fangpin / siamese-pytorch

Implementation of Siamese Networks for image one-shot learning by PyTorch, train and test model on dataset Omniglot
285 stars 62 forks source link
one-shot one-shot-learning pytorch siamese siamese-network

Siamese Networks for One-Shot Learning

A reimplementation of the original paper in pytorch with training and testing on the Omniglot dataset.

requirement

See requirements.txt

run step

experiment result

Loss value is sampled after every 200 batches img My final precision is 89.5% a little smaller than the result of the paper (92%).

The small result difference might be caused by some difference between my implementation and the paper's. I list these differences as follows:

instead of using SGD with momentum I just use ADAM.

Instead of using individual initialization methods, learning rates and regularization rates at different layers I simply use the default setting of pytorch and keep them same.