c18fada581afec3f6febb981c6ff26522dab39d0,keras_rcnn/layers/losses/_rcnn.py,RCNNRegressionLoss,compute_loss,#Any#Any#,50

Before Change

        smooth_l1_sign = tensorflow.cast(tensorflow.less(tensorflow.abs(inside_mul), 1.0 / sigma2), tensorflow.float32)
        smooth_l1_option1 = tensorflow.multiply(tensorflow.multiply(inside_mul, inside_mul), 0.5 * sigma2)
        smooth_l1_option2 = tensorflow.subtract(tensorflow.abs(inside_mul), 0.5 / sigma2)
        smooth_l1_result = tensorflow.add(tensorflow.multiply(smooth_l1_option1, smooth_l1_sign), tensorflow.multiply(smooth_l1_option2, tensorflow.abs(tensorflow.subtract(smooth_l1_sign, 1.0))))
        loss = tensorflow.multiply(outside_weights, smooth_l1_result)
        loss = tensorflow.reduce_mean(tensorflow.reduce_sum(loss, reduction_indices=[1]))

        return loss

After Change

        inside_weights = 1.0
        outside_weights = 1.0
        sigma = 1.0
        sigma2 = keras.backend.square(sigma)
        inside_mul = inside_weights * (output - target)
        smooth_l1_sign = tensorflow.cast(keras.backend.less(keras.backend.abs(inside_mul), 1.0 / sigma2), tensorflow.float32)
        smooth_l1_option1 = (inside_mul * inside_mul) * (0.5 * sigma2)
        smooth_l1_option2 = keras.backend.abs(inside_mul) - (0.5 / sigma2)
        smooth_l1_result = (smooth_l1_option1 * smooth_l1_sign) + (smooth_l1_option2 * keras.backend.abs((smooth_l1_sign - 1.0)))
        loss = outside_weights * smooth_l1_result
        loss = tensorflow.reduce_mean(tensorflow.reduce_sum(loss, reduction_indices=[1]))

        return loss

Instances