focal loss

        def focal_loss(prediction, target, weights=None, alpha=0.22, gamma=2):
            """Compute focal loss for predictions.
            Multi-labels Focal loss formula:
                FL = -alpha * (z-p)^gamma * log(p) -(1-alpha) * p^gamma * log(1-p)
                     ,which alpha = 0.25, gamma = 2, p = sigmoid(x), z = target_tensor.
            Args:
            prediction_tensor: A float tensor of shape [batch_size, num_anchors,
                num_classes] representing the predicted logits for each class
            target_tensor: A float tensor of shape [batch_size, num_anchors,
                num_classes] representing one-hot encoded classification targets
            weights: A float tensor of shape [batch_size, num_anchors]
            alpha: A scalar tensor for focal loss alpha hyper-parameter
            gamma: A scalar tensor for focal loss gamma hyper-parameter
            Returns:
            loss: A (scalar) tensor representing the value of the loss function
            """
            sigmoid_p = prediction
            print (alpha)
            zeros = array_ops.zeros_like(sigmoid_p, dtype=sigmoid_p.dtype) # Get a zero tensor with the given dtype.
            ones = tf.ones_like(sigmoid_p, dtype=sigmoid_p.dtype) #tf.constant(1.0, shape = prediction.get_shape())
            pos_p_sub = array_ops.where(target > zeros, ones - sigmoid_p, zeros)
            neg_p_sub = array_ops.where(target > zeros, zeros, sigmoid_p)


            per_entry_cross_ent = - alpha * (pos_p_sub ** gamma) * tf.log(tf.clip_by_value(sigmoid_p, 1e-8, 1.0)) \
                          - (1 - alpha) * (neg_p_sub ** gamma) * tf.log(tf.clip_by_value(1.0 - sigmoid_p, 1e-8, 1.0))
            return tf.reduce_sum(per_entry_cross_ent)