train_features, train_labels = train_iterator.get_next()
print(train_features)
train_labels = np.expand_dims(train_labels, axis=2)
print(train_labels)
model = keras_model()
model.summary()
model.compile(optimizer='adam',loss='binary_crossentropy',metrics=['accuracy'])
model.fit(x=train_features,y=train_labels, batch_size=BATCH_SIZE,epochs=1, steps_per_epoch=10)
下面是输出：
...tf.Tensor(
[[[ 0 100]
[ 1 101]
[ 2 102]
[ 3 103]]
[[ 1 101]
[ 2 102]
[ 3 103]
[ 4 104]]
[[ 2 102]
[ 3 103]
[ 4 104]
[ 5 105]]], shape=(3, 4, 2), dtype=int64)
[[[0]
[0]
[0]
[0]]
[[1]
[1]
[1]
[1]]
[[0]
[0]
[0]
[0]]]
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (3, 4, 2) 0
_________________________________________________________________
bidirectional (Bidirectional (3, 4, 128) 34304
_________________________________________________________________
time_distributed (TimeDistri (3, 4, 1) 129
=================================================================
Total params: 34,433
Trainable params: 34,433
Non-trainable params: 0
_________________________________________________________________
...
1/10 [==>...........................] - ETA: 10s - loss: 0.6866 - acc: 0.5833
10/10 [==============================] - 1s 124ms/step - loss: 0.6571 - acc: 0.6500
Process finished with exit code 0
我希望这能帮助任何面临类似问题的人。