app.py

import gradio as gr
import torch
from transformers import SegformerForSemanticSegmentation, SegformerImageProcessor
from PIL import Image
import plotly.graph_objects as go
import numpy as np
import os
import torch.nn as nn
from sklearn.metrics import jaccard_score, accuracy_score
from collections import Counter
import matplotlib.pyplot as plt
import seaborn as sns
import torch.nn.functional as F
import seaborn as sns
from functools import partial
from pytorch_grad_cam.utils.image import (
    show_cam_on_image,
    preprocess_image as grad_preprocess,
)
from pytorch_grad_cam import GradCAM
import cv2
import transformers
from torchvision import transforms
import albumentations as A

device = "cuda" if torch.cuda.is_available() else "cpu"
data_folder = "data_sample"
id2label = {
    0: "void",
    1: "flat",
    2: "construction",
    3: "object",
    4: "nature",
    5: "sky",
    6: "human",
    7: "vehicle",
}
label2id = {v: k for k, v in id2label.items()}
num_labels = len(id2label)
checkpoint = "nvidia/segformer-b4-finetuned-cityscapes-1024-1024"
image_processor = SegformerImageProcessor()
state_dict_path = f"runs/{checkpoint}_v1/best_model.pt"
model = SegformerForSemanticSegmentation.from_pretrained(
    checkpoint,
    num_labels=num_labels,
    id2label=id2label,
    label2id=label2id,
    ignore_mismatched_sizes=True,
)
loaded_state_dict = torch.load(
    state_dict_path, map_location=torch.device("cpu"), weights_only=True
)
model.load_state_dict(loaded_state_dict)
model = model.to(device)
model.eval()

# ---- Partie Segmentation


def load_and_prepare_images(image_name, segformer=False):
    image_path = os.path.join(data_folder, "images", image_name)
    mask_name = image_name.replace("_leftImg8bit.png", "_gtFine_labelIds.png")
    mask_path = os.path.join(data_folder, "masks", mask_name)
    fpn_pred_path = os.path.join(data_folder, "resnet101_mask", image_name)

    if not os.path.exists(image_path):
        raise FileNotFoundError(f"Image not found: {image_path}")
    if not os.path.exists(mask_path):
        raise FileNotFoundError(f"Mask not found: {mask_path}")
    if not os.path.exists(fpn_pred_path):
        raise FileNotFoundError(f"FPN prediction not found: {fpn_pred_path}")

    original_image = Image.open(image_path).convert("RGB")
    original = original_image.resize((1024, 512))
    true_mask = np.array(Image.open(mask_path))
    fpn_pred = np.array(Image.open(fpn_pred_path))
    if segformer:
        segformer_pred = predict_segmentation(original)
        return original, true_mask, fpn_pred, segformer_pred

    return original, true_mask, fpn_pred


def predict_segmentation(image):
    # Charger et préparer l'image
    inputs = image_processor(images=image, return_tensors="pt")

    # Utiliser GPU si disponible
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.to(device)

    # Déplacer les inputs sur le bon device et faire la prédiction
    pixel_values = inputs.pixel_values.to(device)

    with torch.no_grad():  # Désactiver le calcul des gradients pour l'inférence
        outputs = model(pixel_values=pixel_values)
        logits = outputs.logits

    # Redimensionner les logits à la taille de l'image d'origine
    upsampled_logits = nn.functional.interpolate(
        logits,
        size=image.size[::-1],  # (height, width)
        mode="bilinear",
        align_corners=False,
    )

    # Obtenir la prédiction finale
    pred_seg = upsampled_logits.argmax(dim=1)[0].cpu().numpy()

    return pred_seg


def process_image(image_name):
    original, true_mask, fpn_pred, segformer_pred = load_and_prepare_images(
        image_name, segformer=True
    )
    true_mask_colored = colorize_mask(true_mask)
    true_mask_colored = Image.fromarray(true_mask_colored.astype("uint8"))
    true_mask_colored = true_mask_colored.resize((1024, 512))
    # fpn_pred_colored = colorize_mask(fpn_pred)
    segformer_pred_colored = colorize_mask(segformer_pred)
    segformer_pred_colored = Image.fromarray(segformer_pred_colored.astype("uint8"))
    segformer_pred_colored = segformer_pred_colored.resize((1024, 512))

    return [
        (original, "Image originale"),
        (true_mask_colored, "Masque réel"),
        (fpn_pred, "Prédiction FPN"),
        (segformer_pred_colored, "Prédiction SegFormer"),
    ]


def create_cityscapes_label_colormap():
    colormap = np.zeros((256, 3), dtype=np.uint8)
    colormap[0] = [78, 82, 110]
    colormap[1] = [128, 64, 128]
    colormap[2] = [154, 156, 153]
    colormap[3] = [168, 167, 18]
    colormap[4] = [80, 108, 28]
    colormap[5] = [112, 164, 196]
    colormap[6] = [168, 28, 52]
    colormap[7] = [16, 18, 112]
    return colormap


