From ab29067d6469473481cc73fe42bcaf69d7633a83 Mon Sep 17 00:00:00 2001
From: Jordan Gong <jordan.gong@protonmail.com>
Date: Tue, 5 Jan 2021 20:20:06 +0800
Subject: Implement Batch All Triplet Loss

---
 utils/triplet_loss.py | 35 +++++++++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)
 create mode 100644 utils/triplet_loss.py

(limited to 'utils')

diff --git a/utils/triplet_loss.py b/utils/triplet_loss.py
new file mode 100644
index 0000000..242be45
--- /dev/null
+++ b/utils/triplet_loss.py
@@ -0,0 +1,35 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class BatchAllTripletLoss(nn.Module):
+    def __init__(self, margin: float = 0.2):
+        super().__init__()
+        self.margin = margin
+
+    def forward(self, x, y):
+        # Duplicate labels for each part
+        p, n, c = x.size()
+        y = y.repeat(p, 1)
+
+        # Euclidean distance p x n x n
+        x_squared_sum = torch.sum(x ** 2, dim=2)
+        x1_squared_sum = x_squared_sum.unsqueeze(1)
+        x2_squared_sum = x_squared_sum.unsqueeze(2)
+        x1_times_x2_sum = x @ x.transpose(1, 2)
+        dist = torch.sqrt(x1_squared_sum - 2 * x1_times_x2_sum + x2_squared_sum)
+
+        hard_positive_mask = y.unsqueeze(1) == y.unsqueeze(2)
+        hard_negative_mask = y.unsqueeze(1) != y.unsqueeze(2)
+        all_hard_positive = dist[hard_positive_mask].view(p, n, -1, 1)
+        all_hard_negative = dist[hard_negative_mask].view(p, n, 1, -1)
+        positive_negative_dist = all_hard_positive - all_hard_negative
+        all_loss = F.relu(self.margin + positive_negative_dist).view(p, -1)
+
+        # Non-zero parted mean
+        parted_loss_mean = all_loss.sum(1) / (all_loss != 0).sum(1)
+        parted_loss_mean[parted_loss_mean == float('Inf')] = 0
+
+        loss = parted_loss_mean.mean()
+        return loss
-- 
cgit v1.2.3