https://github.com/shobrook/seqconv
Graph convolutional operator that uses a LSTM as a filter
https://github.com/shobrook/seqconv
Last synced: 3 months ago
JSON representation
Graph convolutional operator that uses a LSTM as a filter
- Host: GitHub
- URL: https://github.com/shobrook/seqconv
- Owner: shobrook
- License: mit
- Created: 2020-07-21T06:37:57.000Z (about 5 years ago)
- Default Branch: master
- Last Pushed: 2020-08-11T04:26:33.000Z (about 5 years ago)
- Last Synced: 2025-04-27T09:26:58.932Z (5 months ago)
- Language: Python
- Homepage:
- Size: 23.4 KB
- Stars: 10
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# SeqConv
`SeqConv` is a PyTorch implementation of a graph convolutional operator that uses long short-term memory (LSTM) to update node embeddings. This is useful for graph datasets where each node represents a sequence, such as a time series. Node embeddings are updated like so:
Where _φr_ and _φm_ are LSTMs ([`torch.nn.LSTM`](https://pytorch.org/docs/stable/nn.html#torch.nn.LSTM)), and _hΘ_ is a neural network. The outputs of each LSTM are the last hidden state, _hn_, rather than all the output features.
## Installation
This module can be installed with `pip`:
```bash
$ pip install seq_conv
```
## Usage
`SeqConv` is built on PyTorch Geometric and derives from the [`MessagePassing`](https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html#torch_geometric.nn.conv.message_passing.MessagePassing) class. It expects an input graph where each node's "features" is a sequence of vectors. `SeqConv`, similarly to [`NNConv`](https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html#torch_geometric.nn.conv.NNConv), also incorporates any available edge features when collecting messages from a node's neighbors.
**Parameters:**
- **in_channels** (_int_): Number of channels in the input node sequence (e.g. if each node has a sequence of vectors of size n associated with it, then in_channels = n)
- **out_channels** (_int_): Number of channels in the output node embedding
- **edge_nn** (_torch.nn.Module_): A neural network _hΘ_ that maps edge features, `edge_attr`, of shape `[-1, num_edge_features]` to shape `[-1, out_channels]`
- **aggr** (_string_, _optional_): The message aggregation scheme to use ("add", "mean", "max")
- **root_lstm** (_bool_, _optional_): If set to `False`, the layer will not add the LSTM-transformed root node features to the output
- **bias** (_bool_, _optional_): If set to `False`, the layer will not learn an additive bias
- **\*\*kwargs** (_optional_): Additional arguments for `torch.nn.LSTM`
**Example Usage:**
```python
import torch
from seq_conv import SeqConv
# Convolutional layer
conv_layer = SeqConv(
in_channels=1,
out_channels=5,
edge_nn=torch.nn.Linear(2, 5)
)
# Your input graph (see: https://pytorch-geometric.readthedocs.io/en/latest/notes/introduction.html#data-handling-of-graphs)
x = torch.randn((3, 12, 1), dtype=torch.float) # Shape is [num_nodes, seq_len, in_channels]
edge_index = torch.tensor([
[0, 1, 1, 2],
[1, 0, 2, 1]
], dtype=torch.long)
edge_attr = torch.randn((4, 2), dtype=torch.long)
# Your output graph
x = conv_layer(x, edge_index, edge_attr) # Shape is now [3, 5]
```
**To-Do:** Allow stacking of `SeqConv` layers.