Merge pull request #8 from PoCInnovation/feat/prototype-v1

[ADD]: Symetric BCE + binary class

Author: SachaHenneveux

prototype_v1/classifier.py

@@ -1,10 +1,11 @@
 """
-Classifier avec Cross-Entropy
+Classifier Binaire avec Symmetric BCE Loss
 
-Transforme les embeddings contextualisés en prédictions par nœud.
+Transforme les embeddings contextualisés en prédictions binaires par nœud.
+La Symmetric BCE Loss gère automatiquement la symétrie des solutions (MaxCut).
 
 Input:  [batch, n_nodes, hidden_dim]
-Output: [batch, n_nodes, num_classes] (logits)
+Output: [batch, n_nodes] (probabilités entre 0 et 1)
 """
 
 import torch
@@ -14,41 +15,38 @@
 
 class Classifier(nn.Module):
     """
-    Classifier multi-classe pour les problèmes d'optimisation.
+    Classifier binaire pour les problèmes d'optimisation sur graphes.
 
-    Supporte:
-    - MaxCut, Vertex Cover, Independent Set: 2 classes
-    - Graph Coloring: k classes
+    Supporte: MaxCut, Vertex Cover, Independent Set (tous binaires).
+
+    Loss: Symmetric BCE
+        loss = min(BCE(pred, target), BCE(pred, 1-target))
+        → Gère automatiquement la symétrie des solutions
     """
 
-    def __init__(self, hidden_dim=256, max_classes=10, dropout=0.1):
+    def __init__(self, hidden_dim=256, dropout=0.1):
         super().__init__()
 
-        self.hidden_dim = hidden_dim
-        self.max_classes = max_classes
-
         self.layers = nn.Sequential(
             nn.LayerNorm(hidden_dim),
             nn.Linear(hidden_dim, hidden_dim // 2),
             nn.GELU(),
             nn.Dropout(dropout),
-            nn.Linear(hidden_dim // 2, max_classes)
+            nn.Linear(hidden_dim // 2, 1)  # 1 seule sortie → binaire
         )
 
-    def forward(self, x, num_classes=2):
+    def forward(self, x):
         """
         Args:
             x: [batch, n_nodes, hidden_dim] - embeddings contextualisés
-            num_classes: int - nombre de classes
 
         Returns:
-            logits: [batch, n_nodes, num_classes]
-            probs: [batch, n_nodes, num_classes]
-            predictions: [batch, n_nodes]
+            probs: [batch, n_nodes] - probabilités entre 0 et 1
+            predictions: [batch, n_nodes] - 0 ou 1
         """
-        logits = self.layers(x)[:, :, :num_classes]
-        probs = F.softmax(logits, dim=-1)
-        predictions = torch.argmax(logits, dim=-1)
+        logits = self.layers(x).squeeze(-1)  # [batch, n_nodes]
+        probs = torch.sigmoid(logits)         # Sigmoid → [0, 1]
+        predictions = (probs > 0.5).long()    # Seuil → 0 ou 1
 
         return {
             'logits': logits,
@@ -58,47 +56,105 @@ def forward(self, x, num_classes=2):
 
     def compute_loss(self, logits, targets, mask=None):
         """
-        Cross-Entropy Loss.
+        Symmetric BCE Loss.
+
+        Calcule la BCE dans les deux sens (target et 1-target)
+        et garde le minimum → gère la symétrie.
 
         Args:
-            logits: [batch, n_nodes, num_classes]
-            targets: [batch, n_nodes] - classes {0, 1, ..., k-1}
-            mask: [batch, n_nodes] - optionnel
+            logits: [batch, n_nodes] - sorties brutes (avant sigmoid)
+            targets: [batch, n_nodes] - valeurs 0 ou 1
+            mask: [batch, n_nodes] - optionnel (pour graphes de tailles différentes)
 