# Créer la colormap une fois
cityscapes_colormap = create_cityscapes_label_colormap()


def blend_images(original_image, colored_segmentation, alpha=0.6):
    blended_image = Image.blend(original_image, colored_segmentation, alpha)
    return blended_image


def colorize_mask(mask):
    return cityscapes_colormap[mask]


# ---- Fin Partie Segmentation

# def compare_masks(real_mask, fpn_mask, segformer_mask):
#     """
#     Compare les masques prédits par FPN et SegFormer avec le masque réel.
#     Retourne un score IoU et une précision pixel par pixel pour chaque modèle.

#     Args:
#     real_mask (np.array): Le masque réel de référence
#     fpn_mask (np.array): Le masque prédit par le modèle FPN
#     segformer_mask (np.array): Le masque prédit par le modèle SegFormer

#     Returns:
#     dict: Dictionnaire contenant les scores IoU et les précisions pour chaque modèle
#     """

#     assert real_mask.shape == fpn_mask.shape == segformer_mask.shape, "Les masques doivent avoir la même forme"

#     real_flat = real_mask.flatten()
#     fpn_flat = fpn_mask.flatten()
#     segformer_flat = segformer_mask.flatten()

#     # Calcul du score de Jaccard (IoU)
#     iou_fpn = jaccard_score(real_flat, fpn_flat, average='weighted')
#     iou_segformer = jaccard_score(real_flat, segformer_flat, average='weighted')

#     # Calcul de la précision pixel par pixel
#     accuracy_fpn = accuracy_score(real_flat, fpn_flat)
#     accuracy_segformer = accuracy_score(real_flat, segformer_flat)

#     return {
#         'FPN': {'IoU': iou_fpn, 'Precision': accuracy_fpn},
#         'SegFormer': {'IoU': iou_segformer, 'Precision': accuracy_segformer}
#     }

# ---- Partie EDA


def analyse_mask(real_mask, num_labels):
    # Compter les occurrences de chaque classe
    counts = np.bincount(real_mask.ravel(), minlength=num_labels)

    # Calculer le nombre total de pixels
    total_pixels = real_mask.size

    # Calculer les proportions
    class_proportions = counts / total_pixels

    # Créer un dictionnaire avec les proportions
    return dict(enumerate(class_proportions))


def show_eda(image_name):
    original_image, true_mask, _ = load_and_prepare_images(image_name)
    class_proportions = analyse_mask(true_mask, num_labels)
    cityscapes_colormap = create_cityscapes_label_colormap()
    true_mask_colored = colorize_mask(true_mask)
    true_mask_colored = Image.fromarray(true_mask_colored.astype("uint8"))
    true_mask_colored = true_mask_colored.resize((1024, 512))

    # Trier les classes par proportion croissante
    sorted_classes = sorted(
        class_proportions.keys(), key=lambda x: class_proportions[x]
    )

    # Préparer les données pour le barplot
    categories = [id2label[i] for i in sorted_classes]
    values = [class_proportions[i] for i in sorted_classes]
    color_list = [
        f"rgb({cityscapes_colormap[i][0]}, {cityscapes_colormap[i][1]}, {cityscapes_colormap[i][2]})"
        for i in sorted_classes
    ]

    # Distribution des classes avec la colormap personnalisée
    fig = go.Figure()

    fig.add_trace(
        go.Bar(
            x=categories,
            y=values,
            marker_color=color_list,
            text=[f"{v:.2f}" for v in values],
            textposition="outside",
        )
    )

    # Ajouter un titre et des labels, modifier la rotation et la taille de la police
    fig.update_layout(
        title={"text": "Distribution des classes", "font": {"size": 24}},
        xaxis_title={"text": "Catégories", "font": {"size": 18}},
        yaxis_title={"text": "Proportion", "font": {"size": 18}},
        xaxis_tickangle=0,  # Rotation modifiée à -45 degrés
        uniformtext_minsize=12,
        uniformtext_mode="hide",
        font=dict(size=14),
        autosize=True,
        bargap=0.2,
        height=600,
        margin=dict(l=20, r=20, t=50, b=20),
    )

    return original_image, true_mask_colored, fig


# ----Fin Partie EDA

# ----Partie Explication GradCam


class SegformerWrapper(nn.Module):
    def __init__(self, model):
        super().__init__()
        self.model = model

    def forward(self, x):
        output = self.model(x)
        return output.logits


class SemanticSegmentationTarget:
    def __init__(self, category, mask):
        self.category = category
        self.mask = torch.from_numpy(mask)
        if torch.cuda.is_available():
            self.mask = self.mask.cuda()

