rusty1s/pytorch_geometric 9206
Geometric Deep Learning Extension Library for PyTorch
PyTorch Extension Library of Optimized Scatter Operations
PyTorch Extension Library of Optimized Autograd Sparse Matrix Operations
PyTorch Extension Library of Optimized Graph Cluster Algorithms
Generative Adversarial Networks and Autoencoders for 3D Shapes
rusty1s/deep-graph-matching-consensus 112
Implementation of Deep Graph Matching Consensus in PyTorch
Source code for our AAAI paper "Weisfeiler and Leman Go Neural: Higher-order Graph Neural Networks".
rusty1s/pytorch_spline_conv 84
Implementation of the Spline-Based Convolution Operator of SplineCNN in PyTorch
ES6 RESTFul Koa2 API with Mongoose and OAuth2
Embedded Graph Convolutional Neural Networks (EGCNN) in TensorFlow
issue commentrusty1s/pytorch_geometric
torch.jit.frontend.NotSupportedError when creating jittable GNN operators（message passing）
I identified two issues:
- You cannot pass
datato the GNNforwardsincedatais not jittable. You need to pass the used tensors individually, e.g.:
def forward(self, x: Tensor, edge_index: Tensor):
...
- It looks like
Sequentialmakes some problems here. The code runs through if I replaceSequentialwithModuleList, e.g.:
self.layer_seq = nn.ModuleList()
for i in range(num_subgraph_layres):
t1='(Tensor,Tensor)->Tensor'
self.layer_seq.add_module(
GraphLayerProp(in_channels, hidden_unit).jittable(t1))
for layer in self.layer_seq:
x = layer(x, edge_index)
comment created time in 15 hours
push eventrusty1s/pytorch_geometric
commit sha 6304c7c29883cf2aaae0e8c0ffcd12c58dcbb7ed
an gat sampling example using multi-gpu training
commit sha b3ad9b8a1d5080c523314d521c34f85a74d6a496
clean up
commit sha 23a0996d40c4a4d2d519b762b2002eb406e24303
Clean up 2
commit sha caf5f57bf10f9b697b418ea7ec50594ee7a21b73
Merge pull request #1453 from maqy1995/pytorch_geometric_maqy gat sampling example using multi-gpu training
push time in 18 hours
PR merged rusty1s/pytorch_geometric
pr closed time in 18 hours
pull request commentrusty1s/pytorch_geometric
gat sampling example using multi-gpu training
Awesome :) Sorry that it took me so long to merge.
comment created time in 18 hours
push eventmaqy1995/pytorch_geometric
commit sha 23a0996d40c4a4d2d519b762b2002eb406e24303
Clean up 2
push time in 18 hours
push eventmaqy1995/pytorch_geometric
commit sha b3ad9b8a1d5080c523314d521c34f85a74d6a496
clean up
push time in 18 hours
issue commentrusty1s/pytorch_geometric
torch.jit.frontend.NotSupportedError when creating jittable GNN operators（message passing）
Can you show me an example to reproduce?
comment created time in 21 hours
issue commentrusty1s/pytorch_geometric
How to implement other GraphSAGE Aggregator functions like Max Pool?
You can install from source:
pip install git+https://github.com/rusty1s/pytorch_geometric.git
or change the aggr flag after initialization:
conv = SAGEConv(...)
conv.aggr = 'max'
comment created time in 21 hours
issue commentrusty1s/pytorch_geometric
PyTorch Geometric binaries for PPC
Sounds interesting. We currently build our binaries using TravisCI, which looks like to support PPC architectures.
comment created time in 21 hours
issue commentrusty1s/pytorch_geometric
att_r Parameter not updating in GATConv
Thanks for this issue. This is already fixed in master, see here.
comment created time in 21 hours
issue commentrusty1s/pytorch_geometric
Add properties to Dataset object
Hi and thanks for this issue. However, I'm not sure if I understand the problem exactly. What's wrong with your current "workaround"? Note that you are free to save anything you like inside your custom dataset. For example, we already do this for the GEDDataset, see here.
comment created time in 21 hours
issue commentrusty1s/pytorch_geometric
Can I have an example to use GNNExplainer for link prediction?
