delijati / pytorch-siamese

Siamese Network implementation using Pytorch

pytorch-siamese

This is a port of chainer-siamese

Install

This installation requires cuda to be installed.

$ virtualenv /usr/bin/python3.5 env
$ env/bin/pip install http://download.pytorch.org/whl/cu80/torch-0.1.12.post2-cp35-cp35m-linux_x86_64.whl 
$ env/bin/pip install torchvision

Run

$ env/bin/python train_mnist.py --epoch 10

This dumps for every epoch ther current state and creates a result.png.

Run specific model

$ env/bin/python train_mnist.py -m model-epoch-7.pth

Result

pytorch-siamese's Issues

Loss: nan

I tried to run this train_mnist.py, but came across Loss: nan.

Also, it is not helpful to reduce the learning rate.

Do you have a way to make the network converge?

Many thanks!

A question about the implementation in pytorch

Hi, the way you implement siamese network is like that:

output1,output2 = model(x0, x1)
loss = criterion (output1,output2)

And here is the another way:

output1 = model.forward_once(x0)
output2 = model.forward_once(x1)
loss = criterion (output1,output2)

Are they different?

Trying to build a Siamese Network with your code as reference

Hello @delijati
I am trying to build a siamese network very similar to what you have done, I am looking to feed the outputs from the siamese network arms to a fully connected layer which has a loss function of its own.
Do you know how to do it in pytorch? Any kind of help is much appreciated!

Thanks

loss: nan

Hey,

Good work. However maybe you're aware, with default hyperparameter settings, loss: nan.