    def __call__(self, model_output):
        if isinstance(
            model_output, (dict, transformers.modeling_outputs.SemanticSegmenterOutput)
        ):
            logits = (
                model_output["logits"]
                if isinstance(model_output, dict)
                else model_output.logits
            )
        elif isinstance(model_output, torch.Tensor):
            logits = model_output
        else:
            raise ValueError(f"Unexpected model_output type: {type(model_output)}")

        if logits.dim() == 4:  # [batch, classes, height, width]
            return (logits[0, self.category, :, :] * self.mask).sum()
        elif logits.dim() == 3:  # [classes, height, width]
            return (logits[self.category, :, :] * self.mask).sum()
        else:
            raise ValueError(f"Unexpected logits shape: {logits.shape}")


def segformer_reshape_transform_huggingface(tensor, width, height):
    result = tensor.reshape(tensor.size(0), height, width, tensor.size(2))
    result = result.transpose(2, 3).transpose(1, 2)
    return result


def explain_model(image_name, category_name):
    original_image, _, _ = load_and_prepare_images(image_name)
    rgb_img = np.float32(original_image) / 255
    img_tensor = transforms.ToTensor()(rgb_img)
    input_tensor = transforms.Normalize(
        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
    )(img_tensor)
    input_tensor = input_tensor.unsqueeze(0).to(device)
    wrapped_model = SegformerWrapper(model).to(device)
    with torch.no_grad():
        output = wrapped_model(input_tensor)
        upsampled_logits = nn.functional.interpolate(
            output, size=input_tensor.shape[-2:], mode="bilinear", align_corners=False
        )

    normalized_masks = torch.nn.functional.softmax(upsampled_logits, dim=1).cpu()
    category = label2id[category_name]
    mask = normalized_masks[0].argmax(dim=0).numpy()
    mask_float = np.float32(mask == category)
    reshape_transform = partial(
        segformer_reshape_transform_huggingface,  # réorganise les dimensions du tenseur pour qu'elles correspondent au format attendu par GradCAM.
        width=img_tensor.shape[2] // 32,
        height=img_tensor.shape[1] // 32,
    )
    target_layers = [wrapped_model.model.segformer.encoder.layer_norm[-1]]
    mask_float_resized = cv2.resize(mask_float, (output.shape[3], output.shape[2]))
    targets = [SemanticSegmentationTarget(category, mask_float_resized)]
    cam = GradCAM(
        model=wrapped_model,
        target_layers=target_layers,
        reshape_transform=reshape_transform,
    )

    grayscale_cam = cam(input_tensor=input_tensor, targets=targets)
    threshold = 0.01  # Seuil de 1% de sureté
    thresholded_cam = grayscale_cam.copy()
    thresholded_cam[grayscale_cam < threshold] = 0
    if np.max(thresholded_cam) > 0:
        thresholded_cam = thresholded_cam / np.max(thresholded_cam)
    else:
        thresholded_cam = grayscale_cam[0]
    resized_cam = cv2.resize(
        thresholded_cam[0], (input_tensor.shape[3], input_tensor.shape[2])
    )
    masked_cam = resized_cam * mask_float
    if np.max(masked_cam) > 0:
        cam_image = show_cam_on_image(rgb_img, masked_cam, use_rgb=True)
    else:
        cam_image = original_image
    fig, ax = plt.subplots(figsize=(15, 10))
    ax.imshow(cam_image)
    ax.axis("off")
    ax.set_title(f"Masque de chaleur GradCam pour {category_name}", color="white")
    margin = 0.02  # Adjust this value to change the size of the margin
    margin_color = "#0a0f1e"
    fig.subplots_adjust(left=margin, right=1 - margin, top=1 - margin, bottom=margin)
    fig.patch.set_facecolor(margin_color)
    plt.close()

    return fig


# ----Fin Partie Explication GradCam

# ----Partie Data augmentation
import random


def change_image():
    image_dir = (
        "data_sample/images"  # Remplacez par le chemin de votre dossier d'images
    )
    image_list = [f for f in os.listdir(image_dir) if f.endswith(".png")]
    random_image = random.choice(image_list)
    return Image.open(os.path.join(image_dir, random_image))


def apply_augmentation(image, augmentation_names):
    augmentations = {
        "Horizontal Flip": A.HorizontalFlip(p=1),
        "Shift Scale Rotate": A.ShiftScaleRotate(p=1),
        "Random Brightness Contrast": A.RandomBrightnessContrast(p=1),
        "RGB Shift": A.RGBShift(p=1),
        "Blur": A.Blur(blur_limit=(5, 7), p=1),
        "Gaussian Noise": A.GaussNoise(p=1),
        "Grid Distortion": A.GridDistortion(p=1),
        "Random Sun": A.RandomSunFlare(p=1),
    }

    image_array = np.array(image)

    if augmentation_names is not None:
        selected_augs = [
            augmentations[name] for name in augmentation_names if name in augmentations
        ]
        compose = A.Compose(selected_augs)

