PyTorch implementation of Scale-Aware Triplet Networks

This repository contains code for Scale-Aware Triplet Networks based on Learning Deep Descriptors with Scale-Aware Triplet Networks implemented in PyTorch.

Requirements

• Python 3.6
• PyTorch 1.0

Usage

Args:

• theta_glo (float, default: 1.15): Global context in all triplets.
• delta (int, default: 5): Scale correction parameter.
• gamma (float, default: 0.5): Ratio of siamese and triplet.
• scale_aware (bool, default: True): Scale-aware sampling.

Input:

• y_a: Anchor samples.
• y_b: Positive samples. Each positive samples have same class labels to the correspond anchor samples.
• targets: Class labels of y_a and y_b.

criterion = MixedContextLoss(theta_glo=1.15, delta=5, gamma=0.5, scale_aware=True)
y_a = model(input1)
y_p = model(input2)
loss = criterion(y_a, y_p, targets)
optimizer.zero_grad()
loss.backward()
optimizer.step()

Training

MNIST

Use scale-aware siamese loss:

python train.py --gamma 0 --scale-aware True

Use scale-aware triplet loss:

python train.py --gamma 1 --scale-aware True

Use scale-aware mixed context loss (gamma=0.5):

python train.py --gamma 0.5 --scale-aware True

Results

MNIST