temp2.py

class SelectLoss:
    """
        Selection based on Loss values of samples.
        No need of rejection sampling.
    """

    def __init__(self, X, Y, fwd_batch_size, batch_size, _,  loss):
        """
        :param loss: loss function
        :param x_train: training dataN
        :param y_train: training labels
        """
        self.X = X
        self.Y = Y
        self.fwd_batch_size = fwd_batch_size
        self.batch_size = batch_size
        self.loss = loss
        self.A = tf.placeholder('float', shape=[None, 10])
        self.B = tf.placeholder('float', shape=[None, 10])
        self.res = K.get_value (tf.nn.softmax_cross_entropy_with_logits (labels=self.A, logits=self.B))

    def sample(self, model):
        """
            Sort the loss values of the training samples.
            Variables
        """
        idx = np.random.choice (np.arange (0, self.X.shape[0]), size=self.fwd_batch_size, replace=False)
        res = model.predict_proba (self.X[idx])
        with tf.Session () as sess:
            sess.run(self.res, feed_dict={self.A: self.Y[idx], self.B: res})
        res = res / np.sum (res)
        return np.random.choice (idx,
                                 size=self.batch_size,
                                 replace=False,
                                 p=res)