tkipf / c-swm

Contrastive Learning of Structured World Models

Contrastive Learning of Structured World Models

This repository contains the official PyTorch implementation of:

Contrastive Learning of Structured World Models.
Thomas Kipf, Elise van der Pol, Max Welling.
http://arxiv.org/abs/1911.12247

Abstract: A structured understanding of our world in terms of objects, relations, and hierarchies is an important component of human cognition. Learning such a structured world model from raw sensory data remains a challenge. As a step towards this goal, we introduce Contrastively-trained Structured World Models (C-SWMs). C-SWMs utilize a contrastive approach for representation learning in environments with compositional structure. We structure each state embedding as a set of object representations and their relations, modeled by a graph neural network. This allows objects to be discovered from raw pixel observations without direct supervision as part of the learning process. We evaluate C-SWMs on compositional environments involving multiple interacting objects that can be manipulated independently by an agent, simple Atari games, and a multi-object physics simulation. Our experiments demonstrate that C-SWMs can overcome limitations of models based on pixel reconstruction and outperform typical representatives of this model class in highly structured environments, while learning interpretable object-based representations.

Requirements

• Python 3.6 or 3.7
• PyTorch version 1.2
• OpenAI Gym version: 0.12.0 pip install gym==0.12.0
• OpenAI Atari_py version: 0.1.4: pip install atari-py==0.1.4
• Scikit-image version 0.15.0 pip install scikit-image==0.15.0
• Matplotlib version 3.0.2 pip install matplotlib==3.0.2

Generate datasets

2D Shapes:

python data_gen/env.py --env_id ShapesTrain-v0 --fname data/shapes_train.h5 --num_episodes 1000 --seed 1
python data_gen/env.py --env_id ShapesEval-v0 --fname data/shapes_eval.h5 --num_episodes 10000 --seed 2

3D Cubes:

python data_gen/env.py --env_id CubesTrain-v0 --fname data/cubes_train.h5 --num_episodes 1000 --seed 3
python data_gen/env.py --env_id CubesEval-v0 --fname data/cubes_eval.h5 --num_episodes 10000 --seed 4

Atari Pong:

python data_gen/env.py --env_id PongDeterministic-v4 --fname data/pong_train.h5 --num_episodes 1000 --atari --seed 1
python data_gen/env.py --env_id PongDeterministic-v4 --fname data/pong_eval.h5 --num_episodes 100 --atari --seed 2

Space Invaders:

python data_gen/env.py --env_id SpaceInvadersDeterministic-v4 --fname data/spaceinvaders_train.h5 --num_episodes 1000 --atari --seed 1
python data_gen/env.py --env_id SpaceInvadersDeterministic-v4 --fname data/spaceinvaders_eval.h5 --num_episodes 100 --atari --seed 2

3-Body Gravitational Physics:

python data_gen/physics.py --num-episodes 5000 --fname data/balls_train.h5 --seed 1
python data_gen/physics.py --num-episodes 1000 --fname data/balls_eval.h5 --eval --seed 2

Run model training and evaluation

2D Shapes:

python train.py --dataset data/shapes_train.h5 --encoder small --name shapes
python eval.py --dataset data/shapes_eval.h5 --save-folder checkpoints/shapes --num-steps 1

3D Cubes:

python train.py --dataset data/cubes_train.h5 --encoder large --name cubes
python eval.py --dataset data/cubes_eval.h5 --save-folder checkpoints/cubes --num-steps 1

Atari Pong:

python train.py --dataset data/pong_train.h5 --encoder medium --embedding-dim 4 --action-dim 6 --num-objects 3 --copy-action --epochs 200 --name pong
python eval.py --dataset data/pong_eval.h5 --save-folder checkpoints/pong --num-steps 1

Space Invaders:

python train.py --dataset data/spaceinvaders_train.h5 --encoder medium --embedding-dim 4 --action-dim 6 --num-objects 3 --copy-action --epochs 200 --name spaceinvaders
python eval.py --dataset data/spaceinvaders_eval.h5 --save-folder checkpoints/spaceinvaders --num-steps 1

3-Body Gravitational Physics:

python train.py --dataset data/balls_train.h5 --encoder medium --embedding-dim 4 --num-objects 3 --ignore-action --name balls
python eval.py --dataset data/balls_eval.h5 --save-folder checkpoints/balls --num-steps 1

Cite

If you make use of this code in your own work, please cite our paper:

@article{kipf2019contrastive,
  title={Contrastive Learning of Structured World Models}, 
  author={Kipf, Thomas and van der Pol, Elise and Welling, Max}, 
  journal={arXiv preprint arXiv:1911.12247}, 
  year={2019} 
}

c-swm's Issues

Discrepancy between the loss mentioned in the paper and GitHub

According to the paper, the negative component of the contrastive loss is the difference between the negative states (randomly sampled from embedding at timestamp t, (z_{t}~)) and the ground truth state (z_{t+1}).

However, as per the line 113 of modules.py, given no trans, you are effectively taking the difference between randomly sampled from embedding at timestamp t (z_{t}~) and z_{t} (rather than z_{t+1}).