Sure, that can work. However, note that during inference, GraphSAGE operates on the full graph with NeighborSampler size =-1, meaning that you can use a single edge_mask for consecutive layers.
comment created time in 21 hours
issue commentrusty1s/pytorch_geometric
one error when I run function topk() on the last batch
You can use
A = SparseTensor.from_edge_index(edge_index, edge_attr, sparse_sizes=(N, N))
comment created time in a day
issue commentrusty1s/pytorch_geometric
Yes, you can do that. Alternatively, you can use torch_geometric.data.DataLoader to train in mini-batches:
loader = DataLoader([cora] * N, batch_size=...)
for data in loader:
mse_loss(...)
comment created time in a day
push eventrusty1s/pytorch_geometric
commit sha a91892f5ab03256d4de1f3b1a73090cf5d704e8a
added infer block to infer link_prediction.
commit sha d2acf3a474d3174855aaa78c2f77ebaea1f190a1
forward_all usage
commit sha ef1d82b706387ba9f91e7a55d0209518b4224328
clean up
commit sha 35e9eced32271de4e1ab805c1e223bd95123a0ce
Merge pull request #1733 from reshinthadithyan/master Addition of Inference Block in examples/link_pred.py
push time in 2 days
PR merged rusty1s/pytorch_geometric
The existing version of examples/link_pred.py doesn't have a proper inference example. Refer #1699, added a inference block with InnerProductDecoder.forward_all() to give out a probability adjacency matrix. This would give a cleaner view on usage in inference stage.
pr closed time in 2 days
pull request commentrusty1s/pytorch_geometric
Addition of Inference Block in examples/link_pred.py
Thanks a lot! I modified it a bit. In particular, I moved the encoding and decoding part to the model.
comment created time in 2 days
push eventreshinthadithyan/pytorch_geometric
commit sha ef1d82b706387ba9f91e7a55d0209518b4224328
clean up
push time in 2 days
push eventrusty1s/pytorch_geometric
MohammadAmin Khashkhashi Moghaddam
commit sha f5f200759407806c19f00b46919de3ff6eed4385
Accept custom aggregate method for convolution layers Set the aggr method only if it is not already set by the user
commit sha d0d552ce065ca46f42043fa83cb5e1f355ce8156
Merge pull request #1717 from m30m/master Accept custom aggregate method for convolution layers
push time in 2 days
PR merged rusty1s/pytorch_geometric
Set the aggr method only if it is not already set by the user
pr closed time in 2 days
pull request commentrusty1s/pytorch_geometric
Accept custom aggregate method for convolution layers
I see. That makes sense, thanks a lot :)
comment created time in 2 days
issue commentrusty1s/pytorch_scatter
This might be relevant: https://github.com/rusty1s/pytorch_geometric/issues/1593
comment created time in 2 days
issue commentrusty1s/pytorch_geometric
Yes, Eq (7) looks wrong from a dimensionality perspective. You can either append zero vectors to your central node features or use a linear projection to transform incoming messages to the same dimensionality.
IMO, you already learn the weights per feature since gate_x_j computes a weight for each feature.
comment created time in 2 days
push eventsnap-stanford/ogb
commit sha 8548e78b7f74e8cc491246d2833b4e046ebe3110
add dynamic collab experiment
commit sha 6e1424114e985dd19bad800ab7dab53a2852c267
Merge branch 'master' of github.com:snap-stanford/ogb
push time in 2 days
issue commentrusty1s/pytorch_geometric
OSError: [WinError 127] The specified procedure could not be found
This is a weird PyTorch error, and I already encountered it by myself. There are also a lot of users reporting similar problems. I personally fixed it by just setting up a fresh miniconda environment.
comment created time in 2 days
issue commentrusty1s/pytorch_geometric
Can I use GCN Conv to a weighted graph?
Yes, you can simply pass in the edge_weight to GCNConv.forward as an additional attribute.
comment created time in 2 days
issue commentrusty1s/pytorch_geometric
Can you do me a favor and try out SAGEConv instead of GCNConv and see if that succeeds?
comment created time in 2 days
issue commentrusty1s/pytorch_geometric
Not found error for torch_sparse::ptr2ind in torchscript
Thanks, that was indeed a typo, it is fixed now :) I also added this line to the CMake file, and hope that is solves your issues.
comment created time in 2 days
push eventrusty1s/pytorch_sparse
commit sha a1021ebd018cd473e49c66a294e166f834e96c8b
fix
push time in 2 days
issue commentrusty1s/pytorch_geometric
Can I have an example to use GNNExplainer for link prediction?
