3a5ace850931e91c55a692ae7ec716a57e66f4e6,niftynet/layer/loss_segmentation.py,LossFunction,layer_op,#LossFunction#Any#Any#Any#,46

Before Change

        with tf.device("/cpu:0"):

            batch_size = ground_truth.get_shape()[].value
            ground_truth = tf.reshape(ground_truth, [batch_size, -1])
            if weight_map is not None:
                weight_map = tf.reshape(weight_map, [batch_size, -1])
            // assumes same gt and weight across scales

            // prediction should be a list for multi-scale losses
            // single scale ``prediction`` is converted to ``[prediction]``
            if not isinstance(prediction, (list, tuple)):
                prediction = [prediction]

            data_loss = []
            for ind, pred in enumerate(prediction):
                // go through each scale

                loss_batch = []
                for b_ind, pred_b in enumerate(tf.unstack(pred, axis=0)):
                    // go through each image in a batch

                    pred_b = tf.reshape(pred_b, [-1, self._num_classes])
                    if self._softmax:
                        pred_b = tf.nn.softmax(
                            tf.cast(pred_b, dtype=tf.float32))
                    ground_truth_b = ground_truth[b_ind]
                    weight_b = None if weight_map is None else weight_map[b_ind]

                    loss_params = {
                        "prediction": pred_b,
                        "ground_truth": ground_truth_b,

After Change

                    // performs softmax if required
                    if self._softmax:
                        pred_b = tf.cast(pred_b, dtype=tf.float32)
                        pred_b = tf.nn.softmax(pred_b)

                    // reshape pred, ground_truth, weight_map to the same
                    // size: (n_voxels, num_classes)
                    // if the ground_truth has only one channel, the shape
                    // becomes: (n_voxels,)
                    spatial_shape = pred_b.get_shape().as_list()[:-1]
                    ref_shape = spatial_shape + [-1]
                    ground_truth_b = tf.reshape(ground_truth[b_ind], ref_shape)
                    if ground_truth_b.get_shape().as_list()[-1] == 1:
                        ground_truth_b = tf.squeeze(ground_truth_b, axis=-1)
                    if weight_map is not None:
                        weight_b = tf.reshape(weight_map[b_ind], ref_shape)
                        if weight_b.get_shape().as_list()[-1] == 1:
                            weight_b = tf.squeeze(weight_b, axis=-1)
                    else:
                        weight_b = None

                    // preparing loss function parameters
                    loss_params = {
                        "prediction": pred_b,
                        "ground_truth": ground_truth_b,
                        "weight_map": weight_b}

Instances