` def contrastive_loss(self, obs, action, next_obs):

    objs = self.obj_extractor(obs)
    next_objs = self.obj_extractor(next_obs)

    state = self.obj_encoder(objs)
    next_state = self.obj_encoder(next_objs)

    # Sample negative state across episodes at random
    batch_size = state.size(0)
    perm = np.random.permutation(batch_size)
    neg_state = state[perm]

    self.pos_loss = self.energy(state, action, next_state)
    zeros = torch.zeros_like(self.pos_loss)
    
    self.pos_loss = self.pos_loss.mean()
    self.neg_loss = torch.max(
        zeros, self.hinge - self.energy(
            state, action, neg_state, no_trans=True)).mean()

    loss = self.pos_loss + self.neg_loss

    return loss

`
Thus, I feel instead of the state as the first argument of the energy function, next_state should have been the argument. Please let me know if I am misconstruing at any point.

Thanks.

Any scripts for representation visualization?

Hello, thank you for sharing the codes.
I really like the Figures 3 and 4 in your paper, and wonder if there is any scripts for reproducing these results.
Thanks!

Implementation of hinge loss does not match the paper

Hello Kipf, I find there is a discrepancy between the loss mentioned in the paper.

According to Eq(5) in paper, for negative samples, you calculate the Euclidean distance between negative state sample at timestep t and state at timestep t+1.

However, in the code below, state and neg_state are both at timestep t.

self.neg_loss = torch.max(
    zeros, self.hinge - self.energy(
        state, action, neg_state, no_trans=True)).mean()

I noticed that the same question was also asked here.

I want to know if this is a bug ? Does the discrepancy affect the final performance ?

How do you align the corresponding state representation of the manipulated object

How do you determine the corresponding state representation of the manipulated object so that you can enter the unique thermal action code into GNN along with the abstract state of the corresponding target?

Add to path in physics.py

You need to copy the following lines from data_gen/env.py into data_gen/physics.py:

# Get env directory
import sys
from pathlib import Path
if str(Path.cwd()) not in sys.path:
    sys.path.insert(0, str(Path.cwd()))

How do you determine the number objects when it changes through time.

Learn object representations and their relationship is intersting, but it is hard to understand how can you know all objcets at the begining for all examples.

Data Loading Error

When I try to run the code provided here, I end up hitting a RuntimeError here:

c-swm/train.py

Line 104 in e944b24

obs = train_loader.__iter__().next()[0]

when running your Shapes-2D build/train functions: python train.py --dataset data/shapes_train.h5 --encoder small --name shapes

Specifically, the error says: RuntimeError: DataLoader worker is killed by signal: Killed. It seems to be coming from the fact that the dataloader is having trouble multi-processing the training set loading. But when I look through your utils files, I am not seeing why this error would exist.

This same error has occurred on both a windows machine as well as a linux machine.

Here's the full trace:

Traceback (most recent call last):
  File "/home/aadharna/miniconda3/envs/cswm/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 724, in _try_get_data
    data = self.data_queue.get(timeout=timeout)
  File "/home/aadharna/miniconda3/envs/cswm/lib/python3.7/multiprocessing/queues.py", line 104, in get
    if not self._poll(timeout):
  File "/home/aadharna/miniconda3/envs/cswm/lib/python3.7/multiprocessing/connection.py", line 257, in poll
    return self._poll(timeout)
  File "/home/aadharna/miniconda3/envs/cswm/lib/python3.7/multiprocessing/connection.py", line 414, in _poll
    r = wait([self], timeout)
  File "/home/aadharna/miniconda3/envs/cswm/lib/python3.7/multiprocessing/connection.py", line 921, in wait
    ready = selector.select(timeout)
  File "/home/aadharna/miniconda3/envs/cswm/lib/python3.7/selectors.py", line 415, in select
    fd_event_list = self._selector.poll(timeout)
  File "/home/aadharna/miniconda3/envs/cswm/lib/python3.7/site-packages/torch/utils/data/_utils/signal_handling.py", line 66, in handler
    _error_if_any_worker_fails()
RuntimeError: DataLoader worker (pid 19014) is killed by signal: Killed. 

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "train.py", line 104, in <module>
    obs = train_loader.__iter__().next()[0]
  File "/home/aadharna/miniconda3/envs/cswm/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 804, in __next__
    idx, data = self._get_data()
  File "/home/aadharna/miniconda3/envs/cswm/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 771, in _get_data
    success, data = self._try_get_data()
  File "/home/aadharna/miniconda3/envs/cswm/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 737, in _try_get_data
    raise RuntimeError('DataLoader worker (pid(s) {}) exited unexpectedly'.format(pids_str))
RuntimeError: DataLoader worker (pid(s) 19012, 19014) exited unexpectedly
(cswm) aadharna@penguin:~/PycharmProjects/c-swm$ 

I'll continue poking around and update when I find the root cause.

What are the "interactions" in 2d-BoxPushing and 3d-BoxPushing?

Hi, Kipf~I really appreciate your work. I am wondering if the GNN is modeling interactions between objects.

After heading over the codes, I found that when collision happens from object A to object B, there's no bouncing back of object A and pushing forward happens to object B. So what are the interactions in 2d-boxpushing and 3d-boxingpushing experiments. The objects in those two experiments are more like independent with each other.

For 3-body physics, I understand the interactions are the forces among spheres.

Best regards.

What do you use for baseline implementation?

Do you all use unique code for baseline implementations of World Models (with AE and VAE) or do you use a publicly accessible library? Can you point me toward the library/code you used? Would be useful for duplicating results.
(@abaheti95 you'll be interested too)