chihchenghsieh / p700

Analysing graph by Siamese-Network

Siamese-Network for analysing graphs

Grid Graph (a kind of Planar graph is used in this case)

Start from 1 to 20 (20 nodes)

Data Generation

Network Architecture

NELU LSTM

NELU has been implemented.

[NALU Paper]

Result from Pytorch

The learning rate decay has been implemented in this model. We can see the model can only classify the data it has seen. And the accuracy on the validation set is not increasing.

After Applying the L2 Regularization, it still can't be solved

Result from tf.Keras

However, this problem can be solved by using a simple algorithm (Graph_Alg) to create a graph.

Only few batch of training data can boost the accuracy on the validation set to 100%

WithColour

Grid graph with Colours

Generating Positive and Negative pairs

Result from Pytorch

LSTN-NALU

Training

& Validation (Epochs)

Original LSTM

Training

& Validation (Epochs)

Still have a serious overfitting problem.

Without Colour but no restriction on the cycle path

Training result

Simple Path to generate the data

With 2 FC layers

Without FC layer

Consider the last node of a path

Training_result

After setting a root

For PredictingNet

3 Colours

10 Colours (SimplestColorGraph)

Idx (20 Colours)

For Siamese Network

3 Colours

10 Colours (SimplestColorGraph)

Idx (20 Colours)

Both siamese network and PredictingNet have the problem of overfitting.