Beispiel #1
0
def create_loss(out,
                feeds,
                architecture,
                classes_num=1000,
                epsilon=None,
                use_mix=False,
                use_distillation=False,
                use_pure_fp16=False):
    """
    Create a loss for optimization, such as:
        1. CrossEnotry loss
        2. CrossEnotry loss with label smoothing
        3. CrossEnotry loss with mix(mixup, cutmix, fmix)
        4. CrossEnotry loss with label smoothing and (mixup, cutmix, fmix)
        5. GoogLeNet loss

    Args:
        out(variable): model output variable
        feeds(dict): dict of model input variables
        architecture(dict): architecture information,
            name(such as ResNet50) is needed
        classes_num(int): num of classes
        epsilon(float): parameter for label smoothing, 0.0 <= epsilon <= 1.0
        use_mix(bool): whether to use mix(include mixup, cutmix, fmix)
        use_pure_fp16(bool): whether to use pure fp16 data as training parameter

    Returns:
        loss(variable): loss variable
    """
    if use_mix:
        feed_y_a = paddle.reshape(feeds['feed_y_a'], [-1, 1])
        feed_y_b = paddle.reshape(feeds['feed_y_b'], [-1, 1])
        feed_lam = paddle.reshape(feeds['feed_lam'], [-1, 1])
    else:
        target = paddle.reshape(feeds['label'], [-1, 1])

    if architecture["name"] == "GoogLeNet":
        assert len(out) == 3, "GoogLeNet should have 3 outputs"
        loss = GoogLeNetLoss(class_dim=classes_num, epsilon=epsilon)
        return loss(out[0], out[1], out[2], target)

    if use_distillation:
        assert len(out) == 2, ("distillation output length must be 2, "
                               "but got {}".format(len(out)))
        loss = JSDivLoss(class_dim=classes_num, epsilon=epsilon)
        return loss(out[1], out[0])

    if use_mix:
        loss = MixCELoss(class_dim=classes_num, epsilon=epsilon)
        return loss(out, feed_y_a, feed_y_b, feed_lam, use_pure_fp16)
    else:
        loss = CELoss(class_dim=classes_num, epsilon=epsilon)
        return loss(out, target, use_pure_fp16)
Beispiel #2
0
def create_loss(out,
                feeds,
                architecture,
                classes_num=1000,
                epsilon=None,
                mix=False):
    """
    Create a loss for optimization, such as:
        1. CrossEnotry loss
        2. CrossEnotry loss with label smoothing
        3. CrossEnotry loss with mix(mixup, cutmix, fmix)
        4. CrossEnotry loss with label smoothing and (mixup, cutmix, fmix)
        5. GoogLeNet loss

    Args:
        out(variable): model output variable
        feeds(dict): dict of model input variables
        architecture(str): model name, such as ResNet50
        classes_num(int): num of classes
        epsilon(float): parameter for label smoothing, 0.0 <= epsilon <= 1.0
        mix(bool): whether to use mix(include mixup, cutmix, fmix)

    Returns:
        loss(variable): loss variable
    """
    if architecture == "GoogLeNet":
        assert len(out) == 3, "GoogLeNet should have 3 outputs"
        loss = GoogLeNetLoss(class_dim=classes_num, epsilon=epsilon)
        target = feeds['label']
        return loss(out[0], out[1], out[2], target)

    if mix:
        loss = MixCELoss(class_dim=classes_num, epsilon=epsilon)
        feed_y_a = feeds['feed_y_a']
        feed_y_b = feeds['feed_y_b']
        feed_lam = feeds['feed_lam']
        return loss(out, feed_y_a, feed_y_b, feed_lam)
    else:
        loss = CELoss(class_dim=classes_num, epsilon=epsilon)
        target = feeds['label']
        return loss(out, target)