         Returns:
             loss: scalar
         """
-        b, n, c = logits.shape
-        logits_flat = logits.reshape(-1, c)
-        targets_flat = targets.reshape(-1).long()
+        targets = targets.float()
 
         if mask is not None:
-            mask_flat = mask.reshape(-1).float()
-            loss = F.cross_entropy(logits_flat, targets_flat, reduction='none')
-            return (loss * mask_flat).sum() / mask_flat.sum().clamp(min=1)
+            mask = mask.float()
+
+            # Loss directe : pred vs target
+            loss_direct = F.binary_cross_entropy_with_logits(
+                logits, targets, reduction='none'
+            )
+            loss_direct = (loss_direct * mask).sum() / mask.sum().clamp(min=1)
+
+            # Loss inversée : pred vs (1 - target)
+            loss_inverse = F.binary_cross_entropy_with_logits(
+                logits, 1.0 - targets, reduction='none'
+            )
+            loss_inverse = (loss_inverse * mask).sum() / mask.sum().clamp(min=1)
+        else:
+            # Loss directe : pred vs target
+            loss_direct = F.binary_cross_entropy_with_logits(logits, targets)
 
-        return F.cross_entropy(logits_flat, targets_flat)
+            # Loss inversée : pred vs (1 - target)
+            loss_inverse = F.binary_cross_entropy_with_logits(logits, 1.0 - targets)
+
+        # Symmetric : on prend le minimum des deux
+        loss = torch.min(loss_direct, loss_inverse)
+
+        return loss
+
+    def compute_similarity(self, predictions, targets):
+        """
+        Calcule le pourcentage de ressemblance (en tenant compte de la symétrie).
+
+        Args:
+            predictions: [batch, n_nodes] - 0 ou 1
+            targets: [batch, n_nodes] - 0 ou 1
+
+        Returns:
+            similarity: float entre 0 et 1 (1 = parfait)
+        """
+        predictions = predictions.float()
+        targets = targets.float()
+
+        # Ressemblance directe
+        match_direct = (predictions == targets).float().mean()
+
+        # Ressemblance inversée
+        match_inverse = (predictions == (1.0 - targets)).float().mean()
+
+        # Meilleure des deux
+        similarity = torch.max(match_direct, match_inverse)
+
+        return similarity.item()
 
 
 if __name__ == "__main__":
-    print("=== Test Classifier ===")
+    print("=== Test Classifier (Binaire + Symmetric BCE) ===\n")
 
     x = torch.randn(4, 6, 256)  # [batch, n_nodes, hidden_dim]
-    classifier = Classifier(hidden_dim=256, max_classes=10)
+    classifier = Classifier(hidden_dim=256)
 
-    # Test 2 classes (MaxCut)
-    output = classifier(x, num_classes=2)
-    print(f"Logits (2 classes): {output['logits'].shape}")
+    # Forward
+    output = classifier(x)
+    print(f"Logits: {output['logits'].shape}")
+    print(f"Probs: {output['probs'].shape}")
+    print(f"Predictions: {output['predictions'].shape}")
+    print(f"Exemple probs: {output['probs'][0].tolist()}")
+    print(f"Exemple preds: {output['predictions'][0].tolist()}")
 
-    # Test 5 classes (Graph Coloring)
-    output = classifier(x, num_classes=5)
-    print(f"Logits (5 classes): {output['logits'].shape}")
+    # Test Symmetric Loss
+    targets = torch.tensor([[1, 0, 1, 0, 1, 0]] * 4).float()
 
-    # Test loss
-    targets = torch.randint(0, 2, (4, 6))
-    output = classifier(x, num_classes=2)
     loss = classifier.compute_loss(output['logits'], targets)
-    print(f"Loss: {loss.item():.4f}")
-
-    print(f"Params: {sum(p.numel() for p in classifier.parameters()):,}")
+    print(f"\nSymmetric BCE Loss: {loss.item():.4f}")
+
+    # Test symétrie : target et 1-target doivent donner la même loss
+    loss_normal = classifier.compute_loss(output['logits'], targets)
+    loss_inverted = classifier.compute_loss(output['logits'], 1 - targets)
+    print(f"Loss (target normal):  {loss_normal.item():.4f}")
+    print(f"Loss (target inversé): {loss_inverted.item():.4f}")
+    print(f"Égales ? {'✅ OUI' if abs(loss_normal.item() - loss_inverted.item()) < 1e-6 else '❌ NON'}")
+
+    # Test similarité
+    pred = torch.tensor([[0, 1, 0, 1, 0, 1]])
+    target = torch.tensor([[1, 0, 1, 0, 1, 0]])
+    sim = classifier.compute_similarity(pred, target)
+    print(f"\nSimilarité [0,1,0,1,0,1] vs [1,0,1,0,1,0]: {sim:.0%}")
+
+    print(f"\nParams: {sum(p.numel() for p in classifier.parameters()):,}")
     print("✅ OK")

