fix torch cluster

app.py

@@ -3,12 +3,40 @@ import torch
 import numpy as np
 import trimesh
 import plotly.graph_objects as go
-from torch_geometric.data import Data
-from torch_cluster import fps, radius_graph, knn
 from torch_geometric.utils import to_dense_batch
-
 from model import HierarchicalFPSCliffordNet
 
+# --- SAF PYTORCH C++ YEDEKLERİ (Hugging Face için Özel Çözüm) ---
+def knn_pure(x, y, k):
+    dist = torch.cdist(x, y)
+    _, topk_idx = torch.topk(dist, k, dim=1, largest=False)
+    col = torch.arange(x.size(0)).view(-1, 1).expand(-1, k).reshape(-1)
+    row = topk_idx.reshape(-1)
+    return torch.stack([row, col], dim=0)
+
+def radius_graph_pure(pos, r, max_num_neighbors=32):
+    dist = torch.cdist(pos, pos)
+    n = pos.size(0)
+    k = min(max_num_neighbors, n)
+    _, target = torch.topk(dist, k, dim=1, largest=False)
+    source = torch.arange(n).view(-1, 1).expand(-1, k)
+    valid = dist[source, target] < r
+    return torch.stack([source[valid], target[valid]], dim=0)
+
+def fps_pure(pos, ratio=0.5):
+    n = pos.size(0)
+    k = max(1, int(n * ratio))
+    idx = torch.zeros(k, dtype=torch.long, device=pos.device)
+    dist = torch.full((n,), 1e10, device=pos.device)
+    farthest = 0
+    for i in range(k):
+        idx[i] = farthest
+        farthest_point = pos[farthest].view(1, 3)
+        dist_to_farthest = torch.cdist(pos, farthest_point).squeeze(-1)
+        dist = torch.min(dist, dist_to_farthest)
+        farthest = torch.argmax(dist).item()
+    return idx
+
 # --- Kategori Sözlükleri ---
 CATEGORIES = {
     'Airplane': 0, 'Bag': 1, 'Cap': 2, 'Car': 3, 'Chair': 4, 'Earphone': 5, 'Guitar': 6,
@@ -24,7 +52,7 @@ SEG_CLASSES = {
 }
 
 # --- Modeli Yükleme ---
-device = torch.device("cpu") # HF Free Tier her zaman CPU'dur
+device = torch.device("cpu")
 model = HierarchicalFPSCliffordNet(base_channels=12).to(device)
 
 try:
@@ -38,11 +66,9 @@ except Exception as e:
 
 # --- Veri Hazırlama Fonksiyonu ---
 def process_point_cloud(file_path, category_name):
-    # Dosyayı oku ve noktalara çevir
     mesh = trimesh.load(file_path, force='mesh')
     points = np.array(mesh.vertices)
     
-    # 1024 noktaya örnekle (Eğitimdeki gibi)
     if len(points) > 1024:
         idx = np.random.choice(len(points), 1024, replace=False)
         points = points[idx]
@@ -51,8 +77,6 @@ def process_point_cloud(file_path, category_name):
         points = points[idx]
         
     pos = torch.tensor(points, dtype=torch.float32)
-    
-    # Birim Küreye Normalize Et
     pos = pos - pos.mean(dim=0, keepdim=True)
     scale = pos.norm(dim=1).max().clamp(min=1e-8)
     pos = pos / scale
@@ -60,68 +84,52 @@ def process_point_cloud(file_path, category_name):
     batch = torch.zeros(pos.size(0), dtype=torch.long)
     cat_idx = torch.tensor([CATEGORIES[category_name]], dtype=torch.long)
     
-    # Graf İndekslerini Hesapla (Offline Preprocessing simülasyonu)
-    fps_idx_1 = fps(pos, batch, ratio=0.5).long()
+    # Saf PyTorch ile Hızlı Hesaplamalar (C++ Yok)
+    fps_idx_1 = fps_pure(pos, ratio=0.5)
     pos2 = pos[fps_idx_1]
     batch2 = torch.zeros(pos2.size(0), dtype=torch.long)
     
-    fps_idx_2 = fps(pos2, batch2, ratio=0.25).long()
+    fps_idx_2 = fps_pure(pos2, ratio=0.25)
     pos3 = pos2[fps_idx_2]
     batch3 = torch.zeros(pos3.size(0), dtype=torch.long)
 
-    edge_index_1 = radius_graph(pos, r=0.15, batch=batch, max_num_neighbors=16, loop=True)
-    edge_index_2 = radius_graph(pos2, r=0.30, batch=batch2, max_num_neighbors=32, loop=True)
+    edge_index_1 = radius_graph_pure(pos, r=0.15, max_num_neighbors=16)
+    edge_index_2 = radius_graph_pure(pos2, r=0.30, max_num_neighbors=32)
 
