7c8af7a941f49539f2d5eb1b17d61952fc70e2bc,niftynet/layer/resampler.py,ResamplerLayer,_resample_inv_dst_weighting,#ResamplerLayer#,171

Before Change

        b_id = tf.concat([b_id, knots_id], axis=-1)

        samples = tf.gather_nd(inputs, b_id)  // output 8, 2, 2, n_channels
        samples = tf.reduce_sum(
            samples * tf.expand_dims(point_weights, axis=-1), axis=0)
        samples = samples / tf.reduce_sum(point_weights, axis=0)
        return samples

After Change

        point_weights = tf.pow(point_weights, self.power / 2.0)
        point_weights = tf.reciprocal(point_weights)
        // workaround for zero weights
        point_weights = tf.minimum(point_weights, 1e12)

        // transpose to shape inds: [0, -1, others]
        all_coords = tf.transpose(
            all_coords, [0, len(out_size)] + range(1, len(out_size)))
        knots_id = tf.gather_nd(all_coords, weight_id)
        knots_id = tf.transpose(
            knots_id, [0] + range(2, out_spatial_rank + 3) + [1])
        knots_shape = knots_id.get_shape().as_list()
        b_id = tf.reshape(
            tf.range(batch_size),
            [1] + [batch_size] + [1] * (len(knots_shape) - 2))
        b_id = tf.tile(
            b_id, [knots_shape[0]] + [1] + out_spatial_size + [1])
        b_id = tf.concat([b_id, knots_id], axis=-1)

        point_weights = tf.expand_dims(point_weights, axis=-1)
        samples = tf.gather_nd(inputs, b_id)
        samples = tf.reduce_sum(samples * point_weights, axis=0)
        samples = samples / tf.reduce_sum(point_weights, axis=0)
        return samples

Instances