pszemraj's picture
refactor to satellite-based dinoV3
252fe2c verified
raw
history blame
11.1 kB
import gc
from pathlib import Path
import gradio as gr
import matplotlib.cm as cm
import numpy as np
import spaces
import torch
import torch.nn.functional as F
from PIL import Image, ImageOps
from transformers import AutoImageProcessor, AutoModel
# Device configuration with memory management
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
MODEL_MAP = {
"DINOv3 ViT-L/16 Satellite": "facebook/dinov3-vitl16-pretrain-sat493m",
"DINOv3 ViT-L/16 LVD (General Web)": "facebook/dinov3-vitl16-pretrain-lvd1689m",
"⚠️ DINOv3 ViT-7B/16 Satellite": "facebook/dinov3-vit7b16-pretrain-sat493m",
}
DEFAULT_NAME = list(MODEL_MAP.keys())[0]
# Global model state
processor = None
model = None
def cleanup_memory():
"""Aggressive memory cleanup for model switching"""
global processor, model
if model is not None:
del model
model = None
if processor is not None:
del processor
processor = None
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
torch.cuda.synchronize()
def load_model(name):
"""Load model with proper memory management and dtype handling"""
global processor, model
try:
# Clean up existing model
cleanup_memory()
model_id = MODEL_MAP[name]
# Load with auto dtype for optimal performance
processor = AutoImageProcessor.from_pretrained(model_id)
# Determine optimal dtype based on model size and hardware
if "7b" in model_id.lower() and torch.cuda.is_available():
# For 7B model, use bfloat16 if available for memory efficiency
dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
else:
dtype = torch.float32
model = AutoModel.from_pretrained(
model_id,
torch_dtype=dtype,
device_map="auto" if torch.cuda.is_available() else None,
)
if DEVICE == "cuda" and not hasattr(model, "device_map"):
model = model.to(DEVICE)
model.eval()
# Get model info
param_count = sum(p.numel() for p in model.parameters()) / 1e9
dtype_str = str(dtype).split(".")[-1]
return f"✅ Loaded: {name} | {param_count:.1f}B params | {dtype_str} | {DEVICE.upper()}"
except Exception as e:
cleanup_memory()
return f"❌ Failed to load {name}: {str(e)}"
# Initialize default model
load_model(DEFAULT_NAME)
@spaces.GPU(duration=60)
def _extract_grid(img):
"""Extract feature grid from image"""
with torch.inference_mode():
pv = processor(images=img, return_tensors="pt").pixel_values
if DEVICE == "cuda":
pv = pv.to(DEVICE)
out = model(pixel_values=pv)
last = out.last_hidden_state[0].to(torch.float32)
num_reg = getattr(model.config, "num_register_tokens", 0)
p = model.config.patch_size
_, _, Ht, Wt = pv.shape
gh, gw = Ht // p, Wt // p
feats = last[1 + num_reg :, :].reshape(gh, gw, -1).cpu()
return feats, gh, gw
def _overlay(orig, heat01, alpha=0.55, box=None):
"""Create heatmap overlay with improved visualization"""
H, W = orig.height, orig.width
heat = Image.fromarray((heat01 * 255).astype(np.uint8)).resize(
(W, H), resample=Image.LANCZOS
)
# Use a better colormap for satellite imagery
rgba = (cm.get_cmap("turbo")(np.asarray(heat) / 255.0) * 255).astype(np.uint8)
ov = Image.fromarray(rgba, "RGBA")
ov.putalpha(int(alpha * 255))
base = orig.copy().convert("RGBA")
out = Image.alpha_composite(base, ov)
if box:
from PIL import ImageDraw
draw = ImageDraw.Draw(out, "RGBA")
# Enhanced box visualization
draw.rectangle(box, outline=(255, 255, 255, 255), width=3)
draw.rectangle(
(box[0] - 1, box[1] - 1, box[2] + 1, box[3] + 1),
outline=(0, 0, 0, 200),
width=1,
)
return out
def prepare(img):
"""Prepare image and extract features"""
if img is None:
return None
base = ImageOps.exif_transpose(img.convert("RGB"))
feats, gh, gw = _extract_grid(base)
return {"orig": base, "feats": feats, "gh": gh, "gw": gw}
def click(state, opacity, colormap, img_value, evt: gr.SelectData):
"""Handle click events for similarity visualization"""
# If state wasn't prepared, build it now
if state is None and img_value is not None:
state = prepare(img_value)
if not state or evt.index is None:
return img_value, state, None
base, feats, gh, gw = state["orig"], state["feats"], state["gh"], state["gw"]
x, y = evt.index
px_x, px_y = base.width / gw, base.height / gh
i = min(int(x // px_x), gw - 1)
j = min(int(y // px_y), gh - 1)
d = feats.shape[-1]
grid = F.normalize(feats.reshape(-1, d), dim=1)
v = F.normalize(feats[j, i].reshape(1, d), dim=1)
sims = (grid @ v.T).reshape(gh, gw).numpy()
smin, smax = float(sims.min()), float(sims.max())
