clockclock's picture
Update app.py
304b7e6 verified
import gradio as gr
import torch
import torch.nn.functional as F # <-- ADD THIS IMPORT
from transformers import AutoModelForImageClassification, AutoImageProcessor
from PIL import Image
import numpy as np
from captum.attr import LayerGradCam
from captum.attr import visualization as viz
import requests
from io import BytesIO
# --- 1. Load Model and Processor ---
print("Loading model and processor...")
model_id = "Organika/sdxl-detector"
processor = AutoImageProcessor.from_pretrained(model_id)
model = AutoModelForImageClassification.from_pretrained(model_id, torch_dtype=torch.float32)
model.eval()
print("Model and processor loaded successfully.")
# --- 2. FINAL, CORRECTED Explainability (Grad-CAM) Function ---
def generate_heatmap(image_tensor, original_image, target_class_index):
# This wrapper is correct and necessary for Captum to work with Hugging Face models.
def model_forward_wrapper(input_tensor):
outputs = model(pixel_values=input_tensor)
return outputs.logits
# The target layer is also correct for the Swin Transformer.
target_layer = model.swin.layernorm
# Initialize LayerGradCam with the wrapper and the target layer.
lgc = LayerGradCam(model_forward_wrapper, target_layer)
# This call now works and returns the attributions.
attributions = lgc.attribute(image_tensor, target=target_class_index, relu_attributions=True)
# --- THIS IS THE FIX for the Transformer Architecture ---
# Transformer models output a sequence of patch attributions, not a 2D grid.
# We must reshape this sequence into a grid and then upsample it.
# 1. Determine the grid size (e.g., for 49 patches, it's 7x7)
# We remove the batch dimension, and get the number of patches (sequence length).
num_patches = attributions.shape[-1]
grid_size = int(np.sqrt(num_patches))
# 2. Reshape the 1D attributions into a 2D grid.
heatmap = attributions.squeeze(0).squeeze(0).reshape(grid_size, grid_size)
# 3. Upsample the small heatmap to match the original image size for overlay.
# We need to add batch and channel dimensions back for the interpolate function.
heatmap = heatmap.unsqueeze(0).unsqueeze(0)
# Note: original_image.size is (W, H), interpolate needs size as (H, W)
upsampled_heatmap = F.interpolate(heatmap, size=original_image.size[::-1], mode='bilinear', align_corners=False)
# 4. Prepare the final heatmap for visualization
heatmap_for_viz = upsampled_heatmap.squeeze().cpu().detach().numpy()
# The visualization function expects a (H, W, C) shaped numpy array.
# Our heatmap is (H, W), so we add a channel dimension.
visualized_image, _ = viz.visualize_image_attr(
np.expand_dims(heatmap_for_viz, axis=-1),
np.array(original_image),
method="blended_heat_map",
sign="all",
show_colorbar=True,
title="Model Attention Heatmap",
)
return visualized_image
# --- 3. Main Prediction Function (Unchanged) ---
def predict(image_upload: Image.Image, image_url: str):
if image_upload is not None:
input_image = image_upload
print(f"Processing uploaded image of size: {input_image.size}")
elif image_url:
try:
response = requests.get(image_url)
response.raise_for_status()
input_image = Image.open(BytesIO(response.content))
print(f"Processing image from URL: {image_url}")
except Exception as e:
raise gr.Error(f"Could not load image from URL. Please check the link. Error: {e}")
else:
raise gr.Error("Please upload an image or provide a URL to analyze.")
if input_image.mode == 'RGBA':
input_image = input_image.convert('RGB')
inputs = processor(images=input_image, return_tensors="pt")
with torch.no_grad():
outputs = model(**inputs)
logits = outputs.logits
probabilities = torch.nn.functional.softmax(logits, dim=-1)
predicted_class_idx = logits.argmax(-1).item()
confidence_score = probabilities[0][predicted_class_idx].item()
predicted_label = model.config.id2label[predicted_class_idx]
if predicted_label.lower() == 'ai':
explanation = (
f"The model is {confidence_score:.2%} confident that this image is AI-GENERATED.\n\n"
"The heatmap on the right highlights the areas that most influenced this decision. "
"Pay close attention if these hotspots are on unnatural-looking features like hair, eyes, skin texture, or strange background details."
)
else:
explanation = (
f"The model is {confidence_score:.2%} confident that this image is HUMAN-MADE.\n\n"
"The heatmap shows which areas the model found to be most 'natural'. "
"These are likely well-formed, realistic features that AI models often struggle to replicate perfectly."
)
print("Generating heatmap...")
heatmap_image = generate_heatmap(inputs['pixel_values'], input_image, predicted_class_idx)
print("Heatmap generated.")
labels_dict = {model.config.id2label[i]: float(probabilities[0][i]) for i in range(len(model.config.id2label))}
return labels_dict, explanation, heatmap_image
# --- 4. Gradio Interface (Unchanged) ---
with gr.Blocks(theme=gr.themes.Soft()) as demo:
gr.Markdown(
"""
# AI Image Detector with Explainability
Determine if an image is AI-generated or human-made. Upload a file or paste a URL.
This tool uses the [Organika/sdxl-detector](https://huggingface.co/Organika/sdxl-detector) model and provides a **heatmap** to show *why* the model made its decision.
"""
)
with gr.Row():
with gr.Column():
with gr.Tabs():
with gr.TabItem("Upload File"):
input_image_upload = gr.Image(type="pil", label="Upload Your Image")
with gr.TabItem("Use Image URL"):
input_image_url = gr.Textbox(label="Paste Image URL here")
submit_btn = gr.Button("Analyze Image", variant="primary")
with gr.Column():
output_label = gr.Label(label="Prediction")
output_text = gr.Textbox(label="Explanation", lines=6, interactive=False)
output_heatmap = gr.Image(label="Model Attention Heatmap")
submit_btn.click(
fn=predict,
inputs=[input_image_upload, input_image_url],
outputs=[output_label, output_text, output_heatmap]
)
# --- 5. Launch the App (Unchanged) ---
if __name__ == "__main__":
demo.launch(debug=True)