wangxuhongcn / apan Goto Github PK

Dear author of APAN,

Great work on your paper! I'm particularly interested in the model comparison between APAN and the other models, such as TGN, TGAT, JODIE

Could you please provide the code you used to produce the results in the following table?
+------------------+-------------------+-------------------+-------------------+
| Method | Wikipedia | Reddit | Alipay |
+------------------+-------------------+-------------------+-------------------+
| GAE | 74.85 (0.6) | 58.39 (0.5) | - |
| VGAE | 73.67 (0.8) | 57.98 (0.6) | - |
| GAT | 82.34 (0.8) | 64.52 (0.5) | 69.47 (0.4) |
| SAGE | 82.42 (0.7) | 61.24 (0.6) | 67.91 (0.5) |
| CIDNE | 75.89 (0.5) | 59.43 (0.6) | - |
| DyRep | 84.59 (2.2) | 62.91 (2.4) | 65.09 (1.0) |
| JODIE | 83.17 (0.5) | 59.90 (2.1) | 81.89 (0.7) |
| TGAT | 83.69 (0.7) | 65.56 (0.7) | 77.84 (0.9) |
| TGN | 88.56 (0.3) | 68.63 (0.7) | 84.01 (0.9) |
| APAN | 89.86 (0.3) | 65.34 (0.4) | 83.37 (0.7) |
+------------------+-------------------+-------------------+-------------------+

Hi,thanks for providing this nice open-sourced code!

NotImplementedError will be reported when inheriting EdgeCollator to process data and the reason I can not find.

File "/dgl/dataloading/base.py", line 230, in sample_blocks
raise NotImplementedError

Hi Xuhong,

Thanks for providing this nice open-sourced code!

In Section 4.4 of your paper (Arxiv, March version), you mentioned "the number of message passing layer is 2". Is this referring to two-layer of self-attention when aggregating from the mailbox, or two-layer when delivering the mails? In your code, it seems that both these two procedures have one-layer.

Thanks in advance.

Best,
Hongkuan

Hello, teacher, can you give me your ppt of the paper, I would like to use it during the group meeting，Thank you.

Hi Xuhong,

Thanks for providing this nice open-sourced code!

I have missed the mistake of "AttributeError: 'TemporalEdgeCollator' object has no attribute 'block_sampler'"

Thanks in advance.

Hii Xuhong, thanks a lot for your promising and effective work and codes!! I have one concern about how to implement the continual training & inference form of APAN. That is, e.g. assume the model has been trained and saved in data of several months and more data is given now. Only using the pretrained encoder and decoder model seems insufficient and unreasonable for continual training, since it also leverages the history mailbox and last-modified feature of each nodes in graph. Therefore, if I wanna train the model continuely with pretrained model, is it necessary to keep record of the final mailbox and embedding on each nodes at the end of the previous interval? Thanks in advance!! :)

🔨Work Item

Issue Description:
I am facing an error despite following the provided guide and creating a new environment with Python 3.6.13, PyTorch version 1.5.0, DGL version 0.5.2, and numpy version 1.19. The specific error I encounter is as follows:

"""from pytorch_lightning.metrics.functional import accuracy, auroc, average_precision, roc, f1
ModuleNotFoundError: No module named ‘pytorch_lightning’"""

To Reproduce:
Steps to reproduce the behavior:

Create a new environment with Python 3.6.13 using a virtual environment or conda.
Install the required packages with specific versions:
Copy code
pip install torch==1.5.0
pip install dgl==0.5.2
pip install numpy==1.19
Run the code that utilizes the pytorch_lightning module.
Observe the error mentioned above.
Expected behavior:
I expected the code to run without any errors and to find the pytorch_lightning module without any issues.

Environment:

Python version: 3.6.13
PyTorch version: 1.5.0
DGL version: 0.5.2
Numpy version: 1.19

Additional context:
I cannot proceed by installing an older version of PyTorch, DGL, or numpy, as my device requires specific versions. I seek assistance in resolving this issue while using the specified versions and considering my specific device specifications.

Thank you for your help!