Siamese Neural Network

An architecture for comparing pairs of inputs

---

source

DistanceSiamese

 DistanceSiamese (backbone:fastai.torch_core.Module,
                  distance_metric=<function
                  normalized_squared_euclidean_distance>)

Outputs the distance between two inputs in feature space

	Type	Default	Details
backbone	Module		embeds inputs in a feature space
distance_metric	function	normalized_squared_euclidean_distance

---

source

ContrastiveLoss

 ContrastiveLoss (margin:float=1.0, size_average=None, reduce=None,
                  reduction:str='mean')

Same as loss_cls, but flattens input and target.

---

source

normalized_squared_euclidean_distance

 normalized_squared_euclidean_distance (x1, x2)

Squared Euclidean distance over normalized vectors: \[\left\| \frac{x_1}{\|x_1\|}-\frac{x_2}{\|x_2\|} \right\|^2 \]

---

source

DistanceSiamese.plot_distance_histogram

 DistanceSiamese.plot_distance_histogram (pairs_dls:Union[fastai.data.core
                                          .TfmdDL,Dict[str,fastai.data.cor
                                          e.TfmdDL]], label='Distance')

Plots a histogram of intra-class and inter-class distances

from fastai_datasets.all import *

classifier = resnet34(weights=ResNet34_Weights.DEFAULT)
siamese = DistanceSiamese(create_body(model=classifier, cut=-1))

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

When starting with a decent backbone, positive pairs are closer than negative pairs:

siamese.plot_distance_histogram(dls.valid)

Train with contrastive loss:

learn = Learner(dls, siamese, ContrastiveLoss(margin=1.5))
learn.fit_one_cycle(3)

epoch	train_loss	valid_loss	time
0	0.431103	0.475045	00:12
1	0.325288	0.382971	00:16
2	0.272254	0.308985	00:17

siamese.plot_distance_histogram(dls.valid)