Function resize

examples/OpticalFlow/flownet_models.py:115–139

Resize input tensor with unkown input-shape by a factor Args: x (tf.Tensor): tensor NCHW factor (int, optional): resize factor for H, W Note: Differences here against Caffe have huge impacts on the quality of the predictions. Returns: tf.Tensor:

(x, mode, factor=4)

def resize(x, mode, factor=4):
    """Resize input tensor with unkown input-shape by a factor

    Args:
        x (tf.Tensor): tensor NCHW
        factor (int, optional): resize factor for H, W

    Note:
        Differences here against Caffe have huge impacts on the
        quality of the predictions.

    Returns:
        tf.Tensor: resized tensor NCHW
    """
    assert mode in ['bilinear', 'nearest'], mode
    shp = tf.shape(x)[2:] * factor
    # NCHW -> NHWC
    x = tf.transpose(x, [0, 2, 3, 1])
    if mode == 'bilinear':
        x = tf.image.resize_bilinear(x, shp, align_corners=True)
    else:
        # better approximation of what Caffe is doing
        x = tf.image.resize_nearest_neighbor(x, shp, align_corners=False)
    # NHWC -> NCHW
    return tf.transpose(x, [0, 3, 1, 2])


class FlowNetBase(ModelDesc):