问题描述
我正在尝试训练序列到序列自动编码器。输入形状为[sample,time_step,input_dim]，每个样本的时间步长互不相同。这是我模型的实现：
input_dim = 385
output_dim = 32
cells = int(output_dim / 2)
def create_model():
model = Sequential()
model.add(Input(shape=(None,input_dim,),ragged=True))
model.add(Timedistributed(Dense(output_dim,activation='relu')))
model.add(LSTM(cells,return_sequences=True))
model.add(Timedistributed(Dense(output_dim,activation='relu')))
model.add(Timedistributed(Dense(input_dim,activation='sigmoid')))
return model
我的初始数据是格式为[list(),list(),list()...]的python列表，其中每个list()是一个序列，形状为[time_step,input_dim]。我通过以下函数将数据转换为Tensorflow RaggedTensor：
def process_data(data):
data_shape = [len(i) for i in data]
data = [item for sublist in data for item in sublist]
data = to_categorical(data,num_classes=input_dim)
data = tf.RaggedTensor.from_row_lengths(data,data_shape)
return data
这是我的训练代码：
model = create_model()
model.compile(optimizer='adam',loss='mse')
# Load training data for autoencoder
train_ae = np.load('dataset/ADFA-LD/train_ae.npy',allow_pickle=True)
train_ae = process_data(train_ae)
# Load validation data (a combination of normal and malware data)
val_norm = np.load('dataset/ADFA-LD/val_norm.npy',allow_pickle=True)
val_att = np.load('dataset/ADFA-LD/val_att.npy',allow_pickle=True)
y_val = np.concatenate([np.zeros(len(val_norm)),np.ones(len(val_att))])
val_norm = process_data(val_norm)
val_att = process_data(val_att)
x_val = tf.concat([val_norm,val_att],axis=0)
# Train
cb = validate(x_val,y_val)
model.fit(x=train_ae,y=train_ae,epochs=10,callbacks=[cb])
我训练时出现错误：
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\keras\engine\training.py:806 train_function *
return step_function(self,iterator)
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\keras\engine\training.py:796 step_function **
outputs = model.distribute_strategy.run(run_step,args=(data,))
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\distribute\distribute_lib.py:1211 run
return self._extended.call_for_each_replica(fn,args=args,kwargs=kwargs)
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\distribute\distribute_lib.py:2585 call_for_each_replica
return self._call_for_each_replica(fn,args,kwargs)
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\distribute\distribute_lib.py:2945 _call_for_each_replica
return fn(*args,**kwargs)
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\keras\engine\training.py:789 run_step **
outputs = model.train_step(data)
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\keras\engine\training.py:748 train_step
loss = self.compiled_loss(
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\keras\engine\compile_utils.py:204 __call__
loss_value = loss_obj(y_t,y_p,sample_weight=sw)
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\keras\losses.py:149 __call__
losses = ag_call(y_true,y_pred)
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\keras\losses.py:250 call **
y_pred,y_true = tf_losses_util.squeeze_or_expand_dimensions(
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\ops\losses\util.py:71 squeeze_or_expand_dimensions
y_true,y_pred = confusion_matrix.remove_squeezable_dimensions(
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\ops\confusion_matrix.py:62 remove_squeezable_dimensions
predictions = ops.convert_to_tensor(predictions)
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\framework\ops.py:1499 convert_to_tensor
ret = conversion_func(value,dtype=dtype,name=name,as_ref=as_ref)
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\framework\constant_op.py:338 _constant_tensor_conversion_function
return constant(v,name=name)
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\framework\constant_op.py:263 constant
return _constant_impl(value,dtype,shape,name,verify_shape=False,C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\framework\constant_op.py:280 _constant_impl
tensor_util.make_tensor_proto(
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\framework\tensor_util.py:550 make_tensor_proto
raise TypeError("Failed to convert object of type %s to Tensor. "
TypeError: Failed to convert object of type <class 'tensorflow.python.ops.ragged.ragged_tensor.RaggedTensor'> to Tensor. Contents: tf.RaggedTensor(values=Tensor("sequential/time_distributed_2/dense_2/Sigmoid:0",shape=(None,385),dtype=float32),row_splits=Tensor("sequential/time_distributed_2/RaggedFromrowLengths/control_dependency:0",dtype=int64)). Consider casting elements to a supported type.
我假设错误是由于从RaggedTensor到Tensor转换失败而引起的，从日志中您可以看到该错误也是与损失和预测有关的函数引起的：
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\keras\engine\training.py:748 train_step
loss = self.compiled_loss(
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\keras\engine\compile_utils.py:204 __call__
loss_value = loss_obj(y_t,y_pred = confusion_matrix.remove_squeezable_dimensions(
C:\Users\hennm\AppData\Local\Programs\Python\python38\lib\site-packages\tensorflow\python\ops\confusion_matrix.py:62 remove_squeezable_dimensions
predictions = ops.convert_to_tensor(predictions)
所以我认为损失函数的实现方式是尝试将y_pred转换为张量，因此不能使用RaggedTensor类型的输出（预测）。如果是这样，有办法解决吗？
解决方法
可以将参差不齐的张量转换为张量并用于进一步处理吗。
例如：
import tensorflow as tf
mylist = [
[[1,2,3],[5,4,6]],[[2,8,9]]
]
rt = tf.ragged.constant(mylist)
features_set = rt.to_tensor()
features_set
输出：
<tf.Tensor: shape=(2,3),dtype=int32,numpy=
array([[[1,9],[0,0]]],dtype=int32)>