In general, GNNExplainer only learns a single edge mask, and thus GNNExplainer is not trivial to apply in a NeighborSampler scenario. Since GNNExplainer only operates on the L-hop neighborhood around a single node, I do not see the need to apply neighbor sampling here. As an alternative, I suggest to define a model forward function, that operates on a single edge_index and use that one for explaining nodes/links, e.g.:
def forward(self, x, edge_index):
...
def forward_with_sampling(self, x, adjs, link, n_id):
...
comment created time in 2 days
issue commentrusty1s/pytorch_geometric
one error when I run function topk() on the last batch
It's hard for me to say what went wrong solely based on the error message. Can you show me a minimal example to reproduce?
comment created time in 2 days
issue commentrusty1s/pytorch_geometric
Your first operator looks correct to me, but emb_dim should not be static (and actually depends on the output dimension of self.nn of the previous layer).
comment created time in 2 days
issue commentrusty1s/pytorch_geometric
Can you specify what you are trying to achieve?
comment created time in 2 days
push eventrusty1s/pytorch_geometric
commit sha 4b2d17e639d631984997bde57ce8a562b28cc33a
fix docs
push time in 2 days
push eventrusty1s/rusty1s.github.io
commit sha fc684bbd893bb9ad31d4e64413f58a353fb02021
update
push time in 2 days
push eventrusty1s/pytorch_geometric
commit sha b88e3823ba18c70485a5e232aed6bae89bb0cb21
don't read node features on g-plus since they are not consistent
push time in 3 days
push eventrusty1s/pytorch_geometric
commit sha ea3577a15d2eaacdcbb9f0aeeb97f6630ac207cf
fix empty edges
push time in 3 days
issue commentrusty1s/pytorch_geometric
Should be working now :) Sorry for the inconveniences!
comment created time in 4 days
push eventrusty1s/pytorch_geometric
commit sha 708e7e41b72a5bb08c05ba1fab38c387796aed66
fix gplus
push time in 4 days
issue commentrusty1s/himp-gnn
adding tests on regression tasks
That should be straigtforward to implement by training against ground-truth targets with F.mse_loss.
comment created time in 4 days
issue closedrusty1s/himp-gnn
Using PyG 1.5.0 or 1.6.1
got this error during python run:
Traceback (most recent call last): File "train_tox21.py", line 113, in <module> if val_perf > best_val_perf: TypeError: '>' not supported between instances of 'dict' and 'int'
same for Hiv:
Traceback (most recent call last): File "train_hiv.py", line 113, in <module> if val_perf > best_val_perf: TypeError: '>' not supported between instances of 'dict' and 'int'
closed time in 4 days
thegodone
issue commentrusty1s/himp-gnn
Thanks, that's indeed a bug. Fixed it.
comment created time in 4 days
push eventrusty1s/himp-gnn
commit sha 5f8573b1a6f579ae95c918bbd2a44dacf0db2e37
fix test
push time in 4 days
issue commentrusty1s/pytorch_geometric
How can I implement A times X？
Sure:
from torch_sparse import SparseTensor
A = SparseTensor.from_dense(torch.randn(100, 100))
X = torch.randn(100, 32)
out = A @ X
comment created time in 4 days
issue commentrusty1s/pytorch_geometric
That looks correct to me.
comment created time in 4 days
push eventrusty1s/pytorch_geometric
commit sha 0f5b422ce04039b97a9d92b0d93109200d5d650c
linting
push time in 4 days
issue commentrusty1s/pytorch_geometric
GNNExplainer for graph classification problems
Of course, you can use a global pooling layer. I was talking about the hierarchical pooling layer implemented via graclus and max_pool.
comment created time in 4 days
issue commentrusty1s/pytorch_geometric
Concatenation as aggregation function?
