app.py

import gradio as gr
import timm
import torch
from PIL import Image
import requests
from io import BytesIO
import numpy as np
from pytorch_grad_cam import GradCAM, HiResCAM, ScoreCAM, GradCAMPlusPlus, AblationCAM, XGradCAM, EigenCAM, FullGrad
from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget
from pytorch_grad_cam.utils.image import show_cam_on_image
from timm.data import create_transform
from timm.data import infer_imagenet_subset, ImageNetInfo

# List of available timm models
MODELS = timm.list_pretrained()

# List of available GradCAM methods
CAM_METHODS = {
    "GradCAM": GradCAM,
    "HiResCAM": HiResCAM,
    "ScoreCAM": ScoreCAM,
    "GradCAM++": GradCAMPlusPlus,
    "AblationCAM": AblationCAM,
    "XGradCAM": XGradCAM,
    "EigenCAM": EigenCAM,
    "FullGrad": FullGrad
}

class CustomDatasetInfo:
    def __init__(self, label_names, label_descriptions=None):
        self.label_names = label_names
        self.label_descriptions = label_descriptions or label_names

    def index_to_description(self, index, detailed=False):
        if detailed and self.label_descriptions:
            return self.label_descriptions[index]
        return self.label_names[index]

def load_model(model_name):
    model = timm.create_model(model_name, pretrained=True)
    model.eval()
    return model

def process_image(image_path, model):
    if image_path.startswith('http'):
        response = requests.get(image_path)
        image = Image.open(BytesIO(response.content))
    else:
        image = Image.open(image_path)
    
    config = model.pretrained_cfg
    transform = create_transform(
        input_size=config['input_size'],
        crop_pct=config['crop_pct'],
        mean=config['mean'],
        std=config['std'],
        interpolation=config['interpolation'],
        is_training=False
    )
    
    tensor = transform(image).unsqueeze(0)
    return tensor

def get_cam_image(model, image, target_layer, cam_method, target_class):
    if target_class is not None and target_class != "highest scoring":
        target = ClassifierOutputTarget(target_class)
    else:
        target = None
    
    cam = CAM_METHODS[cam_method](model=model, target_layers=[target_layer])
    grayscale_cam = cam(input_tensor=image, targets=[target] if target else None)
    
    config = model.pretrained_cfg
    mean = torch.tensor(config['mean']).view(3, 1, 1)
    std = torch.tensor(config['std']).view(3, 1, 1)
    rgb_img = (image.squeeze(0) * std + mean).permute(1, 2, 0).cpu().numpy()
    rgb_img = np.clip(rgb_img, 0, 1)
    
    cam_image = show_cam_on_image(rgb_img, grayscale_cam[0, :], use_rgb=True)
    return Image.fromarray(cam_image)

def get_feature_info(model):
    if hasattr(model, 'feature_info'):
        return [f['module'] for f in model.feature_info]
    else:
        return []

def get_target_layer(model, target_layer_name):
    if target_layer_name is None:
        return None
    
    try:
        return model.get_submodule(target_layer_name)
    except AttributeError:
        print(f"WARNING: Layer '{target_layer_name}' not found in the model.")
        return None

def get_class_names(model):
    dataset_info = None
    label_names = model.pretrained_cfg.get("label_names", None)
    label_descriptions = model.pretrained_cfg.get("label_descriptions", None)
    if label_names is None:
        imagenet_subset = infer_imagenet_subset(model)
        if imagenet_subset:
            dataset_info = ImageNetInfo(imagenet_subset)
        else:
            label_names = [f"LABEL_{i}" for i in range(model.num_classes)]
    if dataset_info is None:
        dataset_info = CustomDatasetInfo(
            label_names=label_names,
            label_descriptions=label_descriptions,
        )
    return dataset_info

def explain_image(model_name, image_path, cam_method, feature_module, target_class):
    model = load_model(model_name)
    image = process_image(image_path, model)
    
    target_layer = get_target_layer(model, feature_module)
    
    if target_layer is None:
        feature_info = get_feature_info(model)
        if feature_info:
            target_layer = get_target_layer(model, feature_info[-1])
            print(f"Using last feature module: {feature_info[-1]}")
        else:
            for name, module in reversed(list(model.named_modules())):
                if isinstance(module, torch.nn.Conv2d):
                    target_layer = module
                    print(f"Fallback: Using last convolutional layer: {name}")
                    break
    
    if target_layer is None:
        raise ValueError("Could not find a suitable target layer.")
    
    target_class_index = None if target_class == "highest scoring" else int(target_class.split(':')[0])
    cam_image = get_cam_image(model, image, target_layer, cam_method, target_class_index)
    
    with torch.no_grad():
        out = model(image)
    probabilities = out.squeeze(0).softmax(dim=0)
    values, indices = torch.topk(probabilities, 5)  # Top 5 predictions
    dataset_info = get_class_names(model)
    labels = [
        f"{i}: {dataset_info.index_to_description(i.item(), detailed=True)} ({v.item():.2%})"
        for i, v in zip(indices, values)
    ]
    
    return cam_image, "\n".join(labels)

def update_feature_modules(model_name):
    model = load_model(model_name)
    feature_modules = get_feature_info(model)
    return gr.Dropdown(choices=feature_modules, value=feature_modules[-1] if feature_modules else None)

def update_class_dropdown(model_name):
    model = load_model(model_name)
    dataset_info = get_class_names(model)
    class_names = ["highest scoring"] + [f"{i}: {dataset_info.index_to_description(i, detailed=True)}" for i in range(model.num_classes)]
    return gr.Dropdown(choices=class_names, value="highest scoring")

with gr.Blocks() as demo:
    gr.Markdown("# Explainable AI with timm models. NOTE: This is a WIP but some models are functioning.")
    gr.Markdown("Upload an image, select a model, CAM method, and optionally a specific feature module and target class to visualize the explanation.")
    
    with gr.Row():
        with gr.Column():
            model_dropdown = gr.Dropdown(choices=MODELS, label="Select Model")
            image_input = gr.Image(type="filepath", label="Upload Image")
            cam_method_dropdown = gr.Dropdown(choices=list(CAM_METHODS.keys()), label="Select CAM Method")
            feature_module_dropdown = gr.Dropdown(label="Select Feature Module (optional)")
            class_dropdown = gr.Dropdown(label="Select Target Class (optional)")
            explain_button = gr.Button("Explain Image")
        
        with gr.Column():
            output_image = gr.Image(type="pil", label="Explained Image")
            prediction_text = gr.Textbox(label="Top 5 Predictions")
    
    model_dropdown.change(fn=update_feature_modules, inputs=[model_dropdown], outputs=[feature_module_dropdown])
    model_dropdown.change(fn=update_class_dropdown, inputs=[model_dropdown], outputs=[class_dropdown])
    
    explain_button.click(
        fn=explain_image,
        inputs=[model_dropdown, image_input, cam_method_dropdown, feature_module_dropdown, class_dropdown],
        outputs=[output_image, prediction_text]
    )

demo.launch()