How can I count the number of times I call a loss function?-CodePudding

I personalized my own Huber loss function in the way (https://goodboychan.github.io/python/coursera/tensorflow/deeplearning.ai/2022/02/08/01-Tensorflow2-Custom-Loss-Function.html) suggests:

def my_huber_loss(y_true, y_pred):
    threshold = 1.
    error = y_true - y_pred
    is_small_error = tf.abs(error) <= threshold
    small_error_loss = tf.square(error) / 2
    big_error_loss = threshold * (tf.abs(error) - threshold / 2)
    return tf.where(is_small_error, small_error_loss, big_error_loss)

I included it in model.compile(optimizer='adam', loss=my_huber_loss, metrics=['mae']) and training works fine.

Now, I would like to know how many times we call this Huber loss through the training phase, so I did as is there a way to track the number of times a function is called? suggests:

def my_huber_loss(y_true, y_pred):
    threshold = 1.
    error = y_true - y_pred
    is_small_error = tf.abs(error) <= threshold
    small_error_loss = tf.square(error) / 2
    big_error_loss = threshold * (tf.abs(error) - threshold / 2)
    my_huber_loss.counter  =1 #THIS IS THE NEW LINE
    return tf.where(is_small_error, small_error_loss, big_error_loss)

my_huber_loss.counter = 0 #INITIALIZE

However, after the whole training print(my_huber_loss.counter) outputs 3:

results = model.fit(X_train, Y_train, validation_split=0.1, batch_size=1, epochs=numEpochs, callbacks=[earlystopper])
print(my_huber_loss.counter)

Prints 3.

I know this number is not correct, since loss functions should be called more times. In addition, I added the tf.print("--- Called Loss ---") line in my_huber_loss(), and I can see how we call it several times, e.g.:

Epoch 1/2
--- Called Loss ---
   1/1440 [..............................] - ETA: 56:15 - loss: 0.0411 - mae: 0.2357--- Called Loss ---
--- Called Loss ---
   3/1440 [..............................] - ETA: 47s - loss: 0.0398 - mae: 0.2291  --- Called Loss ---
--- Called Loss ---
   5/1440 [..............................] - ETA: 45s - loss: 0.0338 - mae: 0.2096--- Called Loss ---
--- Called Loss ---
   7/1440 [..............................] - ETA: 46s - loss: 0.0338 - mae: 0.2110--- Called Loss ---
--- Called Loss ---
   9/1440 [..............................] - ETA: 44s - loss: 0.0306 - mae: 0.1997--- Called Loss ---
--- Called Loss ---
  11/1440 [..............................] - ETA: 43s - loss: 0.0279 - mae: 0.1893--- Called Loss ---
--- Called Loss ---
  13/1440 [..............................] - ETA: 41s - loss: 0.0265 - mae: 0.1836--- Called Loss ---
--- Called Loss ---
  15/1440 [..............................] - ETA: 41s - loss: 0.0261 - mae: 0.1824--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
  18/1440 [..............................] - ETA: 39s - loss: 0.0250 - mae: 0.1783--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
  21/1440 [..............................] - ETA: 38s - loss: 0.0243 - mae: 0.1764--- Called Loss ---
...

What is going wrong? How can I count the number of times I call a loss function?

CodePudding user response：

I would personally use sub-classing to preserve the state. Regardless, the issue you are facing is due to using Python side effects. The TensorFlow documentation explains why one should not rely on Python side effects like object mutation or list appends. Instead, it is recommended to use the TensorFlow operations such as tf.print, assign_add, etc.

Here is one way to fix the issue using a graph-compatible TensorFlow code:

def my_huber_loss(y_true, y_pred):
    threshold = 1.
    error = y_true - y_pred
    is_small_error = tf.abs(error) <= threshold
    small_error_loss = tf.square(error) / 2
    big_error_loss = threshold * (tf.abs(error) - threshold / 2)
    my_huber_loss.counter.assign_add(1) #graph-compatible tf
    tf.print("--- Called Loss ---")
    return tf.where(is_small_error, small_error_loss, big_error_loss)

A simple model:

model = tf.keras.Sequential([
    tf.keras.layers.Dense(units=1, input_shape=[1,])
])

batch_size=1: The loss function is called 6 times per epoch (data set consisting of 6 examples).

my_huber_loss.counter = tf.Variable(0, trainable=False) #INITIALIZE
model.compile(optimizer='sgd', loss=my_huber_loss)

results = model.fit(xs, ys, batch_size=1, epochs=2, verbose=0)
print(my_huber_loss.counter)

Output:

--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
--- Called Loss ---
<tf.Variable 'Variable:0' shape=() dtype=int32, numpy=12>

batch_size=10: leading to only one batch of 6 examples. The loss function is called once per epoch.

my_huber_loss.counter = tf.Variable(0, trainable=False) #INITIALIZE
model.compile(optimizer='sgd', loss=my_huber_loss)

results = model.fit(xs, ys, batch_size=10, epochs=2, verbose=0)
print(my_huber_loss.counter)

Output:

--- Called Loss ---
--- Called Loss ---
<tf.Variable 'Variable:0' shape=() dtype=int32, numpy=2>