Custom keras loss function using scipy

Question

I would like to implement a custom loss function, and I am using tensorflow with keras backend.

In my loss function for each training sample (2D matrix of size (2048x192) I would like to add a bandpassed version of the corresponding training sample as a constant (non-trainable) value.

I implemented bandpass filter based on How to implement band-pass Butterworth filter with Scipy.signal.butter

from scipy.signal import butter, sosfiltfilt

def butter_bandpass_sos(lowcut, highcut, fs, order=5):
    nyq = 0.5 * fs
    low = lowcut / nyq
    high = highcut / nyq
    sos = butter(order, [low, high], analog=False, btype='band', output='sos')
    return sos

def butter_bandpass_filter_sos(data, lowcut, highcut, fs, order=5):
    sos = butter_bandpass_sos(lowcut, highcut, fs, order=order)
    y = sosfiltfilt(sos, data)
    return y

and for the loss function based on Adding a constant to Loss function in Tensorflow, I implemented:

from tensorflow.python.keras import backend
import tensorflow as tf

lowcut = 2.5e6
highcut = 7.5e6
order = 5
fs = 40e6 
def HP_func(mat):
for i in range(0, 192):
    RF_ch = mat[:, i]
    y = butter_bandpass_filter_sos(RF_ch, lowcut, highcut, fs, order=order)
    mat_band_sos[:, i] = y
return mat_band_sos

def my_custom_loss_HF(y_true,y_pred):
   HF_mat = HP_func(y_true)
   loss = backend.sqrt(tf.keras.losses.mean_squared_error(y_true, y_pred)) + HF_mat
return loss

I have three branches and therefore three losses:

model.compile(loss=['mean_squared_error', my_custom_loss_HF,'mean_squared_error'],
          loss_weights=[1.0, 1.0, 1.0],
          optimizer='Adam',
          metrics=['mae', rmse])

but I am getting this error:

Traceback (most recent call last):
  File "/usr/lib/python3/dist-packages/IPython/core/interactiveshell.py", line 2882, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-19-01efc0717223>", line 6, in <module>
    metrics=['mae', rmse])
  File "/home/z003zpjj/venv/lib/python3.6/site-packages/tensorflow/python/training/tracking/base.py", line 457, in _method_wrapper
    result = method(self, *args, **kwargs)
  File "/home/z003zpjj/venv/lib/python3.6/site-packages/tensorflow/python/keras/engine/training.py", line 337, in compile
    self._compile_weights_loss_and_weighted_metrics()
  File "/home/z003zpjj/venv/lib/python3.6/site-packages/tensorflow/python/training/tracking/base.py", line 457, in _method_wrapper
    result = method(self, *args, **kwargs)
  File "/home/z003zpjj/venv/lib/python3.6/site-packages/tensorflow/python/keras/engine/training.py", line 1710, in _compile_weights_loss_and_weighted_metrics
    self.total_loss = self._prepare_total_loss(masks)
  File "/home/z003zpjj/venv/lib/python3.6/site-packages/tensorflow/python/keras/engine/training.py", line 1770, in _prepare_total_loss
    per_sample_losses = loss_fn.call(y_true, y_pred)
  File "/home/z003zpjj/venv/lib/python3.6/site-packages/tensorflow/python/keras/losses.py", line 215, in call
    return self.fn(y_true, y_pred, **self._fn_kwargs)
  File "<ipython-input-18-a4f1cf924d3f>", line 3, in my_custom_loss_HF
    HF_mat = HP_fun(y_true)
  File "<ipython-input-17-74a2f0e736b9>", line 19, in HP_fun
    y = butter_bandpass_filter_sos(RF_ch, lowcut, highcut, fs, order=order)
  File "<ipython-input-2-4e34aa35b4cd>", line 69, in butter_bandpass_filter_sos
    y = sosfiltfilt(sos, data)
  File "/home/z003zpjj/venv/lib/python3.6/site-packages/scipy/signal/signaltools.py", line 4131, in sosfiltfilt
    x = _validate_x(x)
  File "/home/z003zpjj/venv/lib/python3.6/site-packages/scipy/signal/signaltools.py", line 3926, in _validate_x
    raise ValueError('x must be at least 1D')
ValueError: x must be at least 1D

How can I use the scipy function in my loss function?

score 0 · Answer 1 · answered Jul 27 '21 at 12:40

0

you need to rewrite the function to work with tensors, otherwise it doesnt have a gradient, the network cant be trained. you can use only functions like keras backend.

your function has even a for, forget the logic in loss function.

answered Jul 27 '21 at 12:40

BestDogeStackoverflow

1,037
9
23

Implementing filter using keras backend would be tricky, Is there a possibility that I compute the filtered version outside the loss and then pass the corresponding filtered sample to the loss function? essentially I need to know at each time which sample is being processed, how can I access the sample index? – Farnaz Jul 27 '21 at 13:26
1

the problem is not that the call to the function is made in the loss, even if you put the call in a lambda layer as last layer of the model it will still not work, cause the neural network won't know how to change the weights in the training, your only 2 options are to do this in post processing after the net or to let the neural network learn to do it ( i would try the second) – BestDogeStackoverflow Jul 28 '21 at 09:58

Custom keras loss function using scipy

1 Answers1