init_value : initial value for the scalar

convert into overlapping chunks. Chunk size = 2w, overlap size = w

Matrix multiplication of query and key tensors using with a sliding window attention pattern. This implementation splits the input into overl

Stochastic Depth per sample.

A simple dataloder builder

A simple dataset builder

Args: optimizer (Optimizer): Wrapped optimizer. warmup_epochs (int): Maximum number of iterations for linear warmup

Get all predicitons of the given label. Return empty DataFrame if there is no predcitions with the given label.

shift every row 1 step right, converting columns into diagonals. Example:: chunked_hidden_states: [ 0.4983, 2.6918, -0

pads rows and then flips rows and columns

Same as _sliding_chunks_query_key_matmul but for attn_probs and value tensors. Returned tensor will be of the same shape as `attn_pro

Distance-IoU Loss (Zheng et. al) https://arxiv.org/abs/1911.08287 This is an implementation that assumes a 1D event is represented using

Evaluates a prediction file. For the detection task we measure the interpolated mean average precision to measure the performance of a

Sinusoid position encoding table

Interpolated AP - VOCdevkit from VOC 2011.

0. load config

main function that handles training / inference

create optimizer return a supported optimizer

create scheduler return a supported scheduler All scheduler returned by this function should step every iteration

default_cfg: a reference config (template) cfg_dict: a dictionary of parameters that will be varied during experiments This function wil

Generate batched features and masks from a list of dict items

save checkpoint to file

blur localization results by incorporating side segs. this is known as bounding box voting in object detection literature. sl

Compute the temporal intersection over union between a target segment and all the test segments. Parameters ---------- target_segment

Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002. Taken from https://github.com/facebookresearch/fvcore/blob/

Training the model for one epoch

r"""Fills the input Tensor with values drawn from a truncated normal distribution. The values are effectively drawn from the normal distributi

Truncate feats and time stamps in a dict item data_dict = {'video_id' : str 'feats' : Tensor C x T

Test the model on the validation set

Computes average precision for each class in the subset.

bind to torch interface

Args: optimizer (Optimizer): Wrapped optimizer. warmup_epochs (int): Maximum number of iterations for linear warmup