tkipf/c-swm

Contrastive Learning of Structured World Models

c-swm's Issues

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()))

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)

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?

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 ?

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.

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.

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!

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.