How to Run OpenHgnn?

[andrewkrapivin]

❓ Questions and Help

Hello, we are trying to use OpenHGNN to benchmark different heterogeneous graph neural networks. However, we cannot seem to be able to run openHGNN, as there seems to be some issue with serialization. Following are the code snippets we ran in Jupyter Notebook (installing dependencies with the versions listed on the github page). torch=2.3.0 was not available when running python=3.6, so we are using python=3.11 instead.

!pip install torch==2.3.0 torchvision torchaudio torchdata==0.9.0 !pip install "dgl==2.2.1" -f https://data.dgl.ai/wheels/torch-2.3/repo.html !pip install --upgrade openhgnn==0.7 igraph rdflib lmdb ordered_set

Finally, here is the code that we ran:

from openhgnn import Experiment
from openhgnn import models

def run_experiment(model, dataset, task):
    experiment = Experiment(model=model, dataset=dataset, task=task, use_distributed=False)
    experiment.run()

run_experiment(model=models.SeHGNN, dataset="ogbn-mag", task='node_classification')
# experiment = Experiment(model='RGCN', dataset='acm4GTN', task='node_classification', gpu=-1, lr=0.05, hidden_dim=64,
#                         max_epoch=30, num_layers=3)
# experiment.run()

It appears DGL sparse matrix is not serializable:

DGL backend not selected or invalid.  Assuming PyTorch for now.

Setting the default backend to "pytorch". You can change it in the ~/.dgl/config.json file or export the DGLBACKEND environment variable.  Valid options are: pytorch, mxnet, tensorflow (all lowercase)

/usr/local/lib/python3.11/dist-packages/torchdata/datapipes/__init__.py:18: UserWarning: 
################################################################################
WARNING!
The 'datapipes', 'dataloader2' modules are deprecated and will be removed in a
future torchdata release! Please see https://github.com/pytorch/data/issues/1196
to learn more and leave feedback.
################################################################################

  deprecation_warning()
10 Mar 10:47    INFO  [Config Info]	Model: ABCMeta,	Task: node_classification,	Dataset: ogbn-mag

------------------------------------------------------------------------------
 Basic setup of this experiment: 
     model: ABCMeta    
     dataset: ogbn-mag   
     task: node_classification. 
 This experiment has following parameters. You can use set_params to edit them.
 Use print(experiment) to print this information again.
------------------------------------------------------------------------------
dataset_name: ogbn-mag
device: cpu
gpu: -1
hpo_search_space: None
hpo_trials: 100
load_from_pretrained: False
max_epoch: 1
model_name: ABCMeta
optimizer: Adam
output_dir: ./openhgnn/output/ABCMeta
patience: 1
seed: 0
use_best_config: False
use_database: False
use_distributed: False

Downloading http://snap.stanford.edu/ogb/data/nodeproppred/mag.zip

Downloaded 0.40 GB: 100%|██████████| 413/413 [00:05<00:00, 76.48it/s]

Extracting dataset/mag.zip
Loading necessary files...
This might take a while.
Processing graphs...

100%|██████████| 1/1 [00:00<00:00, 9320.68it/s]

Converting graphs into DGL objects...

100%|██████████| 1/1 [00:37<00:00, 37.47s/it]

Saving...

---------------------------------------------------------------------------

RuntimeError                              Traceback (most recent call last)

<ipython-input-4-0d11cd29a151> in <cell line: 0>()
      6     experiment.run()
      7 
----> 8 run_experiment(model=models.SeHGNN, dataset="ogbn-mag", task='node_classification')
      9 # experiment = Experiment(model='RGCN', dataset='acm4GTN', task='node_classification', gpu=-1, lr=0.05, hidden_dim=64,
     10 #                         max_epoch=30, num_layers=3)

11 frames

<ipython-input-4-0d11cd29a151> in run_experiment(model, dataset, task)
      4 def run_experiment(model, dataset, task):
      5     experiment = Experiment(model=model, dataset=dataset, task=task, use_distributed=False)
----> 6     experiment.run()
      7 
      8 run_experiment(model=models.SeHGNN, dataset="ogbn-mag", task='node_classification')

