Attempting a Dynamic Graph Temporal signal, using your example I hope to extend it and make it work. Thanks for your work these videos are incredibly useful
Great video, very useful! Could you maybe upload a video using temporal data for node classification tasks as there seems to be none available? Thank you!
Great video. Amazing stuff. I have a query. In this use case, it is assumed the distance-based calculations formulate the edge index, and hence, it is constant. How we should proceed if the edges/edge indexes change for every time snapshot.
Hi, I want to use a TGCN to predict the number of incoming requests for each cellular tower in a specific area, using neighboring antennas to improve performance and historical data about the number of requests in that node. The problem is about multivariate and multisteps time series, because I take 3 values and I want to predict n x 3 values in the future. I have a dataset for each phone antenna that contains within it a timestamp (time of request arrival) and 3 other columns, where each column represents the incoming request for a particular feature. The frequency is 15 minutes, so for each antenna I have 3 values every 15 minutes. I already have the adjacency matrix, so I know which antennas are close to each other. My question is this: how do I create a single dataset on which to train a TGCN if I have several datasets (one for each antenna) of the same length (start and end times are the same for all of them)?
Hi! I guess you have to merge the individual antenna datasets into one large dataset. Basically each antenna is one node in your final graph - therefore you have to put the nodes together by building a full graph. Simply assign an ID to each tabular dataset and then do a merge, so that you have all antennas for each timestep. If you have further problems you can also send me an email to deepfindr@gmail.com with attached data. :)
Great videos. Thank you so much, they are very helpful. I do have several questions though. In another video of yours talking about how to include edge features in PyTorch Geometric, you mentioned that models that take 'edge_attr' as an argument can take multi-dimension edge features. Does that also apply to PyTorch Geometric Temporal? For example, you have 3 edge features in this video, what model would you use that can take 3 edge features in Geometric Temporal? I looked into the models included in the Geometric Temporal but not quite sure which one can use multi-dimension edge features since all of them only take 'edge_weight' as an argument.
Hi! Yes you are right, usually only edge_weight is provided. One thing you can do is copy the code of a model, for example this one: pytorch-geometric-temporal.readthedocs.io/en/latest/_modules/torch_geometric_temporal/nn/recurrent/dcrnn.html?highlight=Edge_attr And put it into a separate file. Then you adjust the forward function to also take edge_attr. In the link above, the dcrnn doesn't take edge features, but the dconv layer does, so it should be straight forward to add this functionality. In general you can extend any model to also use edge attr, as discussed here: github.com/benedekrozemberczki/pytorch_geometric_temporal/issues/161
Thank you for this great video. Do you have any good resources about "temporal convolution" implementation for temporal feature extraction using PyTorch/Pytorch-geometric?
Great video !!! I love it ! Can you make a video on Heterogeneous graph with edge attributes with pyGeometric ? Many tutorial exists with homogenous graph but there is lack for heterogenous graph (specially with edge attributes ..)
Great Video! I have a general question regarding the saved graph dataset in 'data.pt ' . could you please help me with how I can retrieve the saved data. Since I saved the data using InMemoryDataset provided by PYG, I wonder how I can load them now?
Hi! Simply call the load function. Have you seen the custom dataset video in my GNN series? There I show how to load the data. This happens in the get function github.com/deepfindr/gnn-project/blob/main/dataset.py
@@DeepFindr Yeah I did that and did StackOverflow too (a lot :( ), I dunno why it is having this issue, okay I will try something more ........ Thanks for the response,, do you have some discord channel or Slack ? would love to be a part of community
![ตีสิบเดย์ [FULL] | เปิดใจ "อาจารย์ เบียร์ ฅนตื่นธรรม" นักสนทนาธรรมสายดุดัน!](http://i.ytimg.com/vi/vqaBI_mnoDk/mqdefault.jpg)
Attempting a Dynamic Graph Temporal signal, using your example I hope to extend it and make it work. Thanks for your work these videos are incredibly useful
Great, just what needed
Amazing stuff buddy! Thank you!
Great video, very useful! Could you maybe upload a video using temporal data for node classification tasks as there seems to be none available? Thank you!
Thank you for such an informative video. please create a video about the converting the dataset of trajectory prediction task for temporal GNNs
great as always!
Great video. Amazing stuff. I have a query. In this use case, it is assumed the distance-based calculations formulate the edge index, and hence, it is constant. How we should proceed if the edges/edge indexes change for every time snapshot.
