search

binxuan / Recurrent-Graph-Neural-Network

Git repository for our submitted paper
0 stars 0 forks source link

Something about recurrent neural network model #1

Open CFF-Dream opened 4 years ago

CFF-Dream commented 4 years ago

Hello, I've read your paper"Inductive Graph Representation learnig with Recurrnt Graph Neural Networks" which benefit me a lot, but there r also some problems confused me for months. The input of the rnn unit should be a tensor which shape like [batch, sequence_len, num_features], it means that the input should be a sequence, but every graph in PPI dataset is complex connects by many nodes. if input the data.x directly means automatically form a sequence according to the order of the node_index?How did u sovle this problem?Or can u share your codes on github?Thank u!

binxuan commented 4 years ago

Hi, thanks for the interest. In practice, you can feed a tensor with shape [batch, sequence_len, num_features] into a RNN at once. But alternatively, you can also update RNN's state iteratively with a for loop like this for i in range(sequence_len): state = RNN([batch, i-th num_features], state)

CFF-Dream commented 4 years ago

Thanks for your reply! Do you mean"sequence_len"=="num_nodes"? And I have wrote the code according to your model, but performance is not good, maybe somgthing goes wrong

binxuan commented 4 years ago

The sequence_len is the model depth (number of layers).

Binxuan

On Sep 10, 2019, at 10:13 AM, CFF-Dream notifications@github.com wrote:

Thanks for your reply! Do you mean"sequence_len"=="num_nodes"? And I have wrote the code according to your model, but performance is not good, maybe somgthing goes wrong