Create an empty ObjectLabels.

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

General interface for updating new predictions and labels.

Returns bbox labels in shape [num_bbox, 5].

Select a 2d convolution implementation based on arguments Creates and returns one of torch.nn.Conv2d, Conv2dSame, MixedConv2d, or CondConv2d.

Activation Layer Factory Fetching activation layers by name with this function allows export or torch script friendly functions to be returne

Build dataloaders.

Get event representation directory.

Extract multi-stage features from the backbone. Input: x: (B, C, H, W), image previous_states: List[(lstm_h, lstm_c)]

Predict object bbox from multi-stage features. Returns: outputs: (B, N, 4 + 1 + num_cls), [(x, y, w, h), obj_conf, cls]

Return a list of valid bbox labels and their idx.

In place conversion to numpy (detach + to cpu + to numpy).

Apply NMS on predicted bboxes. Input: predictions: (B, N, 4 + 1 + num_cls), [(x, y, w, h), obj_conf, cls] Returns: output: L

Subsample the sequence under a folder by a given ratio.

both boxes: [M/N, 4]; xyxy: False --> [x, y, w, h] center format.

draw bboxes in the image img

Build model.

Filter the scores with one or per-class thresholds.

For streaming dataset, we do NO batching. Normal for random dataset.

Get event representation h5 file name.

Load checkpoint from a file. Args: ckp_path (str): Path to checkpoint file. strict (bool, optional): Whether to allow

Store the RNN states to this worker_id (random/streaming data).

seek in the file by ev_count events :param ev_count: seek in the file after ev_count events Note that current time will be se

1) Computes a new smaller canvas size: original canvas scaled by a factor of 1/zoom_in_factor (downscaling) 2) Places the smaller can

1) Scales the input by a factor of 1/zoom_out_factor (i.e. reduces the canvas size) 2) Places the downscaled canvas into the original

Check if all elements in a list are of a certain type.

Check if all elements in a list have the same length.

Delete self.trackers[idx], move it to self.del_trackers.

Load a list of event repr in torch.float32 tensor, each [C, H, W].

Append new model predictions to self.

Construct a bbox from 4 coords (xyxy/xywh).

Get the 4 coords (xyxy/xywh) from a bbox array or tensor.

Create a new tracker.

Stack a list of numpy arrays or tensors.

Read labels from npz file.

Save labels, soft-link ev_repr. save_dir + self.path ├── event_representations_v2 │ └── ev_representation_name

get total duration of video in mus, providing there is no overflow :return: