hin2vec_pytorch

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

hin2vec_pytorch

★ 19 Open GitHub ↗

A pytorch implementation of HIN2vec

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# HIN2vec_pytorch
A pytorch implementation of [HIN2vec](https://github.com/csiesheep/hin2vec)

---

07/13/2021 add binary activation function

---

>*HIN2Vec* learns distributed representations of nodes in heterogeneous information networks (HINs) by capturing the distiguishing metapath relationships between nodes. 
Please refer the paper [here](https://dl.acm.org/citation.cfm?doid=3132847.3132953).

Please find more details in [the original repo](https://github.com/csiesheep/hin2vec)

### Requirements
- python3.6+ 
- pytorch
- numpy
- pandas
- networkx

### Some differences with the original implementation

- use Adam instead of SGD to optimize weights

I find Adam much more effective in training.

- some tricks not implemented
  - sampling the same type of nodes
  - learning rate decay

- may be more robust comparing to the original implementation

Maybe there is something wrong about my parameter setting, but when I do some experiments using the original code,
the program won't continue with `window` set to more than 4. This pytorch code works fine.

- this implementation is slower (It takes about 10 times of the time)

But is surly more easier to understand and modify. (And I would try to use a bigger window size to get comparable or 
better results.)

### Usage
create your own edge.csv referring to `demo_data.csv`

**The input graph is assumed to be undirected by default.**

create a main.py in the project folder, copy the following code and modify it.

```python
    import torch
    import pandas as pd
    import torch.nn as nn
    import torch.optim as optim
    from torch.utils.data import DataLoader
    
    from walker import load_a_HIN_from_pandas
    from model import NSTrainSet, HIN2vec, train

# set method parameters
    window = 4
    walk = 10
    walk_length = 300
    embed_size = 100
    neg = 5
    sigmoid_reg = True 
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    print(f'device = {device}')

# set dataset [PLEASE USE YOUR OWN DATASET TO REPLACE THIS]
    demo_edge = pd.read_csv('./demo_data.csv', index_col=0)

edges = [demo_edge]

print('finish loading edges')

# init HIN
    hin = load_a_HIN_from_pandas(edges)
    hin.window = window

dataset = NSTrainSet(hin.sample(walk_length, walk), hin.node_size, neg=neg)

hin2vec = HIN2vec(hin.node_size, hin.path_size, embed_size, sigmoid_reg)

# load model
    # hin2vec.load_state_dict(torch.load('hin2vec.pt'))

# set training parameters
    n_epoch = 10
    batch_size = 20
    log_interval = 200

data_loader = DataLoader(dataset, batch_size=batch_size, shuffle=True)
    optimizer = optim.AdamW(hin2vec.parameters())  # 原作者使用的是SGD？ 这里使用AdamW
    loss_function = nn.BCELoss()

for epoch in range(n_epoch):
        train(log_interval, hin2vec, device, data_loader, optimizer, loss_function, epoch)

torch.save(hin2vec, 'hin2vec.pt')

# set output parameters [the output file is a bit different from the original code.]
    node_vec_fname = 'node_vec.txt'
    # path_vec_fname = 'meta_path_vec.txt'
    path_vec_fname = None
    
    print(f'saving node embedding vectors to {node_vec_fname}...')
    node_embeds = pd.DataFrame(hin2vec.start_embeds.weight.data.numpy())
    node_embeds.rename(hin.id2node).to_csv(node_vec_fname, sep=' ')

if path_vec_fname:
        print(f'saving meta path embedding vectors to {path_vec_fname}...')
        path_embeds = pd.DataFrame(hin2vec.path_embeds.weight.data.numpy())
        path_embeds.rename(hin.id2path).to_csv(path_vec_fname, sep=' ')
       
    # save model
    # torch.save(hin2vec.state_dict(), 'hin2vec.pt')
```

### About `data_demo.csv`
In this repository, the Heterogeneous Information Network is considered as an undirected weighted simple graph with multiple node types and multiple edge types. And each edge type is presented as "<node_type_A>-<node_type_B>" (e.g. "User-Item"), so there is at most one type of edges for each pair of node types. Besides, isolated nodes are not preferable and the only way to add a node into the graph is to add an edge with this node on one side.

hin2vec_pytorch

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

hin2vec_pytorch

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README