Yes, given that our edge indices are sorted based on destination nodes, you can define a custom aggregation function via aggregate, e.g.:
def aggregate(self, inputs): # [num_edges, num_features]
# Reshape to [num_nodes, k, num_features]
inputs = inputs.view(-1, k, inputs.size(-1))
...
comment created time in 4 days
push eventrusty1s/pytorch_geometric
commit sha bee454aaeeddf5af09532189e3d129577b00dae2
use indices in shortest_path
commit sha 4cf349fc2af4bcfbcbce6fd14667d571733581f1
Merge pull request #1731 from muhanzhang/master Use indices in shortest_path() in seal_link_pred.py
push time in 4 days
PR merged rusty1s/pytorch_geometric
Previously the shortest_path() calculates shortest path distances between all pairwise nodes, but we only need the shortest path distances from all nodes to the src and dst nodes. To avoid redundant computation, I specify indices in shortest_path() directly. This brings 3 times speed up on Cora.
pr closed time in 4 days
pull request commentrusty1s/pytorch_geometric
Use indices in shortest_path() in seal_link_pred.py
Thanks :)
comment created time in 4 days
issue commentrusty1s/pytorch_geometric
GNNExplainer for graph classification problems
The problem is that the pooling approach in the MNIST example is non-deterministic, and can produce a different amount of edges between different runs. Therefore, shapes between edge masks and actual used edges do not longer match. I suggest you try out a model without pooling and see if that succeeds :)
comment created time in 4 days
issue commentrusty1s/pytorch_geometric
You would need to implement the PointConv via scatter_max from the torch-scatter package. Everything else can be implemented via torchlib and torch-cluster.
comment created time in 4 days
issue commentrusty1s/pytorch_geometric
That is indeed compatible, and can be implemented in the message function via torch.cat, e.g.:
def message(self, h_j, z_j):
return torch.cat([h_j, z_j], dim=-1)
comment created time in 4 days
issue commentrusty1s/pytorch_geometric
Can I have an example to use GNNExplainer for link prediction?
I think that the GNNExplainer can be modified to the task of link prediction. You may want to add a new function explain_link to achieve that. Note that at the moment, this is restricted to full-batch subgraphs, and cannot handle multiple adjacencies.
comment created time in 4 days
issue commentrusty1s/pytorch_geometric
Could I create 2 different batches in data
You can pass the follow_batch argument to the `DataLoader, e.g.:
loader = DataLoader(dataset, batch_size, follow_batch=['posx']
comment created time in 4 days
issue commentrusty1s/pytorch_geometric
Questions about the GraphSAGE example
You can also sample the complete 2-hop neighborhood, but this is rather expensive (especially for Reddit). We apply here trick here to compute the node embeddings efficiently: Instead of sampling multiple layers for a mini-batch, we instead compute the node embeddings layer-wise. Doing this exactly two times mimics a two-layer GNN.
comment created time in 4 days
issue commentrusty1s/pytorch_geometric
Your code looks correct to me. Can you apply the suggestion and see which op fails to compute gradients?
Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).
comment created time in 4 days
issue commentrusty1s/pytorch_geometric
Not found error for torch_sparse::ptr2ind in torchscript
I added a header file to torch-sparse. Please try again :)
comment created time in 4 days
push eventrusty1s/pytorch_sparse
commit sha 2bcc13ed3dd2cac1a023601a622307b74efa4db8
added header file
push time in 4 days
issue commentrusty1s/pytorch_geometric
Citeseer dataset isolated nodes
I'm not sure what you are trying to do. What is the exact problem with reading citeseer.cites and citeseer.content?
If you want to run tkipf/pygcn with CiteSeer, it's also maybe a good idea to ask for help there.
comment created time in 4 days
push eventrusty1s/pytorch_geometric
commit sha 66ff7e880b720fc6f4b707360bdf3d19873c3d79
reset numba version
push time in 4 days
issue commentrusty1s/pytorch_geometric
Dataset not found [Reddit and PPI]
Mh, that's interesting, and I believe it has something do to with numba. Can you please try again?
comment created time in 4 days
issue closedrusty1s/pytorch_geometric
🐛 Bug
using torch_geometric.datasets.SNAPDataset with dataset name "ego-twitter" and "ego-gplus" results in an error.
To Reproduce
Steps to reproduce the behavior:
import torch_geometric.datasets as DS
data = DS.SNAPDataset("data/",name="ego-gplus")
<!-- If you have a code sample, error messages, stack traces, please provide it here as well. -->
Expected behavior
e.g.: the "ego-twitter" configuration would lead to:
KeyError: 83943787
Environment
- OS: Linux
- Python version: 3.8.3
- PyTorch version: 1.6.0
- CUDA/cuDNN version: 10.1
- GCC version:
- Any other relevant information:
Additional context
<!-- Add any other context about the problem here. -->
closed time in 7 days
Magolor
issue commentrusty1s/pytorch_geometric
Fixed in master, see https://github.com/rusty1s/pytorch_geometric/commit/c56a50eb2f17a41c4b52d837eb66a44d8efbfee5. Thanks for reporting!
