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

Build a vocab, starting with base vocab containing a few useful tokens.

Eval sequence to sequence models on val splits (iteration selected from training error)

Expects input vocab indices as (batch, seq_len). Also requires a list of lengths for dynamic batching.

Propagate h through the network. h: batch x dim context: batch x seq_len x dim mask: batch x seq_len indices to be masked

Takes a single step in the decoder LSTM (allowing sampling). action: batch x angle_feat_size feature: batch x 36 x (feature_s

Takes a single step in the decoder LSTM (allowing sampling). cand_feat: batch x cand x (feature_size + angle_feat_size) h1: b

Takes a single step in the decoder LSTM (allowing sampling). cand_feat: batch x cand x (feature_size + angle_feat_size) h1: b

:param action_embeds: (batch_size, length, 2052). The feature of the view :param feature: (batch_size, length, 36, 2052). The action

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

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

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

Loads parameters (but not training state)

Loads parameters (but not training state)

Load a new minibatch / episodes.

Snapshot models

Snapshot models

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

Take action (same interface as makeActions)

features: batchsize x length x feature_size insts: batchsize x ins length x feature_size

Evaluate once on each instruction in the current environment

:param iters: :return: path2inst: path_id --> inst (the number from <bos> to <eos>) loss: The XE loss