From 948e0d2ab8aec738b4188c7978c42e2d3d880bab Mon Sep 17 00:00:00 2001
From: zorelcocos-dot <zorelcocos@gmail.com>
Date: Sun, 19 Jul 2026 22:08:08 +0300
Subject: [PATCH] feat: Siamese network baseline for pose similarity

---
 src/poseguide/models/siamese.py | 45 +++++++++++++++++++++++++++++++++
 1 file changed, 45 insertions(+)
 create mode 100644 src/poseguide/models/siamese.py

diff --git a/src/poseguide/models/siamese.py b/src/poseguide/models/siamese.py
new file mode 100644
index 0000000..8716555
--- /dev/null
+++ b/src/poseguide/models/siamese.py
@@ -0,0 +1,45 @@
+import torch
+import torch.nn as nn
+
+class PoseEncoder(nn.Module):
+    def __init__(self, input_dim=66, hidden_dim=128, embed_dim=64):
+        # input_dim=66 for 33 2D mediapipe landmarks
+        super(PoseEncoder, self).__init__()
+        self.net = nn.Sequential(
+            nn.Linear(input_dim, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, embed_dim)
+        )
+        
+    def forward(self, x):
+        return self.net(x)
+
+class SiamesePoseSimilarity(nn.Module):
+    def __init__(self, encoder):
+        super(SiamesePoseSimilarity, self).__init__()
+        self.encoder = encoder
+        
+    def forward(self, pose1, pose2):
+        embed1 = self.encoder(pose1)
+        embed2 = self.encoder(pose2)
+        # Cosine similarity for distance
+        return nn.functional.cosine_similarity(embed1, embed2, dim=-1)
+
+def contrastive_loss(similarity, label, margin=0.5):
+    """
+    label: 1 if similar, -1 if dissimilar
+    """
+    loss_similar = torch.mean((1 - similarity)[label == 1]**2)
+    loss_dissimilar = torch.mean(torch.clamp(similarity - margin, min=0)[label == -1]**2)
+    return loss_similar + loss_dissimilar
+
+if __name__ == "__main__":
+    encoder = PoseEncoder()
+    model = SiamesePoseSimilarity(encoder)
+    
+    # Dummy batched inputs
+    pose_a = torch.randn(4, 66)
+    pose_b = torch.randn(4, 66)
+    sim = model(pose_a, pose_b)
+    print(f"Similarity output shape: {sim.shape}")
+    print("Siamese baseline initialized.")
\ No newline at end of file