Как работать с loss?

Добрый вечер, пара вопросов насчет вот этой функции, вопросы помечены в комментах: def loss(self, X_batch, y_batch, reg): """Logistic Regression loss function Inputs:

X: N x D array of data. Data are D-dimensional rows
y: 1-dimensional array of length N with labels 0-1, for 2 classes Returns: a tuple of:
loss as single float
gradient with respect to weights w; an array of same shape as w """ dw = np.zeros_like(self.w) # initialize the gradient as zero loss = 0
Compute loss and gradient. Your code should not contain python loops.

    pred = self.predict(X_batch)
    loss = hl(y_batch , pred) #hamming function
    dw = X_batch.transpose() * (pred - y_batch).transpose()
    dw /= X_batch.shape[0]

    # Right now the loss is a sum over all training examples, but we want it
    # to be an average instead so we divide by num_train.
    # Note that the same thing must be done with gradient.

<-----Что здесь вообще надо сделать и зачем это??? Почему моя ошибка - это сумма по всем выборкам?

    # Add regularization to the loss and gradient.
    # Note that you have to exclude bias term in regularization.
    <--------------Как исключить "термин смещения"?

    loss *= reg
    dw *= reg

    return loss, dw

vkantor / MIPT_Data_Mining_In_Action_2016

Как работать с loss? #25

Compute loss and gradient. Your code should not contain python loops.