-    assign_index_32 = knn(pos3, pos2, k=3, batch_x=batch3, batch_y=batch2)
-    assign_index_21 = knn(pos2, pos, k=3, batch_x=batch2, batch_y=batch)
+    assign_index_32 = knn_pure(pos3, pos2, k=3)
+    assign_index_21 = knn_pure(pos2, pos, k=3)
 
     dummy_x = torch.zeros(pos3.size(0), 1)
     _, x_dense_mask = to_dense_batch(dummy_x, batch3, max_num_nodes=int(torch.bincount(batch3).max().item()))
     
     return pos, batch, cat_idx, fps_idx_1, fps_idx_2, pos2, batch2, pos3, batch3, edge_index_1, edge_index_2, assign_index_32, assign_index_21, x_dense_mask
 
-# --- Çıkarım (Inference) Fonksiyonu ---
+# --- Çıkarım Fonksiyonu ---
 def predict(file, category_name):
     if file is None:
         return None
         
-    # Veriyi modele uygun formata getir
     (pos, batch, cat_idx, fps_idx_1, fps_idx_2, pos2, batch2, pos3, batch3, 
      edge_index_1, edge_index_2, assign_index_32, assign_index_21, x_dense_mask) = process_point_cloud(file.name, category_name)
     
-    # Modelden Geçir
     with torch.no_grad():
         logits = model(pos, batch, cat_idx, fps_idx_1, fps_idx_2, pos2, batch2, pos3, batch3,
                        edge_index_1, edge_index_2, assign_index_32, assign_index_21, x_dense_mask)
         
-        # O kategoriye ait olmayan parçaları maskele
         valid_parts = SEG_CLASSES[category_name]
         mask = torch.zeros_like(logits, dtype=torch.bool)
         mask[:, valid_parts] = True
         logits = logits.masked_fill(~mask, -10000.0)
-        
-        # Sınıfı Tahmin Et
         preds = logits.argmax(dim=-1).numpy()
     
-    # Plotly ile Görselleştir
     pos_np = pos.numpy()
     fig = go.Figure(data=[go.Scatter3d(
         x=pos_np[:, 0], y=pos_np[:, 1], z=pos_np[:, 2],
         mode='markers',
-        marker=dict(
-            size=4,
-            color=preds,                # Her parçaya farklı renk atar
-            colorscale='Viridis',
-            opacity=0.9,
-            line=dict(width=0.5, color='black') # Noktalara derinlik katar
-        )
+        marker=dict(size=4, color=preds, colorscale='Viridis', opacity=0.9, line=dict(width=0.5, color='black'))
     )])
     
-    fig.update_layout(
-        margin=dict(l=0, r=0, b=0, t=0),
-        scene=dict(xaxis=dict(visible=False), yaxis=dict(visible=False), zaxis=dict(visible=False))
-    )
-    
+    fig.update_layout(margin=dict(l=0, r=0, b=0, t=0), scene=dict(xaxis=dict(visible=False), yaxis=dict(visible=False), zaxis=dict(visible=False)))
     return fig
 
 # --- Gradio Arayüzü ---

model.py

@@ -3,10 +3,8 @@ import torch.nn as nn
 import torch.nn.functional as F
 from torch_geometric.nn import MessagePassing
 from torch_geometric.utils import to_dense_batch
-from torch_cluster import fps, radius_graph, knn
 import math
 
-# --- TRITON & CPU FALLBACK KONTROLÜ ---
 try:
     import triton
     import triton.language as tl
@@ -15,7 +13,6 @@ except ImportError:
     HAS_TRITON = False
 
 def pytorch_clifford_product(a, b):
-    """Hugging Face Free Tier (CPU) için saf PyTorch 8D Geometrik Çarpım"""
     res = torch.zeros_like(a)
     res[..., 0] = a[...,0]*b[...,0] + a[...,1]*b[...,1] + a[...,2]*b[...,2] + a[...,3]*b[...,3] - a[...,4]*b[...,4] - a[...,5]*b[...,5] - a[...,6]*b[...,6] - a[...,7]*b[...,7]
     res[..., 1] = a[...,0]*b[...,1] + a[...,1]*b[...,0] - a[...,2]*b[...,4] + a[...,3]*b[...,6] + a[...,4]*b[...,2] - a[...,5]*b[...,7] - a[...,6]*b[...,3] - a[...,7]*b[...,5]
@@ -28,13 +25,10 @@ def pytorch_clifford_product(a, b):
     return res
 
 def smart_clifford_product(a, b):
-    # Eğer ortamda GPU ve Triton yoksa, CPU versiyonunu kullan.
     if HAS_TRITON:
-        # Eğitilmiş Triton kodlarınız buraya gelir (Çıkarım için genellikle CPU yeterlidir)
         pass 
     return pytorch_clifford_product(a, b)
 
-
 class CliffordDiracLayer(MessagePassing):
     def __init__(self, channels: int):
         super().__init__(aggr="add", node_dim=0)

requirements.txt

@@ -1,8 +1,5 @@
 torch
 torch-geometric
-torch-scatter
-torch-sparse
-torch-cluster
 gradio
 plotly
 numpy