# app.py import torch import gradio as gr import numpy as np from PIL import Image import torchvision.transforms.functional as TF from matplotlib import colormaps from transformers import AutoModel import os # ---------------------------- # Configuration # ---------------------------- # The model will be downloaded from the Hugging Face Hub # Using the specific revision that works well with transformers AutoModel MODEL_ID = "facebook/dinov3-vith16plus" PATCH_SIZE = 16 DEVICE = "cuda" if torch.cuda.is_available() else "cpu" # Normalization constants (standard for ImageNet) IMAGENET_MEAN = (0.485, 0.456, 0.406) IMAGENET_STD = (0.229, 0.224, 0.225) # ---------------------------- # Model Loading (runs once at startup) # ---------------------------- def load_model_from_hub(): """Loads the DINOv3 model from the Hugging Face Hub.""" print(f"Loading model '{MODEL_ID}' from Hugging Face Hub...") try: # Use your HF token if the model is gated # You can set this as a secret in your Hugging Face Space settings token = os.environ.get("HF_TOKEN") model = AutoModel.from_pretrained(MODEL_ID, token=token, trust_remote_code=True) model.to(DEVICE).eval() print(f"✅ Model loaded successfully on device: {DEVICE}") return model except Exception as e: print(f"❌ Failed to load model: {e}") # This will display an error message in the Gradio interface raise gr.Error(f"Could not load model from Hub. If it's a gated model, ensure you have access and have set your HF_TOKEN secret in the Space settings. Error: {e}") # Load the model globally when the app starts model = load_model_from_hub() # ---------------------------- # Helper Functions # ---------------------------- def resize_to_grid(img: Image.Image, long_side: int, patch: int) -> torch.Tensor: """Resizes an image to dimensions that are multiples of the patch size.""" w, h = img.size scale = long_side / max(h, w) new_h = max(patch, int(round(h * scale))) new_w = max(patch, int(round(w * scale))) new_h = ((new_h + patch - 1) // patch) * patch new_w = ((new_w + patch - 1) // patch) * patch return TF.to_tensor(TF.resize(img.convert("RGB"), (new_h, new_w))) def colorize(data: np.ndarray, cmap_name: str = 'viridis') -> Image.Image: """Converts a 2D numpy array to a colored PIL image.""" x = data.astype(np.float32) x = (x - x.min()) / (x.max() - x.min() + 1e-8) cmap = colormaps.get_cmap(cmap_name) rgb = (cmap(x)[..., :3] * 255).astype(np.uint8) return Image.fromarray(rgb) def blend(base: Image.Image, heat: Image.Image, alpha: float) -> Image.Image: """Blends a heatmap onto a base image.""" base = base.convert("RGBA") heat = heat.convert("RGBA") return Image.blend(base, heat, alpha=alpha) # ---------------------------- # Core Gradio Function # ---------------------------- @torch.inference_mode() def generate_pca_visuals( image_pil: Image.Image, resolution: int, cmap_name: str, overlay_alpha: float, progress=gr.Progress(track_tqdm=True) ): """Main function to generate PCA visuals.""" if model is None: raise gr.Error("DINOv3 model is not available. Check the startup logs.") if image_pil is None: return None, None, "Please upload an image and click Generate.", None, None # 1. Image Preprocessing progress(0.2, desc="Resizing and preprocessing image...") image_tensor = resize_to_grid(image_pil, resolution, PATCH_SIZE) t_norm = TF.normalize(image_tensor, IMAGENET_MEAN, IMAGENET_STD).unsqueeze(0).to(DEVICE) original_processed_image = TF.to_pil_image(image_tensor) _, _, H, W = t_norm.shape Hp, Wp = H // PATCH_SIZE, W // PATCH_SIZE # 2. Feature Extraction progress(0.5, desc="🦖 Extracting features with DINOv3...") outputs = model(t_norm) # 💡 FIX: The model output includes a [CLS] token AND 4 register tokens. # We must skip all of them (total 5) to get only the patch embeddings. # The original code only skipped 1, causing the size mismatch. n_special_tokens = 5 # 1 [CLS] token + 4 register tokens patch_embeddings = outputs.last_hidden_state.squeeze(0)[n_special_tokens:, :] # 3. PCA Calculation progress(0.8, desc="🔬 Performing PCA...") X_centered = patch_embeddings.float() - patch_embeddings.float().mean(0, keepdim=True) U, S, V = torch.pca_lowrank(X_centered, q=3, center=False) # 💡 IMPROVEMENT: Stabilize the signs of the eigenvectors for deterministic output. # This prevents the colors from randomly inverting on different runs. for i in range(V.shape[1]): max_abs_idx = torch.argmax(torch.abs(V[:, i])) if V[max_abs_idx, i] < 0: V[:, i] *= -1 scores = X_centered @ V[:, :3] # 4. Explained Variance total_variance = (X_centered ** 2).sum() explained_variance = [float((s**2) / total_variance) for s in S] variance_text = ( f"**📊 Explained Variance Ratios:**\n\n" f"- **PC1:** {explained_variance[0]:.2%}\n" f"- **PC2:** {explained_variance[1]:.2%}\n" f"- **PC3:** {explained_variance[2]:.2%}" ) # 5. Create Visualizations # This part should now work correctly as `scores` has the right shape (Hp*Wp, 3) pc1_map = scores[:, 0].reshape(Hp, Wp).cpu().numpy() pc1_image_raw = colorize(pc1_map, cmap_name) pc_rgb_map = scores.reshape(Hp, Wp, 3).cpu().numpy() min_vals = pc_rgb_map.reshape(-1, 3).min(axis=0) max_vals = pc_rgb_map.reshape(-1, 3).max(axis=0) pc_rgb_map = (pc_rgb_map - min_vals) / (max_vals - min_vals + 1e-8) pc_rgb_image_raw = Image.fromarray((pc_rgb_map * 255).astype(np.uint8)) target_size = original_processed_image.size pc1_image_smooth = pc1_image_raw.resize(target_size, Image.Resampling.BICUBIC) pc_rgb_image_smooth = pc_rgb_image_raw.resize(target_size, Image.Resampling.BICUBIC) blended_image = blend(original_processed_image, pc1_image_smooth, overlay_alpha) progress(1.0, desc="✅ Done!") return pc1_image_smooth, pc_rgb_image_smooth, variance_text, blended_image, original_processed_image # ---------------------------- # Gradio Interface # ---------------------------- with gr.Blocks(theme=gr.themes.Soft(), title="DINOv3 PCA Explorer") as demo: gr.Markdown( """ # 🦖 DINOv3 PCA Explorer Upload an image to visualize the principal components of its patch features. This reveals the main axes of semantic variation within the image as understood by the model. """ ) with gr.Row(): with gr.Column(scale=2): # Added a default image URL for convenience input_image = gr.Image(type="pil", label="Upload Image", value="https://images.squarespace-cdn.com/content/v1/607f89e638219e13eee71b1e/1684821560422-SD5V37BAG28BURTLIXUQ/michael-sum-LEpfefQf4rU-unsplash.jpg") with gr.Accordion("⚙️ Visualization Controls", open=True): resolution_slider = gr.Slider( minimum=224, maximum=1024, value=512, step=16, label="Processing Resolution", info="Higher values capture more detail but are slower." ) cmap_dropdown = gr.Dropdown( ['viridis', 'magma', 'inferno', 'plasma', 'cividis', 'jet'], value='viridis', label="Heatmap Colormap" ) alpha_slider = gr.Slider( minimum=0, maximum=1, value=0.5, label="Overlay Opacity" ) run_button = gr.Button("🚀 Generate PCA Visuals", variant="primary") with gr.Column(scale=3): with gr.Tabs(): with gr.TabItem("🖼️ Overlay"): gr.Markdown("Visualize the main heatmap blended with the original image.") output_blended = gr.Image(label="PC1 Heatmap Overlay") output_processed = gr.Image(label="Original Processed Image (at selected resolution)") with gr.TabItem("📊 PCA Outputs"): gr.Markdown("View the raw outputs of the Principal Component Analysis.") output_pc1 = gr.Image(label="PC1 Heatmap (Smoothed)") output_rgb = gr.Image(label="Top 3 PCs as RGB (Smoothed)") output_variance = gr.Markdown(label="Explained Variance") run_button.click( fn=generate_pca_visuals, inputs=[input_image, resolution_slider, cmap_dropdown, alpha_slider], outputs=[output_pc1, output_rgb, output_variance, output_blended, output_processed] ) if __name__ == "__main__": demo.launch()