Train for a given number of iterations

Loads parameters (but not training state)

Interface between Panoramic view and Egocentric view It will convert the action panoramic view action a_t to equivalent egocentric v

Break sentence into a list of words and punctuation

:param viewpoints: [[], [], ....(batch_size)]. Only for dropout viewpoint :param get_first_feat: whether output the first feat

:param train_ml: The weight to train with maximum likelihood :param train_rl: whether use RL in training :param reset:

:param x: start :param y: end :return: the path from x to y [v1, v2, ..., v_n, y]

Store the minibach in 'self.batch' :param tile_one: Tile the one into batch_size :return: None

Extract instructions from a list of observations and sort by descending sequence length (to enable PyTorch packing).

:param splits: A list of split. if the split is "something@5000", it will use a random 5000 data from the data :return:

Reset the data index to beginning of epoch. Primarily for testing. You must still call reset() for a new episode.

Extract precomputed features into variable.

Extracts all n-grams upto a given maximum order from an input segment. Args: segment: text segment from which n-grams will be extracted. ma

Calculate error based on the final position in trajectory, and also the closest position (oracle stopping rule). The path co

Determine next action on the shortest path to goal, for supervised training.

This is used for debug

Get list of states augmented with precomputed image features. rgb field will be empty. Agent's current view [0-35] (set only when vie

:param sampling: if not, use argmax. else use softmax_multinomial :param train: Whether in the train mode :return: if sampli

Load connectivity graph for each scan

Return a list of dicts containing instr_id:'xx', path:[(viewpointId, heading_rad, elevation_rad)]

Evaluate each agent trajectory based on how close it got to the goal location

The dijkstra algorithm. Was called beam search to be consistent with existing work. But it actually finds the Exact K paths w

load graph from self.scan, Store the graph {scan_id: graph} in self.graphs Store the shortest path {scan_id: {view_id_x: {vie

Extract teacher actions into variable. :param obs: The observation. :param ended: Whether the action seq is ended :re

Extract teacher actions into variable. :param obs: The observation. :param ended: Whether the action seq is ended :re

:param speaker: The speaker to be used in searching. :return: { "scan": XXX "instr_id":XXX,

Computes BLEU score of translated segments against one or more references. Args: reference_corpus: list of lists of references for each transla

Euclidean distance between two graph poses

Run simple baselines on each split.

See "utils.Tokenizer.encode_sentence(...)" for "instr_encoding" details

Initialize to zero cell states and hidden states.

Loads parameters (but not training state)

Take an action using the full state dependent action interface (with batched input). Every action element should be an (index, heading,

Snapshot models

:param inst: The id inst :return: Remove the potential <BOS> and <EOS> If no <EOS> return empty list

1. Load pretrained image feature 2. Init the Simulator. :param feature_store: The name of file stored the feature. :p

Initialize layer.