Constructing a CNN model for a 2D input with Keras

manpreet Best Answer 2 years ago

I know that this kind of question has been already asked a ton of com/tag/times">times before, but i haven't found a satisfactory answer or example that could aid me with the solution i'm looking for.

I'm pretty new to keras and data science in general, so excuse me if any of this sounds obvious or wrong.

I'm trying to create a CNN model that predicts a value from a 2D table (1M+, 16), the table has the following structure:

    Date  Name  [... 13 columns ... ]  Result
    1     A     ...                     0.5
    1     A     ...                     0.3
    1     A     ...                    -0.1
    ...
    1     B     ...                     0.4
    1     B     ...                    -0.8
    1     C     ...                     0.7
    ...
    2     A     ...                     0.15
    2     A     ...                     0.38
    ...

I have already tried several approaches, but the closest one was this:

def split_sequences(x_, y_ , n_steps):
  X, y = com/tag/list">list(), com/tag/list">list()
  for i in range(len(x_)):
      # find the end of this pattern
      end_ix = i + n_steps
      # check if we are beyond the dataset
      if end_ix > len(x_):
          break
      # gather com/tag/input">input and output parts of the pattern
      seq_x, seq_y = x_[i:end_ix, :], y_[end_ix-1]
      X.append(seq_x)
      y.append(seq_y)
  return array(X), array(y)

n_steps = 5
# X and Y are the training sets, val_X and val_Y are the validation sets
X_, Y_ = split_sequences(X,Y,n_steps)
n_features = X_.shape[2]
val_X_, val_Y_ = split_sequences(val_X,val_Y,n_steps)
print(X_.shape)
print(Y_.shape)
print(val_X_.shape)
print(val_Y_.shape)


->(1804642, 5, 15)
->(1804642,)
->(601545, 5, 15)
->(601545,)


model = Sequential()
model.add(Conv1D(filters=64, kernel_size=2, activation
                                                

                                                                                                        
                                                    
                                                
                                                


                                                    

                                                        
                                                            0
                                                            views 

                                                            0
                                                                likes 
                                                            
                                                            

                                                            0
                                                            shares

                                                            


    

         Facebook
        
         Twitter
        
        
         Linked In 
        
          WhatsApp