heat01 = (sims - smin) / (smax - smin + 1e-12)
# Update colormap dynamically
cm_func = cm.get_cmap(colormap.lower())
rgba = (cm_func(heat01) * 255).astype(np.uint8)
ov = Image.fromarray(rgba, "RGBA")
ov.putalpha(int(opacity * 255))
base_rgba = base.copy().convert("RGBA")
box = (int(i * px_x), int(j * px_y), int((i + 1) * px_x), int((j + 1) * px_y))
out = Image.alpha_composite(base_rgba, ov)
if box:
from PIL import ImageDraw
draw = ImageDraw.Draw(out, "RGBA")
draw.rectangle(box, outline=(255, 255, 255, 255), width=3)
draw.rectangle(
(box[0] - 1, box[1] - 1, box[2] + 1, box[3] + 1),
outline=(0, 0, 0, 200),
width=1,
)
# Stats for info panel
stats = f"""📊 **Similarity Statistics**
- Min: {smin:.3f}
- Max: {smax:.3f}
- Range: {smax - smin:.3f}
- Patch: ({i}, {j})
- Grid: {gw}×{gh}"""
return out, state, stats
def reset():
"""Reset the interface"""
return None, None, None
# Build the interface
with gr.Blocks(
theme=gr.themes.Soft(
primary_hue="blue",
secondary_hue="gray",
neutral_hue="gray",
font=gr.themes.GoogleFont("Inter"),
),
css="""
.container {max-width: 1200px; margin: auto;}
.header {text-align: center; padding: 20px;}
.info-box {
background: rgba(0,0,0,0.03);
border-radius: 8px;
padding: 12px;
margin: 10px 0;
border-left: 4px solid #2563eb;
}
""",
) as demo:
gr.HTML(
"""
<div class="header">
<h1>🛰️ DINOv3 Satellite Vision: Interactive Patch Similarity</h1>
<p style="font-size: 1.1em; color: #666;">
Explore how DINOv3 models trained on satellite imagery understand visual patterns
</p>
</div>
"""
)
with gr.Row():
with gr.Column(scale=3):
gr.Markdown(
"""
### How it works
1. **Select a model** - Satellite-pretrained models are optimized for aerial/satellite imagery
2. **Upload or select an image** - Works best with satellite, aerial, or outdoor scenes
3. **Click any region** - See how similar other patches are to your selection
4. **Adjust visualization** - Fine-tune opacity and colormap for clarity
"""
)
with gr.Column(scale=2):
gr.HTML(
"""
<div class="info-box">
<b>💡 Model Info:</b><br>
• <b>Satellite models</b>: Trained on 493M satellite images<br>
• <b>LVD model</b>: Trained on 1.7B diverse images<br>
• <b>7B model</b>: Massive capacity, slower but more nuanced
</div>
"""
)
with gr.Row():
with gr.Column(scale=1):
model_choice = gr.Dropdown(
choices=list(MODEL_MAP.keys()),
value=DEFAULT_NAME,
label="🤖 Model Selection",
info="Satellite models excel at geographic and structural patterns",
)
status = gr.Textbox(
label="📡 Model Status",
value=f"Ready: {DEFAULT_NAME}",
interactive=False,
lines=1,
)
with gr.Row():
opacity = gr.Slider(
0.2,
0.9,
0.55,
step=0.05,
label="🎨 Heatmap Opacity",
info="Balance between image and similarity map",
)
colormap = gr.Dropdown(
choices=["Turbo", "Inferno", "Viridis", "Plasma", "Magma", "Jet"],
value="Turbo",
label="🌈 Colormap",
info="Different maps for different contrasts",
)
info_panel = gr.Markdown(value=None, label="Statistics", visible=True)
with gr.Row():
reset_btn = gr.Button("🔄 Reset", variant="secondary", scale=1)
clear_btn = gr.ClearButton(value="🗑️ Clear All", scale=1)
with gr.Column(scale=2):
img = gr.Image(
type="pil",
label="Interactive Canvas (Click to explore)",
interactive=True,
height=600,
show_download_button=True,
show_share_button=False,
)
state = gr.State()
# Examples focused on satellite-relevant imagery
gr.Examples(
examples=[
[_filepath.name]
for _filepath in Path.cwd().iterdir()
if _filepath.suffix.lower() in [".jpg", ".png", ".webp"]
],
inputs=img,
fn=prepare,
outputs=[state],
label="Example Images",
examples_per_page=6,
cache_examples=False,
)
# Event handlers
model_choice.change(
load_model, inputs=model_choice, outputs=status, show_progress="full"
)
img.upload(prepare, inputs=img, outputs=state, show_progress="minimal")
img.select(
click,
inputs=[state, opacity, colormap, img],
outputs=[img, state, info_panel],
show_progress="minimal",
)
reset_btn.click(reset, outputs=[img, state, info_panel], show_progress=False)
clear_btn.add([img, state, info_panel])
gr.Markdown(
"""
---
<div style="text-align: center; color: #666; font-size: 0.9em;">
<b>Performance Notes:</b> Satellite models are optimized for geographic patterns, land use classification,
and structural analysis. The 7B model provides exceptional detail but requires significant compute.
<br><br>
Built with DINOv3 | Optimized for satellite and aerial imagery analysis
</div>
"""
)
if __name__ == "__main__":
demo.launch(share=False, show_error=True)