comment created time in 7 days
issue closedrusty1s/pytorch_geometric
Dataset not found [Reddit and PPI]
Hi, The dataset for cluster examples (cluster_gcn_reddit.py and cluster_gcn_PPI.py) throws "HTTP Error 404: Not Found" For reddit, it tries to download from Downloading https://s3.us-east-2.amazonaws.com/dgl.ai/dataset/reddit.zip Has the dataset moved, or I am doing something wrong? Thanks.
closed time in 7 days
snash4
push eventrusty1s/pytorch_geometric
commit sha c56a50eb2f17a41c4b52d837eb66a44d8efbfee5
fix idx numerical issues
push time in 7 days
issue commentrusty1s/pytorch_geometric
Working with large graph for node prediction!
Since your graphs are rather dense with a rather low number of nodes, I suggest you look into the newly introduced SparseTensor format, see here.
If you want to make use of mini-batch training, we provide several options:
comment created time in 7 days
issue commentrusty1s/pytorch_geometric
There are no C++ examples of PyTorch Geometric. PyG is purely Python-based. Nonetheless, with release 1.6.0, we support TorchScript, which should help you to run your models in non-Python environments. You can find some basic examples here and a short tutorial here.
comment created time in 7 days
issue commentrusty1s/pytorch_geometric
Citeseer dataset isolated nodes
You can find the same pre-processing technique as in tkipf/gcn here.
comment created time in 7 days
issue commentrusty1s/pytorch_geometric
Not found error for torch_sparse::ptr2ind in torchscript
That looks good to me. Does that work for you?
comment created time in 7 days
issue commentrusty1s/pytorch_geometric
GEDDataset num_classes request leads to AttributeError
You can access them via data.x.
comment created time in 7 days
issue commentrusty1s/pytorch_geometric
How to get value pass through using knn?
That's correct. knn computation has now backward computation since its output is discrete. I'm not sure what you are trying do to, but this should not be the reason for any errors you are receiving during back-propagation.
comment created time in 7 days
issue commentrusty1s/pytorch_geometric
Can you show me your model to reproduce this?
comment created time in 7 days
push eventrusty1s/pytorch_sparse
commit sha a9d4d4695fb58aa0a2696443ab8f9dd1d068c19c
typo
push time in 7 days
issue commentrusty1s/pytorch_geometric
Metis [segmentation Fault on ClusterData ]
Can you try to make your graph undirected and see if that fixes the issue?
comment created time in 7 days
push eventrusty1s/pytorch_geometric
commit sha 68ca38e28c1b86cf267fab8499c5f3e52d8ec410
fix broken urls
push time in 7 days
issue commentrusty1s/pytorch_geometric
Dataset not found [Reddit and PPI]
This is now fixed in master. You can simply re-install via:
pip install git+https://github.com/rusty1s/pytorch_geometric.git
Thanks for reporting!
comment created time in 7 days
issue commentrusty1s/pytorch_sparse
No, wheels are not uploaded to PyPI since different CUDA versions require different wheels (which is not covered by PyPI).
comment created time in 8 days
issue commentrusty1s/pytorch_geometric
[Roadmap] PyTorch Geometric 1.6.1 and 1.7.0
cugraph conversions is a good idea, I added it to the roadmap. The SuperpixelDataset is already fixed in master.
comment created time in 8 days
issue commentrusty1s/pytorch_geometric
Not found error for torch_sparse::ptr2ind in torchscript
For executing in a C++ environment, please ensure that you also have the C++ extensions installed, see here.
comment created time in 8 days
issue commentrusty1s/pytorch_geometric
Seems to be a bug. I will look into it.
comment created time in 8 days
pull request commentrusty1s/pytorch_geometric
Accept custom aggregate method for convolution layers
Thanks for the PR. While I agree that the aggr keyword is missing for a few operators, I feel like allowing the user to modify the aggregation type for any kind of operator is not particular useful, e.g., for GCNConv or GATConv which are defined to perform weighted mean normalization.
comment created time in 8 days
issue commentrusty1s/pytorch_geometric
Understanding examples/link_pred.py
You can use the Decoder to reconstruct such probabilistic adjacency matrix.
