Custom lost function

AI Discussions
1

This is my code, but I got loss function comes after sometime inf and than nan can you please tell me where is the mistake

def custom_loss(x_train):
def loss(y_true, y_pred):
#some constants
a, b, c, d = 1, 1, 500, 70

    ## Getting x_batch from labels of y_true and sorting them in increasing order of indices
    y_true_num = y_true.numpy()
    y_pred_num = y_pred.numpy()
    # print(y_true_num)
    sorted_indices = np.argsort(y_true_num[:,1])
    y_true_num = y_true_num[sorted_indices]
    y_pred_num = y_pred_num[sorted_indices]
    matching_indices_mask = np.isin(x_train[:,1], y_true_num[:, 1])
    x_batch = x_train[matching_indices_mask]
    sorted_indices = np.argsort(x_batch[:,1])
    x_batch = x_batch[:,0][sorted_indices]
    y_true_num = y_true_num[:,0]
    y_pred_num = y_pred_num[:,0]
    
    # print(y_true_num,x_batch)
    # print("length - ",len(y_true_num),len(x_batch))
    
    ## Finding the trans conductance Gm of y_true and y_pred
    assert len(y_true_num) == len(x_batch)
    
    y_true_grad = np.array([])
    y_pred_grad = np.array([])
    for i in range(len(y_true_num)-1):
        val = 0 if x_batch[i+1]==x_batch[i] else (y_true_num[i+1]-y_true_num[i])/(x_batch[i+1]-x_batch[i])
        val2 = 0 if x_batch[i+1]==x_batch[i] else (y_pred_num[i+1]-y_pred_num[i])/(x_batch[i+1]-x_batch[i])
        y_true_grad = np.append(y_true_grad,[val])
        y_pred_grad = np.append(y_pred_grad,[val2])


    ## Finally Calculating the loss
    # print("grads - ",len(y_true_grad),len(y_pred_grad))
    y_true_grad = tf.convert_to_tensor(y_true_grad)
    y_pred_grad = tf.convert_to_tensor(y_pred_grad)
    
    
    normal_loss = tf.reduce_mean(tf.square(tf.boolean_mask((y_true - y_pred) / y_true, ~tf.math.is_nan((y_true - y_pred) / y_true))))
    gradient_loss = tf.reduce_mean(tf.square(tf.boolean_mask(y_true_grad - y_pred_grad, ~tf.math.is_nan(y_true_grad - y_pred_grad))))
    # normal_loss = a * tf.reduce_mean(tf.square((y_true - y_pred) / y_true))
    # gradient_loss = c * tf.reduce_mean(tf.square(y_true_grad - y_pred_grad))
    normal_loss = tf.cast(normal_loss,dtype=tf.float64)
    print(normal_loss,gradient_loss)
    # print("Data types - normal_loss:", normal_loss.dtype, "gradient_loss:", gradient_loss.dtype)
    total_loss = normal_loss + gradient_loss
    return total_loss

return loss
1 Like
2

Can you describe what your loss function is trying to do?

For example:

  • What loss method are you trying to implement?
  • What are you going to use it for?
  • What problem does your custom loss function help to solve?
3

Neural_Network-Based_and_Modeling_With_High_Accuracy_and_Potential_Model_Speed.pdf (2.0 MB)

I want to implement the loss function of this curve (I-V) tell me how to apply it.

1 Like
4

and also I have the data of I-V curve of N - channel MOSFET

1 Like
5

I want to have the three differential di/dvg to be smooth

1 Like
6

That’s quite interesting, thanks for posting the paper.
Unfortunately I’m not able to assist with this.
Perhaps someone else from the community will contribute.

1 Like
7

can you tell me what the code is doing wrong ??

1 Like
8

or just tag someone who might can help?

1 Like