SAILx

This repo contains code to generate artificial data described in: A new dataset and model for learning to understand navigational instructions

https://arxiv.org/abs/1805.07952

To generate the fixed dataset:

sh sailx.sh

To generate the data used in the efficiency experiments:

sh tasks.sh

You can generate data for specific tasks by using the generatedata.jl. It will create a folder for each subtask and generate instructions.json and corresponding maps.json.

Example: julia generate_data.jl --num 15000 --folder ../unique_sailx/ --unique --tasks turn_to_x --seed 123789

--num : number of instances
--tasks: list of the tasks
--folder: parent folder to save the data
--seed: random seed
--unique: the combination of the instruction and corresponding path (including the configuration of visual properties) is unique for each instance
--ratio: default is [0.0]. If you want to split data into train, dev and test splits, give the ratio (e.g 0.7 0.15 0.15)
--ofolder: If the ratio is given, then --folder argument is used to input folder. ofolder argument is used as the parent folder to save the data

List of possible tasks:

• turn_to_x
• move_to_x
• combined_12 (sample from [turn_to_x, move_to_x])
• turn_and_move_to_x
• lang_only
• combined_1245 (sample from [turn_to_x, move_to_x, turn_and_move_to_x, lang_only])
• move_until
• orient
• describe
• move_vis_turn_lang
• turn_vis_move_lang
• move_lang_turn_vis
• turn_lang_move_vis
• move_vis_turn_vis
• turn_vis_move_vis
• any_combination (sample from [move_vis_turn_lang, turn_vis_move_lang, move_lang_turn_vis, turn_lang_move_vis, move_vis_turn_vis, turn_vis_move_vis])
• norestriction

Instruction:

id : id
fname : the file name
text : tokenized version of the instruction
map : the name of the map
path : a list of (x,y,orientation) tuples

Map:

name : randomly generated name
nodes : A dictionary where keys are the locations as (x,y) tuples and values are ids of items
edges : A dictionary as (x1,y1) => (x2, y2) => [wall id, floor id],
    where (x1, y1) and (x2, y2) are nodes and [wall id, floor id] represents the wall paintings and flooring. 

Ids of attributes:

Items = Dict("stool" => 1, "chair" => 2, "easel" => 3,
    "hatrack" => 4, "lamp" => 5, "sofa" => 6, "" => 7)

Walls = Dict("butterfly" => 1, "fish" => 2, "tower" => 3)

Floors = Dict("blue" => 1, "brick" => 2, "concrete" => 3, "flower" => 4,
    "grass" => 5, "gravel" => 6, "wood" => 7, "yellow" => 8)

Dependencies (Julia Packages):

Logging
ArgParse
JLD
JSON
DataStructures