dinov3 example

dinov3_example.py

@@ -0,0 +1,165 @@
+import torch
+from transformers import AutoModel, AutoImageProcessor
+from PIL import Image
+import numpy as np
+
+
+def dinov3_basic():
+    print("=== DINOv3 Basic Example ===")
+
+    # model loading
+    model_name = "facebook/dinov3-vits16-pretrain-lvd1689m"
+    print(f"Loading model: {model_name}")
+
+    try:
+        model = AutoModel.from_pretrained(model_name)
+        processor = AutoImageProcessor.from_pretrained(model_name)
+        print(f"✓ Model loaded successfully")
+        print(f"  Model type: {model.config.model_type}")
+        print(f"  Hidden size: {model.config.hidden_size}")
+        print(f"  Image size: {model.config.image_size}")
+    except Exception as e:
+        print(f"✗ Model loading failed: {e}")
+        return False
+
+    print("Testing inference with dummy image...")
+    try:
+        dummy_image = Image.new('RGB', (224, 224), color=(128, 64, 192))
+
+        # Process the image
+        inputs = processor(images=dummy_image, return_tensors="pt")
+        print(f"  Input tensor shape: {inputs.pixel_values.shape}")
+
+        # Run inference
+        model.eval()
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # Check outputs
+        print(f"✓ Inference successful")
+        if hasattr(outputs, 'pooler_output') and outputs.pooler_output is not None:
+            print(f"  Pooler output shape: {outputs.pooler_output.shape}")
+            print(f"  Pooler output (CLS token): {outputs.pooler_output.shape}")
+        if hasattr(outputs, 'last_hidden_state'):
+            print(f"  Last hidden state shape: {outputs.last_hidden_state.shape}")
+            print(f"  CLS token shape: {outputs.last_hidden_state[:, 0].shape}")
+            print(f"  Patch tokens shape: {outputs.last_hidden_state[:, 1:].shape}")
+
+        return True
+
+    except Exception as e:
+        print(f"✗ Inference failed: {e}")
+        return False
+
+
+def dinov3_variants():
+    """different DINOv3 model variants availability."""
+    print("=== Testing DINOv3 Model Variants ===")
+
+    variants = [
+        ("facebook/dinov3-vits16-pretrain-lvd1689m", "ViT-S/16", 384),
+        ("facebook/dinov3-vitb16-pretrain-lvd1689m", "ViT-B/16", 768),
+        ("facebook/dinov3-vitl16-pretrain-lvd1689m", "ViT-L/16", 1024),
+    ]
+
+    available_variants = []
+    for model_name, display_name, expected_dim in variants:
+        try:
+            print(f"Testing {display_name} ({model_name})...")
+            model = AutoModel.from_pretrained(model_name)
+            actual_dim = model.config.hidden_size
+
+            print(f"✓ {display_name}")
+            print(f"  Expected dim: {expected_dim}, Actual dim: {actual_dim}")
+            if actual_dim == expected_dim:
+                print(f"  ✓ Dimension matches expected")
+            else:
+                print(f"  ⚠ Dimension mismatch!")
+
+            available_variants.append((model_name, display_name, actual_dim))
+
+        except Exception as e:
+            print(f"✗ {display_name} failed: {e}")
+
+    return available_variants
+
+
+def eva_compatibility():
+    """compatibility with eva-probe patterns."""
+    print("=== Testing eva-probe Compatibility ===")
+
+    model_name = "facebook/dinov3-vits16-pretrain-lvd1689m"
+
+    try:
+        model = AutoModel.from_pretrained(model_name)
+        processor = AutoImageProcessor.from_pretrained(model_name)
+
+        # batch processing (similar to eva's batch sizes)
+        batch_size = 2
+        dummy_images = [Image.new('RGB', (224, 224), color=(i*50, 100, 150)) for i in range(batch_size)]
+
+        inputs = processor(images=dummy_images, return_tensors="pt")
+        print(f"  Batch input shape: {inputs.pixel_values.shape}")
+
+        model.eval()
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # both CLS token (classification) and patch tokens (segmentation) modes
+        if hasattr(outputs, 'pooler_output') and outputs.pooler_output is not None:
+            cls_features = outputs.pooler_output
+            print(f"✓ CLS features (classification): {cls_features.shape}")
+
+        if hasattr(outputs, 'last_hidden_state'):
+            # CLS token (first token)
+            cls_from_hidden = outputs.last_hidden_state[:, 0]
+            print(f"✓ CLS from hidden state: {cls_from_hidden.shape}")
+
+            # Patch tokens (remaining tokens)
+            patch_tokens = outputs.last_hidden_state[:, 1:]
+            print(f"✓ Patch tokens (segmentation): {patch_tokens.shape}")
+
+            # Calculate expected patch count (14x14 for 224x224 input with patch size 16)
+            expected_patches = (224 // 16) ** 2  # 196 patches
+            actual_patches = patch_tokens.shape[1]
+            print(f"  Expected patches: {expected_patches}, Actual patches: {actual_patches}")
+
+        return True
+
+    except Exception as e:
+        print(f"✗ eva-probe compatibility failed: {e}")
+        return False
+
+
+def main():
+    """Run all DINOv3 validation tests."""
+    print("DINOv3 Integration Validation for eva-probe")
+    print("=" * 50)
+
+    # basic functionality
+    basic_success = dinov3_basic()
+
+    # model variants
+    available_variants = dinov3_variants()
+
+    # eva compatibility
+    eva_success = eva_compatibility()
+
+    # Summary
+    print("" + "=" * 50)
+    print("VALIDATION SUMMARY")
+    print("=" * 50)
+    print(f"Basic functionality: {'✓ PASS' if basic_success else '✗ FAIL'}")
+    print(f"Available variants: {len(available_variants)}")
+    for model_name, display_name, dim in available_variants:
+        print(f"  - {display_name}: {dim} dimensions")
+    print(f"eva-probe compatibility: {'✓ PASS' if eva_success else '✗ FAIL'}")
+
+    if basic_success and eva_success and len(available_variants) > 0:
+        print("🎉 DINOv3 is ready for eva-probe integration!")
+    else:
+        print("⚠️  Some issues detected. Please review the output above.")
+
+
+if __name__ == "__main__":
+    main()

pyproject.toml

@@ -47,6 +47,8 @@ dependencies = [
     "rich>=13.7.1",
     "torchmetrics>=1.6.0",
     "nibabel>=3.2.2",
+    "huggingface-hub>=0.34.4",
+    "torchvision>=0.23.0",
 ]
 
 [project.urls]
@@ -237,3 +239,6 @@ omit = [
 
 [tool.coverage.comment]
 layout = "header, diff, flags, files, footer"
+
+[tool.uv.sources]
+transformers = { git = "https://github.com/huggingface/transformers.git" }