Initialize the graph and session, if this has not yet been done.

Args: image: A 4-D tensor of shape `[batch, in_height, in_width, in_channels]`. kernels: A 4-D of shape `

Args: image: A 4-D tensor of shape `[batch, in_height, in_width, in_channels]`. kernels: A 6-D of shape `

:param enc: time, batch, z_dim :param params: :param tlen: length of horizon :return: e in time-major

Create the TensorFlow graph needed to compute this operation. This should write ops to the default graph and return `None`.

L2 distance between tensors true and pred. Args: true: the ground truth image. pred: the predicted image. keep_axis:

Hack that uses pr_curve summaries for 2D plots. Args: x: 1-D tensor with values in increasing order. y: 1-D tensor with stat

To generate a gif from image files, first generate palette from images and then generate the gif from the images and the palette. ffmpeg

we use separate trajectories for the encoder than from the ones used for prediction training :param inputs: dict with tensors in *time-major*

Possibly less computationally efficient but more memory efficent than upsampled2d.

:param image_tensor: float 32 array of Batch x Height x Width x Channel immages (range 0 - 1) :return: pre-processed images (ready to input t