0

I have designed a network for multi-task learning. Here I have some shared layers and specific layers for 5 different tasks.

 input_dim_train=X_train.shape[1]
 n_input_dim = int(input_dim_train)
 shared_model = Dense(input_dim_train)(inputs)
 shared_model = Dense(n_input_dim, activation='relu')(shared_model)
 shared_model = BatchNormalization(input_dim=n_input_dim)(shared_model)
 shared_model = Dropout(0.1)(shared_model)

 shared_model = Dense(int(n_input_dim//4), activation='relu')(shared_model)
 shared_model = BatchNormalization(input_dim=n_input_dim)(shared_model)
 shared_model = Dropout(0.1)(shared_model)

 shared_model = Dense(int(n_input_dim//12), activation='relu')(shared_model)
 shared_model = BatchNormalization(input_dim=n_input_dim)(shared_model)
 shared_model = Dropout(0.1)(shared_model)

 shared_model = Dense(int(n_input_dim//20), activation='relu')(shared_model)
 shared_model = BatchNormalization()(shared_model)
 shared_model = Dropout(0.1)(shared_model)

 output1 = Dense(3, activation='softmax',name='target_1',)(shared_model)
 output2 = Dense(2, activation='softmax',name='target_2')(shared_model)
 output3 = Dense(4, activation='softmax',name='target_3')(shared_model)
 output4 = Dense(2, activation='softmax',name='target_4')(shared_model)
 output5 = Dense(3, activation='softmax',name='target_5')(shared_model)

 model = Model(inputs=inputs, outputs=[output1, output2,output3,output4,output5])

 model.compile(optimizer='adam', loss='categorical_crossentropy',metrics=['accuracy',f1_metric])



 history=model.fit(X_train, [y_train_task1, y_train_task2,y_train_task3,y_train_task4,y_train_task5],
              validation_data=(X_test,[y_test_task1,y_test_task2,y_test_task3,y_test_task4,y_test_task5]),verbose='auto',epochs=n_epoch,batch_size=n_batch_size)

 task1_pred,task2_pred,task3_pred,task4_pred,task5_pred  = model.predict(X_test)

The network somehow works well and gives the results as follows:

enter image description here

Now, my question is: How can I measure the effect of e.g., Task1 on Task2 in this multi task learning? Indeed In want to calculate the task affinity score between the tasks in this problem. Is there any way to calculate the inter-task affinity score during the training process--over different epochs?

jeny ericsoon
  • 131
  • 7
  • You can train single-task models with data you have and get the results of them and compare the results with your multi-task model. – A.Najafi Nov 23 '21 at 15:56
  • This way doesn't provide any knowledge about the effect of Task1 on Task2. Basically in this way we can compare multi-task with individual learning? – jeny ericsoon Nov 23 '21 at 16:11
  • 1
    yes, As much I know this is the only way! – A.Najafi Nov 23 '21 at 16:18

0 Answers0