comment created time in 8 days
issue commentrusty1s/pytorch_geometric
The idea would be to "scatter" pixel feature information of a superpixel into a single representation. Consider that you have a "mask" idx holding indices to which superpixel a pixel belongs. Then you can do the following to train your model in an end-to-end fashion:
idx = ... # [height, width] holding indices from 0 to N - 1 superixels
img = ... # [height, width, num_features]
node_x = scatter(img.view(-1, num_features), idx.view(-1), dim=0, dim_size=N, reduce="mean")
comment created time in 8 days
issue commentrusty1s/pytorch_geometric
My guess is that, data objects of different graphs may have different feature dimensionalities, e.g., data1.x.size(-1) == data2.x.size(-1) and data1.edge_attr.size(-1) == data2.edge_attr.size(-1). Can you verify that this is the case?
comment created time in 8 days
issue commentrusty1s/pytorch_geometric
Message passing with more than two nodes
Are you referring to hypergraphs? This is currently not really covered by the MessagePassing interface, so you may need to write logic for that by yourself.
comment created time in 8 days
push eventrusty1s/pytorch_geometric
commit sha e3fb65d2703fe5ca073a255d956e41364f8d89ef
Change accuracy return type to float
commit sha ffa52fc77e622044d96282ee3e28dc9776bb39d0
Merge pull request #1712 from LukasZahradnik/master Change accuracy return type to float
push time in 10 days
PR merged rusty1s/pytorch_geometric
While going through the documentation, I noticed that torch_geometric.utils.accuracy has wrong return type (int).
pr closed time in 10 days
push eventrusty1s/pytorch_geometric
commit sha ee17659ebc16d17e778c00bc1195d386b44cdff4
Fix typo in total graphs in ZINC subset
commit sha 50af85b0e8fc4741fa09ebdc260b7e586666b723
Merge pull request #1711 from vijaydwivedi75/master Fix typo in total graphs in ZINC subset
push time in 11 days
PR merged rusty1s/pytorch_geometric
The ZINC subset used in Benchmarking GNNs has 12,000 molecular graphs - 10,000 train/1,000 val/1,000 test.
PR for this minor correction.
pr closed time in 11 days
pull request commentrusty1s/pytorch_geometric
Fix typo in total graphs in ZINC subset
Yeah, thanks :)
comment created time in 11 days
push eventrusty1s/pytorch_geometric
commit sha 9faf964400d08d9345548d95e183551dfb38ec1c
typo
push time in 12 days
push eventrusty1s/pytorch_geometric
commit sha f7c0591007fdf8cf7faac16c0b865900318dde0a
Update README.md
commit sha b2429c0c7d0d3900f74f98b81e42773d972f911e
Update autoencoder.py
commit sha 3bd9fde4c63e252bc8e967515eca2f1664aca6fe
Update autoencoder.py
commit sha f11eabc65072d34b431cc85419c27c0b0d00950c
clean up
commit sha fb1d3f989688e4ca6989350dc58638f574d98e52
Merge pull request #1656 from GuillaumeSalha/linear_gae Linear Graph AE and VAE
push time in 12 days
PR merged rusty1s/pytorch_geometric
Dear @rusty1s,
Once again, congratulations to you and your collaborators for this amazing work!
I am the lead author of the article Keep It Simple: Graph Autoencoders Without Graph Convolutional Networks, presented at the NeurIPS 2019 workshop on Graph Representation Learning. In this article, as well as in an extended conference version recently accepted at ECML-PKDD 2020, we studied simpler versions of Graph Autoencoders (AE) and Graph Variational Autoencoders (VAE), referred to as Linear Graph AE and VAE.
In a nutshell, we replaced the multi-layer GCN encoders from standard models (Kipf and Welling 2016) by simpler linear models w.r.t. the adjacency matrix of the graph (Z = AW, or Z = AXW in case of node-level features X), involving a unique weight matrix to tune and no activation function. We showed that they often appear as effective and simpler alternatives to GCN-based models, for various graphs including the widely used Cora, Citeseer and Pubmed citation networks, achieving similar or even better performances on downstream tasks such as link prediction or node clustering. In our work, we also tried to provide insights on when to use multi-layer GCN encoders and when, on the contrary, these encoders might actually be unnecessarily complex.
We released a TensorFlow implementation of Linear Graph AE and VAE a few months ago. However, some users recently asked for a PyTorch one, and I think that your pytorch-geometric project is the best place to include it.