prototype_v1/explication_bitstring_comparaison.txt

@@ -0,0 +1,54 @@
+Format QAOA (Qiskit)
+
+result = solver.solve(problem)
+print(result.x)  # numpy array: [0, 1, 0, 1]
+print(type(result.x))  # <class 'numpy.ndarray'>
+Format Notre Modèle
+
+output = model(x, edge_index, problem_id=0)
+print(output['predictions'])  # tensor([[0, 1, 0, 1]])
+print(type(output['predictions']))  # <class 'torch.Tensor'>
+
+
+Sont-ils comparables ?
+         QAOA	      Notre Modèle
+Type	numpy.ndarray	torch.Tensor
+Shape	[n_nodes]	  [batch, n_nodes]
+Valeurs	{0, 1}	        {0, 1}
+
+Signification	Nœud i dans set 0 ou 1	Nœud i dans set 0 ou 1
+OUI mais il faut convertir :
+
+
+# QAOA → Tensor pour comparaison
+qaoa_target = torch.tensor(result.x)  # [0, 1, 0, 1]
+
+# Notre modèle → squeeze pour enlever batch dim
+model_pred = output['predictions'].squeeze(0)  # [0, 1, 0, 1]
+
+# Maintenant comparable !
+⚠️ ATTENTION : Symétrie du problème !
+Pour MaxCut, il y a une subtilité :
+
+
+Solution [0, 1, 0, 1] = Set A: {0, 2}, Set B: {1, 3}
+Solution [1, 0, 1, 0] = Set A: {1, 3}, Set B: {0, 2}
+
+CE SONT LES MÊMES SOLUTIONS ! (juste inversées)
+Donc si :
+
+
+QAOA dit:    [0, 1, 0, 1]
+Modèle dit:  [1, 0, 1, 0]
+
+→ C'est CORRECT ! Même partition, juste labels inversés.
+Solutions pour gérer la symétrie
+Option 1 : Normaliser (forcer à commencer par 0)
+
+def normalize_bitstring(bits):
+    if bits[0] == 1:
+        return 1 - bits  # Inverser
+    return bits
+
+target = normalize_bitstring(qaoa_result)
+pred = normalize_bitstring(model_pred)

prototype_v1/model.py

@@ -5,7 +5,7 @@
     Graph → GNN Encoder → E_local, E_global
     problem_id → Lookup Table → E_prob
     Concat [E_global || E_local || E_prob] → Transformer → embeddings contextualisés
-    Classifier → logits → Cross-Entropy Loss → Backpropagation
+    Classifier binaire → probs → Symmetric BCE Loss → Backpropagation
 """
 
 import torch
@@ -20,11 +20,12 @@ class QuantumGraphModel(nn.Module):
     """
     Modèle complet pour résoudre des problèmes d'optimisation sur graphes.
 
-    Supporte:
+    Supporte (tous binaires):
     - MaxCut (2 classes)
     - Vertex Cover (2 classes)
     - Independent Set (2 classes)
-    - Graph Coloring (k classes)
+
+    Loss: Symmetric BCE (gère la symétrie des solutions)
     """
 
     def __init__(
@@ -36,7 +37,6 @@ def __init__(
         transformer_layers=4,
         num_heads=8,
         num_problems=10,
-        max_classes=10,
         dropout=0.1
     ):
         super().__init__()
@@ -67,24 +67,22 @@ def __init__(
             dropout=dropout
         )
 
