:param s: :return:

:param x: [B x num_channels x T] :return: [B x num_channels x T]

Given a list of items, it summarizes them in a way that no two neighboring values are the same. It also returns the size of each section.

Predicts the temporal action probabilities S using the intermediate representation Z :param Z: intermediate representation [1 x D' x

It will use nearest neighbor interpolation to make the prediction the same size as the target.

:param A: [1 x C x K] :param A_hat: :param weight: :param loss_type: kl_div or L1 :param activation: :return:

Calculates the Mean over frames segmentation metric. :param predictions: List of numpy arrays. Each array is assumed to do 1D. It should cont

:param A: [1 x C x K] :param A_hat: :return:

:param A: :param A_hat: :param weight: :return:

:param A: :param A_hat: :return:

Normalizes the absolute value of z_where to the range that is appropriate for the network. :param T: :param pis: :par

:param cfg: :param dataset: :return:

args: :param p: Tensor same size as q :param q: Tensor same size as p :returns: kl divergence between the `p` and `q`

Parse input arguments

:param T: 1:int :param L: [1 x T']:float :return: [1 x T']:float

relu(s - w) + relu(- w - s)

Set Constrained Temporal Transformation :param W: masks [1 x M x T] :param S: S intermediate temporal representation over input video X [

parameters: no_features: if set to True, this method will *not* actually load_model the features from disk. This is u

:param main_source: :param indices: :param num_iter: 0: how_many=main_source_len/len(self.indices), othe

feat_list: a file containing the relative path to numpy files with video features in them, separated by new line video fea