app.py

import os
import numpy as np
import gradio as gr
import cv2
import tensorflow as tf
import keras
from keras.models import Model
from huggingface_hub import hf_hub_download
import pandas as pd
from PIL import Image

# ---- Configuration ----
CLASS_NAMES = ['akiec', 'bcc', 'bkl', 'df', 'nv', 'vasc', 'mel']
label_to_index = {name: i for i, name in enumerate(CLASS_NAMES)}
diagnosis_map = {
    'akiec': 'Bénin', 'bcc': 'Malin', 'bkl': 'Bénin', 'df': 'Bénin',
    'nv': 'Bénin', 'vasc': 'Bénin', 'mel': 'Malin'
}

# ---- Téléchargement modèles ----
resnet_path = hf_hub_download(repo_id="ericjedha/resnet50", filename="Resnet50.keras")
densenet_path = hf_hub_download(repo_id="ericjedha/densenet201", filename="Densenet201.keras")

model_resnet50 = keras.saving.load_model(resnet_path, compile=False)
model_densenet = keras.saving.load_model(densenet_path, compile=False)
model_xcept = keras.saving.load_model("Xception.keras", compile=False)

# ---- Préprocesseurs ----
from tensorflow.keras.applications.xception import preprocess_input as preprocess_xception
from tensorflow.keras.applications.resnet50 import preprocess_input as preprocess_resnet
from tensorflow.keras.applications.densenet import preprocess_input as preprocess_densenet

# ---- Fonctions utilitaires robustes pour les modèles Keras ----
def get_primary_input_name(model):
    if isinstance(model.inputs, list) and len(model.inputs) > 0:
        return model.inputs[0].name.split(':')[0]
    return "input_1" 

def safe_forward(model, x):
    input_name = get_primary_input_name(model)
    return model({input_name: x}, training=False).numpy()

# ---- Prédiction ----
def predict_single(image_pil, weights=(0.45, 0.25, 0.30)):
    img_np = np.array(image_pil)
    img_299 = cv2.resize(img_np, (299, 299))
    img_224 = cv2.resize(img_np, (224, 224))
    
    bx = preprocess_xception(np.expand_dims(img_299, axis=0))
    br = preprocess_resnet(np.expand_dims(img_224, axis=0))
    bd = preprocess_densenet(np.expand_dims(img_224, axis=0))

    pred_x = safe_forward(model_xcept, bx)
    pred_r = safe_forward(model_resnet50, br)
    pred_d = safe_forward(model_densenet, bd)

    preds = (weights[0] * pred_x + weights[1] * pred_r + weights[2] * pred_d)
    mel_idx = label_to_index['mel']
    preds[:, mel_idx] = (0.5 * preds[:, mel_idx] + 0.5 * pred_d[:, mel_idx])
    return preds[0]


# ---- Grad-CAM ----
def make_gradcam(image_pil, model, last_conv_layer_name="conv5_block32_concat", class_index=None):
    img_np = np.array(image_pil)
    img_resized = cv2.resize(img_np, (224, 224))
    img_array_preprocessed = preprocess_densenet(np.expand_dims(img_resized, axis=0))
    
    grad_model = Model(model.inputs, [model.get_layer(last_conv_layer_name).output, model.output])
    
    input_name = get_primary_input_name(model)
    input_for_model = {input_name: img_array_preprocessed}

    with tf.GradientTape() as tape:
        last_conv_layer_output, preds = grad_model(input_for_model, training=False)
        if isinstance(preds, list):
            preds = preds[0]
        if class_index is None:
            class_index = tf.argmax(preds[0])
        class_channel = preds[:, class_index]

    grads = tape.gradient(class_channel, last_conv_layer_output)
    
    # ===== LA CORRECTION FINALE EST ICI : on ne moyenne que sur les axes spatiaux =====
    pooled_grads = tf.reduce_mean(grads, axis=(0, 1))
    # ===================================================================================
    
    last_conv_layer_output = last_conv_layer_output[0]
    heatmap = last_conv_layer_output @ pooled_grads[..., tf.newaxis]
    heatmap = tf.squeeze(heatmap)
    heatmap = tf.maximum(heatmap, 0) / (tf.math.reduce_max(heatmap) + 1e-8)
    heatmap = heatmap.numpy()

    heatmap = np.uint8(255 * heatmap)
    # applyColorMap prend une image 1 canal et retourne une image 3 canaux (BGR)
    heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET)
    
    img_bgr = cv2.cvtColor(img_resized, cv2.COLOR_RGB2BGR)
    # Les deux images ont maintenant la même taille (224, 224) et 3 canaux.
    superimposed_img = cv2.addWeighted(img_bgr, 0.6, heatmap, 0.4, 0)
    
    return cv2.cvtColor(superimposed_img, cv2.COLOR_BGR2RGB)

# ---- Fonction Gradio ----
def gradio_predict(image_pil):
    if image_pil is None:
        return "Veuillez uploader une image.", None, None
    try:
        probs = predict_single(image_pil)
        benign_prob = sum(probs[i] for i, cls in enumerate(CLASS_NAMES) if diagnosis_map[cls] == "Bénin")
        malign_prob = sum(probs[i] for i, cls in enumerate(CLASS_NAMES) if diagnosis_map[cls] == "Malin")
        global_diag = "Bénin" if benign_prob >= malign_prob else "Malin"
        
        confidences = {CLASS_NAMES[i]: float(probs[i]) for i in range(len(CLASS_NAMES))}
        df = pd.DataFrame.from_dict(confidences, orient='index', columns=['Probabilité'])
        df = df.sort_values(by='Probabilité', ascending=False)
        df.index.name = "Classe"
        df = df.reset_index()

        top_class_idx = np.argmax(probs)
        gradcam_img = make_gradcam(image_pil, model_densenet, class_index=top_class_idx)

        return global_diag, df, gradcam_img
    except Exception as e:
        print(f"Erreur dans gradio_predict : {e}")
        import traceback
        traceback.print_exc()
        return "Erreur lors du traitement de l'image.", None, None

# ---- Gradio UI ----
example_paths = ["exemple1.jpg", "exemple2.jpg", "exemple3.jpg"]
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# Analyse de lésions cutanées (Ensemble de modèles + Grad-CAM)")
    gr.Markdown("Cet outil propose une prédiction de la nature de la lésion (Bénin/Malin) avec explication visuelle.")
    with gr.Row():
        with gr.Column(scale=1):
            input_image = gr.Image(type="pil", label="Uploader une image de lésion")
            submit_btn = gr.Button("Analyser", variant="primary")
            gr.Examples(examples=example_paths, inputs=input_image)
        with gr.Column(scale=1):
            output_label = gr.Label(label="Diagnostic global")
            output_plot = gr.BarPlot(label="Probabilités par classe", x="Classe", y="Probabilité", y_lim=[0, 1])
            output_gradcam = gr.Image(label="Visualisation Grad-CAM")
    submit_btn.click(fn=gradio_predict, inputs=input_image, outputs=[output_label, output_plot, output_gradcam])

if __name__ == "__main__":
    demo.launch()