Therefore, I am opening this PR to propose a modification of examples/autoencoder.py, aiming at including a working example of linear graph AE and VAE with application to link prediction, building upon the existing GAE and VGAE classes. With the proposed implementation, I manage to reach better results than standard GAE/VGAE on the provided Cora, Citeseer and Pubmed datasets from the example file, while being faster. So, I sincerely think this can be of interest for some future users. I also updated the README file accordingly.
Best,
Guillaume
pr closed time in 12 days
pull request commentrusty1s/pytorch_geometric
Thank you :)
comment created time in 12 days
push eventGuillaumeSalha/pytorch_geometric
commit sha f11eabc65072d34b431cc85419c27c0b0d00950c
clean up
push time in 12 days
push eventrusty1s/pytorch_geometric
commit sha 26866fcc3516a17ee2f38504afd9e82901c1d287
Update softmax.py I propose a new version of softmax function in Utils module. The version can reduce a little running time by 3 times, ^ _ ^. ```python from torch import Tensor from torch_scatter import scatter import torch from time import time # set N = 4000 src = torch.rand(N*(N+1)//2) index = [] for i in range(N): tmp = [i]*(i+1) index = index+tmp index = torch.tensor(index,dtype=torch.long ) # norm t1 = time() out = src out = out.exp() out_sum = scatter(out, index, dim=0, dim_size=N, reduce='sum')[index] t2 = time() # local max local_out = src - scatter(src, index, dim=0, dim_size=N, reduce='max')[index] local_out = local_out.exp() local_out_sum = scatter(local_out, index, dim=0, dim_size=N, reduce='sum')[index] t3 = time() # global max global_out = src - src.max() global_out = global_out.exp() global_out_sum = scatter(global_out, index, dim=0, dim_size=N, reduce='sum')[index] t4 = time() print('norm', t2 - t1) print('local softmax', t3 - t2) print('global softmax', t4 - t3) a = out / (out_sum + 1e-16) b = local_out / (local_out_sum + 1e-16) c = global_out / (global_out_sum + 1e-16) eps = 0.0000001 print((torch.abs(a-c)>eps).sum()) print((torch.abs(a-b)>eps).sum()) # ============================================================================= # result # norm 0.04488039016723633 # local softmax 0.1874983310699463 # global softmax 0.06482648849487305 # tensor(0) # tensor(0) # ============================================================================= ```
commit sha 910817862dd01c539e6ad32930fe3d528f5487d2
clean up
commit sha 421fd37ac54c0220134e0f80b29ea7038982ca7c
fix test
commit sha b59619a7980da7edcc5d74f0065fe3b26ecc543c
Merge pull request #1707 from Book1996/patch-1 Update softmax.py
push time in 12 days
PR merged rusty1s/pytorch_geometric
I propose a new version of softmax function in Utils module. The version can reduce a little running time by 3 times, ^ _ ^.
from torch import Tensor
from torch_scatter import scatter
import torch
from time import time
# set
N = 4000
src = torch.rand(N*(N+1)//2)
index = []
for i in range(N):
tmp = [i]*(i+1)
index = index+tmp
index = torch.tensor(index,dtype=torch.long )
# norm
t1 = time()
out = src
out = out.exp()
out_sum = scatter(out, index, dim=0, dim_size=N, reduce='sum')[index]
t2 = time()
# local max
local_out = src - scatter(src, index, dim=0, dim_size=N, reduce='max')[index]
local_out = local_out.exp()
local_out_sum = scatter(local_out, index, dim=0, dim_size=N, reduce='sum')[index]
t3 = time()
# global max
global_out = src - src.max()
global_out = global_out.exp()
global_out_sum = scatter(global_out, index, dim=0, dim_size=N, reduce='sum')[index]
t4 = time()
print('norm', t2 - t1)
print('local softmax', t3 - t2)
print('global softmax', t4 - t3)
a = out / (out_sum + 1e-16)
b = local_out / (local_out_sum + 1e-16)
c = global_out / (global_out_sum + 1e-16)
eps = 0.0000001
print((torch.abs(a-c)>eps).sum())
print((torch.abs(a-b)>eps).sum())
# =============================================================================
# result
# norm 0.04488039016723633
# local softmax 0.1874983310699463
# global softmax 0.06482648849487305
# tensor(0)
# tensor(0)
# =============================================================================
pr closed time in 12 days