I'm making a model from sx-stackoverflow of SNAP, it only has source target and timestamp, no clue how to make a dataset for TGN from that
please make a video on, how we can convert SMILES notations to graph .👍👍
why we need additional_edge_index?
Thanks for the video! Do you know of any guide that shows how to use landmarks extracted from videos using mediapipe as a temporal graph?
hi ! thanks to your nice tutorial, could u make the 'biparitite manner' graph dataset too?
Hi, I want to use a TGCN to predict the number of incoming requests for each cellular tower in a specific area, using neighboring antennas to improve performance and historical data about the number of requests in that node. The problem is about multivariate and multisteps time series, because I take 3 values and I want to predict n x 3 values in the future. I have a dataset for each phone antenna that contains within it a timestamp (time of request arrival) and 3 other columns, where each column represents the incoming request for a particular feature. The frequency is 15 minutes, so for each antenna I have 3 values every 15 minutes. I already have the adjacency matrix, so I know which antennas are close to each other. My question is this: how do I create a single dataset on which to train a TGCN if I have several datasets (one for each antenna) of the same length (start and end times are the same for all of them)?
Hi! I guess you have to merge the individual antenna datasets into one large dataset. Basically each antenna is one node in your final graph - therefore you have to put the nodes together by building a full graph. Simply assign an ID to each tabular dataset and then do a merge, so that you have all antennas for each timestep.
If you have further problems you can also send me an email to deepfindr@gmail.com with attached data. :)
@@DeepFindr Hi thank you for your help but I think that I didn't understand so much. I've send you an email, thank you so much.
Great videos. Thank you so much, they are very helpful. I do have several questions though. In another video of yours talking about how to include edge features in PyTorch Geometric, you mentioned that models that take 'edge_attr' as an argument can take multi-dimension edge features. Does that also apply to PyTorch Geometric Temporal? For example, you have 3 edge features in this video, what model would you use that can take 3 edge features in Geometric Temporal? I looked into the models included in the Geometric Temporal but not quite sure which one can use multi-dimension edge features since all of them only take 'edge_weight' as an argument.
Hi! Yes you are right, usually only edge_weight is provided.
One thing you can do is copy the code of a model, for example this one: pytorch-geometric-temporal.readthedocs.io/en/latest/_modules/torch_geometric_temporal/nn/recurrent/dcrnn.html?highlight=Edge_attr
And put it into a separate file.
Then you adjust the forward function to also take edge_attr. In the link above, the dcrnn doesn't take edge features, but the dconv layer does, so it should be straight forward to add this functionality.
In general you can extend any model to also use edge attr, as discussed here: github.com/benedekrozemberczki/pytorch_geometric_temporal/issues/161
Thank you for this great video. Do you have any good resources about "temporal convolution" implementation for temporal feature extraction using PyTorch/Pytorch-geometric?
Hi! I have a video about spatio-temporal GNNs if that is what you are looking for :)
Great video !!! I love it ! Can you make a video on Heterogeneous graph with edge attributes with pyGeometric ?
Many tutorial exists with homogenous graph but there is lack for heterogenous graph (specially with edge attributes ..)
Hi! Did you have a look at the previous video to this one? :)
There I also quickly talk about heterogeneous graphs and have an example.
@@DeepFindr Yes I did ! But my sentence was about heterogeneous architecture and model training !
Ahh ok :) yes I note it down. Maybe in the future :)
Have you seen the PyG documentation of this topic? It's pretty intuitive :?
Great Video! I have a general question regarding the saved graph dataset in 'data.pt ' . could you please help me with how I can retrieve the saved data. Since I saved the data using InMemoryDataset provided by PYG, I wonder how I can load them now?
Hi! Simply call the load function. Have you seen the custom dataset video in my GNN series? There I show how to load the data. This happens in the get function
github.com/deepfindr/gnn-project/blob/main/dataset.py
Thanks!
Thank you!!
Thank you
Hi, I was experimenting with the colab and I n the Extract Labels and Build Dataset Section I am getting an error on my local env: "TypeError: '
Hi! Did you convert your date string into a datetime object? Thats most likely the problem
@@DeepFindr Yeah I did that and did StackOverflow too (a lot :( ), I dunno why it is having this issue, okay I will try something more ........
Thanks for the response,, do you have some discord channel or Slack ? would love to be a part of community
Unfortunately I don't have time to maintain one :(
Did you use the strptime function to convert your string? Also try to double check if the conversion works using type()
@@DeepFindr Yeah I guess that's what is causing the problem, I will try to do something else for the string conversion
gerat ,👽
Great video but the background music is loud
.