Hello Siamese Network

Siamese Network

• Contains two or more identical subnetworks used to generate feature vectors for each input and compare them
• Loss function: binary crossentropy (work but usually not effective), triplet function, contrastive function, mean square error, etc.
• Application: duplicate detection, anomalies detection, face recognition, etc.
• Original Siamese Network

• A modified Siamese Network

The loss function is triplet.

Notions: A: anchor image. B: positive image (the same class as A). C: negative image (different class from that of A)

Experiment

Training procedure:

• Step 1: Generate training set from MNIST. This training set consists of triples. A triple is defined as (image 1, image 2, its similarity). If image 1 and image 2 are similar (based on their labels), their similarity is assigned to 1. Otherwise, their similarity is assigned to 0.

I also generate a test set in the same way.

In total, the training set has 100k triples. The test set has 50k triples. The training set and test set are balanced.

The source code of this step is provided in dataset_generation.py. Set the paths of TRAINING_SET and TEST_SET correctly. The paths of the training set and the test set are represented by MYTRAINING and MYTEST, respectively.

• Step 2: Train on 70k triples. Evaluate the remaining 30k triples. There are 40 epochs. Of course, the number of epochs could be larger. However, because I am a lazy person, I chose this epoch based on my intuition.

• Step 3: Choose the model achieving the highest accuracy on the validation set.

• Step 4: Compute the accuracy of the model on the whole training set and test set.

The accuracy of the models could be better. I would run more epochs in the future.

Model

I modified the original architecture of Siamese Network a little bit. The output vectors of the mirror parts are concatenated. Originally, these output vectors are computed by using L2 distance.

MNIST

Model	Loss	Training set	Test set	Result
M1 (main1.py)	binary_crossentropy	0.98794	0.9539	link
M2 (main2.py)	mse	0.99195	0.9734	link
M3 (main3.py)	contrastive loss	0.99069	0.9751	link

Fashion-MNIST

Model	Loss	Training set	Test set	Result
F1 (main1.py)	binary_crossentropy	0.93579	0.91416	link
F2 (main2.py)	mse	0.93968	0.91858	link
F3 (main3.py)	contrastive loss	0.94351	0.9107	link

How to use the trained models?

Just fed two images into the trained models. The output is the probability. The higher probability, the more similarity between the two images.

For example, I display 5 comparisons as follows:

Interesting Links

https://www.kaggle.com/martinpiotte/whale-recognition-model-with-score-0-78563#Siamese-Neural-Network-architecture

http://slazebni.cs.illinois.edu/spring17/lec09_similarity.pdf