CODEWITHSUNDEEP
tensorflowkerasloss-functionmodel-fitting
mssandes
mssandes

Reputation: 55

Building Custom Loss and accessing data from y_pred

I'm working on a regression problem to reconstruct the information to mitigate crosstalk effects from neighbors' sensors, and I want to change the loss function to evaluate the standard deviation in an intermediate step inside my custom loss function.

Here is the code that I want to implement:

My y_pred as shape (nEvents, 100)

I found many examples to create my loss function as below:

import backend as K
def my_loss_func1(y_true, y_pred):
    return K.mean(K.square(y_pred - y_actual) + K.square(layer), axis=-1)

But in those examples, I didn't find how to access data inside the y_pred (TensorFlow variable). In all examples, it just uses a function from the backend to evaluate the loss passing y_pred, y_true as a parameter.

What I need to do is:

def my_loss_func2(y_true, y_pred):
    samples = y_pred( <HOW DO TO GET THE VALUES?> )
    cells   = int(samples.shape[1]/4)

    AmpRec  = np.tensordot(samples.reshape(samples.shape[0], signals, nSamp),ai, axes=(2,0))
    TimeRec = np.tensordot(samples.reshape(samples.shape[0], signals, nSamp),ai, axes=(2,0))/AmpRec
    
    return K.std(TimeRec)

I want to improve and update my model to adjust weights, taking into account the optimal_filter output concerning the timeRec spread (std).

Has anybody evaluated the loss function accessing y_pred like I want to do?

Upvotes: 1

Views: 140

Answers (1)

mssandes
mssandes

Reputation: 55

Today, a few days after...

I found a solution, reading the post how to print Keras tensors.

There exist some methods to manipulate a tensor, like:

K.eva(tensor)
K.print_tensor(tensor)
...

But to enable those, it's necessary to add run_eagerly on the model compile method:

model.compile(..., run_eagerly = True)

After this single step, you can access tensor information to manipulate as you need.

Upvotes: 1

Related Questions