Filter with objectness or class confidence scores.

Filters boxes according to the paper rule. To note: the default represents our threshold when evaluating GEN4 resolution (1280x720) To note:

Filter bbox as in data pre-processing.

Get the bbox_tracker for the given bbox_idx.

Compose the name used in wandb run's name and ckp path.

Compose the name used in wandb run's name and ckp path.

Calculate whether the center of an object is located in a fixed range of an anchor. This is used to avert inappropriate matching. It

Creates figure to visualize gradients flowing through different layers in the net during training. Can be used for checking for possible gradient

Returns a batch of bbox labels, [num_frames, num_bbox, 5]. Each frame is padded to the max number of boxes.

Get xxx/objframe_idx_2_repr_idx.npy file name.

Collect the predicted bbox's IoU with GTs, and their cls/obj_scores.

Load a list of sub-seq (train) or one full event seq (val/test).

Generate sequences such that we do always have labels **within** (not necessarily at the end!) each sample of the sequence. This is

Returns the current bounding box estimate (with class_id).

Sub-sample labels from a long sequence.

Build the native PyTorch Lightning WandB logger.

Get a zipped datapipe that load a batch of data at every `next()`.

Hard Mish Experimental, based on notes by Mish author Diganta Misra at https://github.com/digantamisra98/H-Mish/blob/0da20d4bc58e696b6803f2

Hard Mish Experimental, based on notes by Mish author Diganta Misra at https://github.com/digantamisra98/H-Mish/blob/0da20d4bc58e696b6803f2

Computes IOU between two bboxes in the form [x,y,w,h,(cls_id)] both bbox are in shape (N, 4/5) where N is the number of bboxes If class

Unflatten a 1D list to a list of list.

Flatten a list of list to a 1D list. Should also record the original shape so that we can unflatten it later.

Mish: A Self Regularized Non-Monotonic Neural Activation Function - https://arxiv.org/abs/1908.08681 NOTE: I don't have a working inplace variant

Mish: A Self Regularized Non-Monotonic Neural Activation Function - https://arxiv.org/abs/1908.08681

Decay the tracklet's confidence.

N C H W -> N H W C

N H W C -> N C H W

Pad the labels to length N and stack them for return.

Convert the model prediction after post-processing to pseudo GT. 1. Filter with objectness score threshold. 2. Filter with class score thresh

Run model over one full event sequence.

Sample new augmentation parameters that will be consistently applied among the items. This function only works with augmentations that are in

Computes a set of top-left coordinates from which the top-left corner of the zoom window can be sampled such that the zoom window is guaranteed t

Read data from original and newly generated dataset.

Detach all.

Reset the hidden states of all/partial inputs (RNN) to 0.

reset at beginning of file