用Keras实现对抗训练+lstm的一个报错

猪仔很忙

我正在尝试建立结合对抗训练的LSTM模型，代码如下：

1. def gradient_operation(args):
   
2.     y_true = args[0]
   
3.     y_pred = args[1]
   
4.     v_final = args[2]
   
5.     pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred))
   
6.     pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred))
   
7.     loss = -K.mean(0.75 * K.pow(1. - pt_1, 0) * K.log(pt_1)) - K.mean((1 - 0.75) * K.pow(pt_0, 0) * K.log(1. - pt_0))
   
8.     perturb = K.gradients(loss, v_final)
   
9.     adv_v_final = K.gradients(loss, v_final) + v_final
   
10.     return adv_v_final
    
11. 
    
12. def build_train(datas, machineID, modelName):
    
13.     train_data, train_labels, val_data, val_labels, test_data, test_labels = datas
    
14. 
    
15.     num_samples = train_data.shape[0]
    
16.     time_step = train_data.shape[1]
    
17.     feature_dim = train_data.shape[2]
    
18. 
    
19.     trace_input = Input(shape=(time_step, feature_dim))
    
20.     label_input = Input(shape=(1,))
    
21. 
    
22.     if modelName == 'FEMT_LSTM':
    
23.         v_final = RNN_SEPARATE_2(time_step, feature_dim)(trace_input)
    
24.     else:
    
25.         v_final = RNN(time_step, feature_dim)(trace_input)
    
26.     pred = Dense(1, activation='sigmoid', name='pred')(v_final)
    
27.     v_final_adv = Lambda(gradient_operation)([label_input, pred, v_final])
    
28.     adv_pred = Dense(1, activation='sigmoid', name='adv_pred')(v_final_adv)
    
29.     model = Model(inputs=[trace_input, label_input], outputs=[pred, adv_pred])
    
30.     model.compile(optimizer=optimizers.Adam(lr=0.001, clipvalue=15),
    
31.                   loss={'pred': myloss(alpha=0.75, gamma=0), 'adv_pred': myloss(alpha=0.75, gamma=0)},
    
32.                   loss_weights={'pred': 1., 'adv_pred': 0.1})
    
33. 
    
34.     print('Train...')
    
35.     model_save_path = './lib/model_cp_rnn_{}'.format(machineID)
    
36. 
    
37.     call_backs = [Call_back_0(valid_data=[val_data, val_labels, test_data, test_labels], # test_data, test_labels
    
38.                               model_save_path=model_save_path),
    
39.                   ReduceLROnPlateau(monitor='val_loss', factor=0.8, patience=4, mode='min'), # 4
    
40.                   ModelCheckpoint(filepath=model_save_path, monitor='val_loss', save_best_only=True, save_weights_only=True, mode='min')]
    
41. 
    
42.     model.fit({'trace_input': train_data, 'label_input': train_labels},
    
43.             y=[train_labels, train_labels],
    
44.             batch_size=64,
    
45.             epochs=60, # 30 60
    
46.             callbacks=call_backs,
    
47.             validation_data=[val_data, val_labels])

复制代码

目标函数是：loss=(-1/N) Σ ylogy1 + (1-y)log(1-y1) + 0.01 * ((-1/N)Σ ylogy2 + (1-y)log(1-y2))，y是真实值，对应代码中的label_input、y_true；y_1是非对抗样本的预测结果，对应pred、y_pred；y_2是对抗样本的预测结果，对应adv_pred。

训练数据的每个样本是一个10*7矩阵的时序数据（暂时称为x），预测分类结果y_2 = f(x)。在模型的中间过程，我想通过求梯度制造一个噪声（见图中的adv_v_final），并将其与v_final相加，再用相加结果进行激活得到分类预测值。

为了避免出现NoneType' object has no attribute '_inbound_nodes'这个报错，我把求噪声的过程放入一个Lambda层。但是现在出现报错：

1. Traceback (most recent call last):
   
2.   File "<input>", line 1, in <module>
   
3.   File "D:\JetBrains\PyCharm 2020.1.2\plugins\python\helpers\pydev\_pydev_bundle\pydev_umd.py", line 197, in runfile
   
4.     pydev_imports.execfile(filename, global_vars, local_vars)  # execute the script
   
5.   File "D:\JetBrains\PyCharm 2020.1.2\plugins\python\helpers\pydev\_pydev_imps\_pydev_execfile.py", line 18, in execfile
   
6.     exec(compile(contents+"\n", file, 'exec'), glob, loc)
   
7.   File "D:/PycharmProjects/turningpoint-master/main.py", line 67, in <module>
   
8.     modelName=modelName)
   
9.   File "D:\PycharmProjects\turningpoint-master\baseline_main_rnn.py", line 37, in train_rnn_turning_point
   
10.     clf = rnn_turning_point.build_train(datas, machineID, modelName)
    
11.   File "D:\PycharmProjects\turningpoint-master\model\rnn_turning_point.py", line 52, in build_train
    
12.     perturbation = Lambda(gradient_operation)([label_input, pred, v_final])
    
13.   File "D:\Anaconda3\envs\edogawaAi\lib\site-packages\keras\engine\base_layer.py", line 474, in __call__
    
14.     output_shape = self.compute_output_shape(input_shape)
    
15.   File "D:\Anaconda3\envs\edogawaAi\lib\site-packages\keras\layers\core.py", line 654, in compute_output_shape
    
16.     return [K.int_shape(x_elem) for x_elem in x]
    
17.   File "D:\Anaconda3\envs\edogawaAi\lib\site-packages\keras\layers\core.py", line 654, in <listcomp>
    
18.     return [K.int_shape(x_elem) for x_elem in x]
    
19.   File "D:\Anaconda3\envs\edogawaAi\lib\site-packages\keras\backend\tensorflow_backend.py", line 593, in int_shape
    
20.     return tuple(x.get_shape().as_list())
    
21. AttributeError: 'NoneType' object has no attribute 'get_shape'

复制代码

我是一个机器学习和python的小白 除了这个报错，代码逻辑可能也有问题，对此也请各位不吝赐教：）

猪仔很忙

抱歉，我应该发到python交流区