similarity-learning 0.0.2

A fastai based framework for similarity learning

similarity-learning

Docs

See https://irad-zehavi.github.io/similarity-learning/

Install

pip install similarity_learning

How to use

As an nbdev library, similarity_learning supports import * (without importing unwanted symbols):

from similarity_learning.all import *

Now we can train a pair-matcher. First let’s construct dataloaders of pairs:

from fastai.vision.all import *

from fastai_datasets.all import *

pairs = Pairs(Imagenette(160), .1)
dls = pairs.dls(after_item=Resize(128),
                after_batch=Normalize.from_stats(*imagenet_stats))

<div>
  <progress value='196' class='' max='196' style='width:300px; height:20px; vertical-align: middle;'></progress>
  100.00% [196/196 00:00&lt;00:00]
</div>

To get quick results, we can use the body of a pretrained model as a backbone for our Siamese neural network:

classifier = resnet34(weights=ResNet34_Weights.DEFAULT)
siamese = ThresholdSiamese(create_body(model=classifier, cut=-1)).to(dls.device)
siamese.fit_threshold(dls.train)

<div>
  <progress value='14' class='' max='14' style='width:300px; height:20px; vertical-align: middle;'></progress>
  100.00% [14/14 00:05&lt;00:00]
</div>


(1.0299999713897705, 0.8895089626312256)

Let’s see how good it is:

learn = Learner(dls, siamese, metrics=accuracy)
learn.validate()

(#2) [0.541471004486084,0.9005101919174194]

learn.show_results()

Not bad, but we can do better with finetuning:

learn.fit(5, 1e-4)
learn.validate()

(#2) [0.2768465280532837,0.9464285969734192]

learn.show_results()

We can also consider the distribution of feature-space distances compared to the decision threshold:

siamese.plot_distance_histogram(dls.valid)

<div>
  <progress value='7' class='' max='7' style='width:300px; height:20px; vertical-align: middle;'></progress>
  100.00% [7/7 00:02&lt;00:00]
</div>

See the rest of the docs for more examples, including more visualizations, comparison of loss functions, and facial recognition.