yzhou359 / 3dindoor-scenegraphnet Goto Github PK

It seems that the link to the sgn-data-train.zip file has been invalid. Could you please upload it again ?

Here is the sample data we use.
bedroom_data.zip

This is the command I run.

python main.py --name matterport_train --train_cat --room_type bedroom --num_train_rooms 200 --num_test_rooms 22

I'm getting the following error.

Traceback (most recent call last):
  File "main.py", line 63, in <module>
    M.train(epoch)
  File "/home/fcr/research/3DIndoor-SceneGraphNet/SceneGraphNet/train.py", line 481, in train
    self._training_pass(self.valid_rooms_train, epoch, is_training=True)
  File "/home/fcr/research/3DIndoor-SceneGraphNet/SceneGraphNet/train.py", line 280, in _training_pass
    subtree_to_leaf_path = self.find_root_to_leaf_node_path(node_list, cur_node=sub_tree_root_node)
  File "/home/fcr/research/3DIndoor-SceneGraphNet/SceneGraphNet/train.py", line 103, in find_root_to_leaf_node_path
    child_node_list = self.find_root_to_leaf_node_path(node_list, child)
  File "/home/fcr/research/3DIndoor-SceneGraphNet/SceneGraphNet/train.py", line 103, in find_root_to_leaf_node_path
    child_node_list = self.find_root_to_leaf_node_path(node_list, child)
  File "/home/fcr/research/3DIndoor-SceneGraphNet/SceneGraphNet/train.py", line 108, in find_root_to_leaf_node_path
    child_node_list = self.find_root_to_leaf_node_path(node_list, key)
  File "/home/fcr/research/3DIndoor-SceneGraphNet/SceneGraphNet/train.py", line 108, in find_root_to_leaf_node_path
    child_node_list = self.find_root_to_leaf_node_path(node_list, key)
  File "/home/fcr/research/3DIndoor-SceneGraphNet/SceneGraphNet/train.py", line 108, in find_root_to_leaf_node_path
    child_node_list = self.find_root_to_leaf_node_path(node_list, key)
  [Previous line repeated 991 more times]
  File "/home/fcr/research/3DIndoor-SceneGraphNet/SceneGraphNet/train.py", line 97, in find_root_to_leaf_node_path
    len(node_list[cur_node]['support']) > 0):
RecursionError: maximum recursion depth exceeded in comparison

Can you help me debug this issue? I'm not sure what's wrong. I followed the instructions on the data_structurees.md document, but maybe there are things that don't line up.

Hi, i find in your dataset, the node's 'rotation' field is not same as 'transform' field in the original dataset. Did you know what's the meaning of this field and how to use it? If you know, please tell me. Thank you!

Hi Yang,

Thanks for providing the well-written code, which I learnt a lot from. Just would like to make an enquiry about the gpu memory usage & training speed from your side, as I noticed 0 gpu memory consumed (are they running in cpu???) in my machine (with slow training speed). Could you please give me your confirmation on whether this is a normal case caused by TorchFold? Thanks in advance!

In the data_structure readme, this is the definition of co-occurence:

co-occurrence list : a list of node's ids which are in neighbor of this node, especially for root and wall nodes (notice that this is not the next-to relation mentioned in the paper)

What is the definition of neighborhood? I thought that co-occurence is defined as objects co-occurring in a scene with a reference object in the same scene. However, based on the diagram in the readme, the walls are chosen to be connected to a subset of objects.

How do we replicate this structure if we are working with other datasets?

I was just wondering if you have code to support iterative scene synthesis. Do you just create a bunch of trees which represent points of the room and do multiple inferences on that? Is there a clearer approach that can just take in an input spot, given a scene?

Hello @yzhou359,

Thank you for sharing the code and data.

I am trying to train the SceneGraphNet model by running the command you have given on the GitHub page. I am using the 'torchfold.py' file that you have shared. Still, I got the following error after training for a few hours:

TRAIN JOB_NAME 6 : (3033/5000:11.0) CAT Loss: 2.8641, Acc_1: 1.0000, Acc_3: 1.0000, Acc_5: 1.0000,Dim Loss: 96.30447665, dim acc: 1.00

Traceback (most recent call last):
File "main.py", line 61, in <module>

File "/3DIndoor-SceneGraphNet/SceneGraphNet/train.py", line 481, in train
self._training_pass(self.valid_rooms_train, epoch, is_training=True)

File "/3DIndoor-SceneGraphNet/SceneGraphNet/train.py", line 306, in _training_pass
rand_path_fold, rand_path_node_name_order = self.model.encode_tree_fold(enc_fold, sub_node_list, rand_path, opt_parser)

File "/3DIndoor-SceneGraphNet/SceneGraphNet/model.py", line 415, in encode_tree_fold
encode_node(node_list, leaf_node=tree_leaf_node, step=i)

File "/3DIndoor-SceneGraphNet/SceneGraphNet/model.py", line 334, in encode_node
to_torch(dis_feat))

File "/torchfold/torchfold.py", line 96, in add
if args not in self.cached_nodes[op]:

RuntimeError: bool value of Tensor with more than one value is ambiguous

Could you please help me resolve this error?

Thank you,
Supriya

Hi,

Does SUNCG Dataset provide images corresponding to each point cloud scene?

Hello,
We were trying to train the network on your original code, but we have some trouble that the training time is so long. It takes about 2 hours per epoch using bedroom dataset on a GeForce 1080Ti GPU or a GeForce 2060 GPU.
I'm wondering if you could tell me how many GPU you used for the training? And what about the batchsize and the learning rate?

Hello, can you please release your script to create the scene graph from suncg? it would be really helpful :)
thanks!

Hi, could you give some hints on how to visualize the scene after prediction as in Figure 7 in the paper? That would be very helpful for analyzing the prediction results, thank you.

When is it uploaded?

I want to study this code :)