/usr/local/lib/python3.11/dist-packages/openhgnn/experiment.py in run(self)
    210             hpo_experiment(self.config, trainerflow)
    211         else:
--> 212             flow = build_flow(self.config, trainerflow)
    213             result = flow.train()
    214             if hasattr(self.config, 'line_profiler_func'):

/usr/local/lib/python3.11/dist-packages/openhgnn/trainerflow/__init__.py in build_flow(args, flow_name)
     45     if not try_import_flow(flow_name):
     46         exit(1)
---> 47     return FLOW_REGISTRY[flow_name](args)
     48 
     49 

/usr/local/lib/python3.11/dist-packages/openhgnn/trainerflow/node_classification.py in __init__(self, args)
     30         """
     31 
---> 32         super(NodeClassification, self).__init__(args)
     33         self.args.category = self.task.dataset.category
     34         self.category = self.args.category

/usr/local/lib/python3.11/dist-packages/openhgnn/trainerflow/base_flow.py in __init__(self, args)
     63         self.model = args.model
     64         self.device = args.device
---> 65         self.task = build_task(args)
     66         self.max_epoch = args.max_epoch
     67         self.optimizer = None

/usr/local/lib/python3.11/dist-packages/openhgnn/tasks/__init__.py in build_task(args)
     37     if not try_import_task(args.task):
     38         exit(1)
---> 39     return TASK_REGISTRY[args.task](args)
     40 
     41 

/usr/local/lib/python3.11/dist-packages/openhgnn/tasks/node_classification.py in __init__(self, args)
     32         super(NodeClassification, self).__init__()
     33         self.logger = args.logger
---> 34         self.dataset = build_dataset(args.dataset, 'node_classification', 
     35                                      logger=self.logger,args = args)
     36         if args.graphbolt:

/usr/local/lib/python3.11/dist-packages/openhgnn/dataset/__init__.py in build_dataset(dataset, task, *args, **kwargs)
    209         _dataset = 'grail_'+ task
    210         return DATASET_REGISTRY[_dataset](dataset, logger=kwargs['logger'],args=kwargs['args'])
--> 211     return DATASET_REGISTRY[_dataset](dataset, logger=kwargs['logger'])
    212 
    213 

/usr/local/lib/python3.11/dist-packages/openhgnn/dataset/NodeClassificationDataset.py in __init__(self, dataset_name, *args, **kwargs)
    778             self.category], split_idx["test"][self.category]
    779         self.g, self.label_dict = dataset[0]
--> 780         self.SeHGNN_g = self.mag4sehgnn(dataset)
    781         self.g = self.mag4HGT(self.g)
    782         self.label = self.label_dict[self.category].squeeze(dim=-1)

/usr/local/lib/python3.11/dist-packages/openhgnn/dataset/NodeClassificationDataset.py in mag4sehgnn(self, dataset)
    850         if not os.path.exists(diag_name):
    851             PP_rm_diag = row_norm(PP)
--> 852             th.save(PP_rm_diag, diag_name)
    853 
    854         diag_name = f'ogbn-mag_PPP_diag.pt'

/usr/local/lib/python3.11/dist-packages/torch/serialization.py in save(obj, f, pickle_module, pickle_protocol, _use_new_zipfile_serialization, _disable_byteorder_record)
    626     if _use_new_zipfile_serialization:
    627         with _open_zipfile_writer(f) as opened_zipfile:
--> 628             _save(obj, opened_zipfile, pickle_module, pickle_protocol, _disable_byteorder_record)
    629             return
    630     else:

/usr/local/lib/python3.11/dist-packages/torch/serialization.py in _save(obj, zip_file, pickle_module, pickle_protocol, _disable_byteorder_record)
    838     pickler = pickle_module.Pickler(data_buf, protocol=pickle_protocol)
    839     pickler.persistent_id = persistent_id
--> 840     pickler.dump(obj)
    841     data_value = data_buf.getvalue()
    842     zip_file.write_record('data.pkl', data_value, len(data_value))

RuntimeError: Tried to serialize object __torch__.torch.classes.dgl_sparse.SparseMatrix which does not have a __getstate__ method defined!