Skip to content

Latest commit

 

History

History
100 lines (81 loc) · 4.83 KB

README.md

File metadata and controls

100 lines (81 loc) · 4.83 KB
Raw

LINK

Code and data for our RSS 2021 paper "Learning Instance-Level N-Ary Knowledge for Robots Operating in Everyday Environments"

drawing

Data

The raw data can be downloaded here. Unzip the data to /data.

Two python dictionaries object_data and object_instance_data are stored in the pickle file /data/LINK_dataset/object_data.pkl. We discuss the structure and content of each dictionary below.

Object Instances

Object instances are stored in object_data dictionary. Each entry in the dictionary represents an object instance, specifically an Amazon product. Each object instance has the following attributes:

- id [int]: an unique identifier assigned to this object 
- class [str]: the class of the object (e.g., cup, mug)
- name [str]: name of the Amazon product (e.g., zova Bathroom Tumbler 340ml, 3 pieces)
- price [float]: Amazon price 
- weight [float]: weight
- size [list of float]: x, y, z dimension
- material [str]: material (e.g., plastic)
- color [str]: color (e.g., white)
- transparency [str]: transparency (e.g., opaque)
- dimension [list of str]: words describing its dimension (e.g., thick and short)
- physical_property [list of str]: physical properties (e.g., fragile and hard)
- shape [list of str]: words describing its shape (e.g., holow and curved)

Situated Object Instances

Sitauted object instances are stored in object_instance_data dictionary. Each entry in the dictionary represents an object situated in a specific environmental context (e.g., a half-full coffee mug on a kitchen table). Each situated object instance has the following attributes describing its state:

- id [int]: points back to the object class
- iid [int]: an unique identifier assigned to this object instance
- room [str]: where this object instance is found (e.g., kitchen)
- state_description [str]: raw phrases and sentences from mturk workers describing the state of the object instance
- specific_place [str]: the more specific location of this object instance (e.g., on table)
- cleaniness [str]: cleaniness (e.g., dirty)
- dampness [str]: wetness (e.g., wet)
- fullness [str]: whether the object is full, specifically for containers (e.g., empty)
- temperature [str]: temperature of the object (e.g., cold)

Visualize Data

To analyze the correlations between any two object properties (e.g., dampness and purity), use the following script

python plot_heatmap.py --property_1 dampness --property_2 purity

Convert Data To Role-Value Format

object_data and object_instance_data can be combined and converted to the role-value format. In that format, each unique object is expressed as a list of role-value pairs, where each role is a property type and each value is the corresponding property. Here is one example:

[('class', 'bottle'), ('color', 'blue'), ('color', 'clear'), ('dampness', 'dry'), ('dimension', 'long'), ('dimension', 'narrow'), ('material', 'plastic'), ('purity', 'normal'), ('room', 'dining_room'), ('shape', 'irregular'), ('spatial_distribution', 'full'), ('temperature', 'room_temperature'), ('transparency', 'transparent')]

To complete the conversion, first modify the config file configs/data/data_non_repeating_10_value_negative_expanded.yaml (e.g., change paths) and then run the following script:

python main/build_role_value_data.py
  • non_repeating means that the created data split will have no test leakage.
  • 10_value_negative means 10 negative examples are sampled by randomly perturbing one value at a time for each positive example in the training set.
  • expanded means literal property types (i.e., price, weight, and size) are discretized and included in the data.

Experiments

Our transformer model and the baseline models can be evaluated on missing value prediction (i.e., predicting one missing value given all other role-value pairs of an instance).

  • Transformer (Ours)

    To run the model, modify the config file configs/transformer/run_transformer_non_repeating_10_value_negative_expanded.yaml and run the following script:

        python run_Transformer.py

  • NaLP

    To run the model, modify the config file configs/nalp/run_nalp_non_repeating_10_value_negative_expanded.yaml and run the following script:

        python run_NaLP.py

  • Frequency model

    To run the model, modify the config file configs/frequency/run_frequency_non_repeating_10_value_negative_expanded.yaml and run the following script:

        python run_frequency_model.py

  • Random

    To run the model, modify the config file configs/random/run_random_non_repeating_10_value_negative_expanded.yaml and run the following script:

        python run_random_model.py