-        # 4. Classifier
+        # 4. Classifier (binaire)
         self.classifier = Classifier(
             hidden_dim=hidden_dim,
-            max_classes=max_classes,
             dropout=dropout
         )
 
-    def forward(self, x, edge_index, problem_id, batch=None, num_classes=2):
+    def forward(self, x, edge_index, problem_id, batch=None):
         """
         Args:
             x: [n_nodes, node_input_dim]
             edge_index: [2, n_edges]
             problem_id: int ou [batch_size]
             batch: [n_nodes] (optionnel)
-            num_classes: int
 
         Returns:
-            dict avec logits, probs, predictions
+            dict avec logits, probs, predictions (tous [batch, n_nodes])
         """
         # 1. GNN Encoder → E_local, E_global
         e_local, e_global = self.encoder(x, edge_index, batch)
@@ -105,8 +103,8 @@ def forward(self, x, edge_index, problem_id, batch=None, num_classes=2):
         # 3. Transformer → embeddings contextualisés
         contextualized = self.transformer(e_local, e_global, e_prob)
 
-        # 4. Classifier → logits, probs, predictions
-        output = self.classifier(contextualized, num_classes=num_classes)
+        # 4. Classifier binaire → probs, predictions
+        output = self.classifier(contextualized)
 
         return output
 
@@ -126,14 +124,19 @@ def _batch_node_embeddings(self, e_local, batch, batch_size):
         return out
 
     def compute_loss(self, logits, targets, mask=None):
-        """Cross-Entropy Loss"""
+        """Symmetric BCE Loss"""
         return self.classifier.compute_loss(logits, targets, mask)
 
-    def forward_with_loss(self, x, edge_index, problem_id, targets, batch=None, num_classes=2):
+    def compute_similarity(self, predictions, targets):
+        """Pourcentage de ressemblance (avec symétrie)"""
+        return self.classifier.compute_similarity(predictions, targets)
+
+    def forward_with_loss(self, x, edge_index, problem_id, targets, batch=None):
         """Forward + Loss en une seule passe"""
-        output = self.forward(x, edge_index, problem_id, batch, num_classes)
+        output = self.forward(x, edge_index, problem_id, batch)
         loss = self.compute_loss(output['logits'], targets)
-        return output, loss
+        similarity = self.compute_similarity(output['predictions'], targets)
+        return output, loss, similarity
 
 
 if __name__ == "__main__":
@@ -159,24 +162,27 @@ def forward_with_loss(self, x, edge_index, problem_id, targets, batch=None, num_
     print(f"Paramètres: {sum(p.numel() for p in model.parameters()):,}")
 
     # Forward (MaxCut)
-    output = model(x, edge_index, problem_id=0, num_classes=2)
-    print(f"\nMaxCut (2 classes):")
-    print(f"  Logits: {output['logits'].shape}")
+    output = model(x, edge_index, problem_id=0)
+    print(f"\nMaxCut:")
+    print(f"  Probs: {output['probs'].shape}")
     print(f"  Predictions: {output['predictions']}")
 
-    # Loss
+    # Symmetric Loss
     targets = torch.tensor([[1, 0, 1, 0, 1, 0]])
     loss = model.compute_loss(output['logits'], targets)
     print(f"  Loss: {loss.item():.4f}")
 
+    # Test symétrie
+    loss_inv = model.compute_loss(output['logits'], 1 - targets)
+    print(f"  Loss inversée: {loss_inv.item():.4f}")
+    print(f"  Symétrie OK ? {'✅' if abs(loss.item() - loss_inv.item()) < 1e-6 else '❌'}")
+
+    # Similarité
+    sim = model.compute_similarity(output['predictions'], targets)
+    print(f"  Similarité: {sim:.0%}")
+
     # Backprop
     loss.backward()
     print("  Backprop OK")
 
-    # Graph Coloring
-    output = model(x, edge_index, problem_id=3, num_classes=5)
-    print(f"\nGraph Coloring (5 classes):")
-    print(f"  Logits: {output['logits'].shape}")
-    print(f"  Predictions: {output['predictions']}")
-
     print("\n✅ Tous les tests passés!")