Right now I have two networks f and g, the first trained on task 1 and the second on task 2. I labelled my data as either beloning to task 1 or to task 2. How can I build the following (trainable) custom architecture:
x -> decide if 1 or 2 -> pass to f or g accordingly?
I've never used such a branched architecture before...
I tried to demonstrate what you need with a Sample Code shown below. Please let me know if this is not what you are looking for and give more details, and I will be Happy to help you.
As per the question, we are trying to achieve 2 Tasks, Task 1 --> Regression (Feedforward Neural Networks) and Task 2 --> CNN. We shall form 2 Datasets from the existing Dataset based on the Label, whether it belongs to Task 1 --> Data_T1 and Task 2 --> Data_T2.
Then using Functional API, we can pass Multiple Inputs and we can get Multiple Outputs.
Code is shown below:
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Conv2D, Dense, Flatten
import pandas as pd
F1 = [1,2,3,4,5,6,7,8,9,10]
F2 = [1,2,3,4,5,6,7,8,9,10]
F3 = [1,2,3,4,5,6,7,8,9,10]
Task = ['t1', 't1', 't2', 't1', 't2', 't2', 't2', 't1', 't1', 't2']
Dict = {'F1': F1, 'F2':F2, 'F3':F3, 'Task':Task} # Column Task tells us whether the Data belongs to Task1 or Task2
Data = pd.DataFrame(Dict) #Create a Dummy Data Frame
Data_T1 = Data[Data['Task']=='t1']
Data_T1 = Data_T1.drop(columns = ['Task'])
Data_T2 = Data[Data['Task']=='t2']
Data_T2 = Data_T2.drop(columns = ['Task'])
Input1 = ...
Input2 = ...
Number_Of_Classes = 3
# Regression Model
D1 = Dense(10, activation = 'relu')(Input1)
Out_Task1 = Dense(1, activation = 'linear')
# CNN Model
Conv1 = Conv2D(16, (3,3), activation = 'relu')(Input2)
Conv2 = Conv2D(32, (3,3, activation = 'relu'))(Conv1)
Flatten = Flatten()(Conv2)
D2_1 = Dense(10, activation = 'relu')
Out_Task2 = Dense(Number_Of_Classes, activation = 'softmax')
model = Model(inputs = [Input1, Input2], outputs = [Out_Task1, Out_Task2])
model.compile....
model.fit([Data_T1, Data_T2], .....)