        # Appliquer la composition d'augmentations
        augmented = compose(image=image_array)
        return Image.fromarray(augmented["image"])
    else:
        return image


# ---- Fin Partie Data augmentation

image_list = [
    f for f in os.listdir(os.path.join(data_folder, "images")) if f.endswith(".png")
]
category_list = list(id2label.values())
image_name = "dusseldorf_000012_000019_leftImg8bit.png"
default_image = os.path.join(data_folder, "images", image_name)

my_theme = gr.Theme.from_hub("YenLai/Superhuman")
with gr.Blocks(title="Preuve de concept", theme=my_theme) as demo:
    gr.Markdown("# Projet 10 - Développer une preuve de concept")
    with gr.Tab("Prédictions"):
        gr.Markdown("## Comparaison de segmentation d'images Cityscapes")
        gr.Markdown(
            "### Sélectionnez une image pour voir la comparaison entre le masque réel, la prédiction FPN et la prédiction SegFormer."
        )

        image_input = gr.Dropdown(choices=image_list, label="Sélectionnez une image")

        gallery_output = gr.Gallery(
            label="Résultats de segmentation",
            show_label=True,
            elem_id="gallery",
            columns=[2],
            rows=[2],
            object_fit="contain",
            height="512px",
            min_width="1024px",
        )

        image_input.change(fn=process_image, inputs=image_input, outputs=gallery_output)

    with gr.Tab("EDA"):
        gr.Markdown("## Analyse Exploratoire des données Cityscapes")
        gr.Markdown(
            "### Visualisations de la distribution de chaque classe selon l'image choisie."
        )
        eda_image_input = gr.Dropdown(
            choices=image_list,
            label="Sélectionnez une image",
        )

        with gr.Row():
            original_image_output = gr.Image(type="pil", label="Image originale")
            original_mask_output = gr.Image(type="pil", label="Masque original")
        class_distribution_plot = gr.Plot(label="Distribution des classes")
        eda_image_input.change(
            fn=show_eda,
            inputs=eda_image_input,
            outputs=[
                original_image_output,
                original_mask_output,
                class_distribution_plot,
            ],
        )

    with gr.Tab("Explication SegFormer"):
        gr.Markdown("## Explication du modèle SegFormer")
        gr.Markdown(
            "### La méthode Grad-CAM est une technique populaire de visualisation qui est utile pour comprendre comment un réseau neuronal convolutif a été conduit à prendre une décision de classification. Elle est spécifique à chaque classe, ce qui signifie qu’elle peut produire une visualisation distincte pour chaque classe présente dans l’image."
        )
        gr.Markdown(
            "### NB: Si l'image s'affiche sans masque, c'est que le modèle ne trouve pas de zones significatives pour une catégorie donnée."
        )

        with gr.Row():
            explain_image_input = gr.Dropdown(
                choices=image_list, label="Sélectionnez une image"
            )
            explain_category_input = gr.Dropdown(
                choices=category_list, label="Sélectionnez une catégorie"
            )

        explain_button = gr.Button("Expliquer")
        explain_output = gr.Plot(label="Explication SegFormer", min_width=200)
        explain_button.click(
            fn=explain_model,
            inputs=[explain_image_input, explain_category_input],
            outputs=explain_output,
        )

    with gr.Tab("Data Augmentation"):
        gr.Markdown("## Visualisation de l'augmentation de données")
        gr.Markdown(
            "### Sélectionnez une ou plusieurs augmentations pour l'appliquer à l'image."
        )
        gr.Markdown("### Vous pouvez également changer d'image.")

        with gr.Row():
            image_display = gr.Image(
                value=default_image,
                label="Image",
                show_download_button=False,
                interactive=False,
            )
            augmented_image = gr.Image(label="Image Augmentée")

        with gr.Row():
            change_image_button = gr.Button("Changer image")
            augmentation_dropdown = gr.Dropdown(
                choices=[
                    "Horizontal Flip",
                    "Shift Scale Rotate",
                    "Random Brightness Contrast",
                    "RGB Shift",
                    "Blur",
                    "Gaussian Noise",
                    "Grid Distortion",
                    "Random Sun",
                ],
                label="Sélectionnez une augmentation",
                multiselect=True,
            )
            apply_button = gr.Button("Appliquer l'augmentation")

        change_image_button.click(fn=change_image, outputs=image_display)

        apply_button.click(
            fn=apply_augmentation,
            inputs=[image_display, augmentation_dropdown],
            outputs=augmented_image,
        )


# Lancer l'application
demo.launch(favicon_path="static/favicon.ico", share=True)