!pip install -q datasets transformers evaluate wandb
# datasets==3.0.0, transformers==4.44.2, evaluate==0.4.3, wandb==0.18.1 ile test edildi

from huggingface_hub import notebook_login

notebook_login()

sidewalk_dataset_identifier = "segments/sidewalk-semantic"

from datasets import load_dataset

dataset = load_dataset(sidewalk_dataset_identifier)

dataset

dataset = dataset.shuffle(seed=42)
dataset = dataset["train"].train_test_split(test_size=0.2)
train_ds = dataset["train"]
test_ds = dataset["test"]

image = train_ds[0]
image

>>> import json
>>> from huggingface_hub import hf_hub_download

>>> filename = "id2label.json"
>>> id2label = json.load(
...     open(hf_hub_download(repo_id=sidewalk_dataset_identifier, filename=filename, repo_type="dataset"), "r")
... )
>>> id2label = {int(k): v for k, v in id2label.items()}
>>> label2id = {v: k for k, v in id2label.items()}

>>> num_labels = len(id2label)
>>> print("Id2label:", id2label)

sidewalk_palette = [
    [0, 0, 0],  # unlabeled / etiketlenmemiş
    [216, 82, 24],  # flat-road / düz-yol
    [255, 255, 0],  # flat-sidewalk / düz-kaldırım
    [125, 46, 141],  # flat-crosswalk / düz-yaya geçidi
    [118, 171, 47],  # flat-cyclinglane / düz-bisiklet yolu
    [161, 19, 46],  # flat-parkingdriveway / düz-park yeri/garaj yolu
    [255, 0, 0],  # flat-railtrack / düz-ray hattı
    [0, 128, 128],  # flat-curb / düz-kaldırım kenarı
    [190, 190, 0],  # human-person / insan-kişi
    [0, 255, 0],  # human-rider / insan-sürücü
    [0, 0, 255],  # vehicle-car / araç-araba
    [170, 0, 255],  # vehicle-truck / araç-kamyon
    [84, 84, 0],  # vehicle-bus / araç-otobüs
    [84, 170, 0],  # vehicle-tramtrain / araç-tramvay/tren
    [84, 255, 0],  # vehicle-motorcycle / araç-motosiklet
    [170, 84, 0],  # vehicle-bicycle / araç-bisiklet
    [170, 170, 0],  # vehicle-caravan / araç-karavan
    [170, 255, 0],  # vehicle-cartrailer / araç-römork
    [255, 84, 0],  # construction-building / inşaat-bina
    [255, 170, 0],  # construction-door / inşaat-kapı
    [255, 255, 0],  # construction-wall / inşaat-duvar
    [33, 138, 200],  # construction-fenceguardrail / inşaat-çit/korkuluk
    [0, 170, 127],  # construction-bridge / inşaat-köprü
    [0, 255, 127],  # construction-tunnel / inşaat-tünel
    [84, 0, 127],  # construction-stairs / inşaat-merdiven
    [84, 84, 127],  # object-pole / nesne-direk
    [84, 170, 127],  # object-trafficsign / nesne-trafik levhası
    [84, 255, 127],  # object-trafficlight / nesne-trafik ışığı
    [170, 0, 127],  # nature-vegetation / doğa-bitki örtüsü
    [170, 84, 127],  # nature-terrain / doğa-arazi
    [170, 170, 127],  # sky / gökyüzü
    [170, 255, 127],  # void-ground / boşluk-zemin
    [255, 0, 127],  # void-dynamic / boşluk-dinamik
    [255, 84, 127],  # void-static / boşluk-statik
    [255, 170, 127],  # void-unclear / boşluk-belirsiz
]

>>> from matplotlib import pyplot as plt
>>> import numpy as np
>>> from PIL import Image
>>> import matplotlib.patches as patches

>>> # Legend'i(Açıklama/lejant) ayrı olarak oluştrup ve gösterelim
>>> fig, ax = plt.subplots(figsize=(18, 2))

>>> legend_patches = [
...     patches.Patch(color=np.array(color) / 255, label=label)
...     for label, color in zip(id2label.values(), sidewalk_palette)
... ]

>>> ax.legend(handles=legend_patches, loc="center", bbox_to_anchor=(0.5, 0.5), ncol=5, fontsize=8)
>>> ax.axis("off")

>>> plt.show()

>>> for i in range(5):
...     image = train_ds[i]

...     fig, ax = plt.subplots(1, 3, figsize=(18, 6))

...     # Orjinal resmi göster
...     ax[0].imshow(image["pixel_values"])
...     ax[0].set_title("Orjinal Resim")
...     ax[0].axis("off")

...     mask_np = np.array(image["label"])

...     # Boş bir RGB resim oluştur
...     colored_mask = np.zeros((mask_np.shape[0], mask_np.shape[1], 3), dtype=np.uint8)

...     # Her etikete denk gelen maske için bir renk atayalım
...     for label_id, color in enumerate(sidewalk_palette):
...         colored_mask[mask_np == label_id] = color

...     colored_mask_img = Image.fromarray(colored_mask, "RGB")

...     # Segmentasyon maskesini göster
...     ax[1].imshow(colored_mask_img)
...     ax[1].set_title("Segmentasyon Maskesi")
...     ax[1].axis("off")

...     # Şeffaflığı desteklemek için orijinal görüntüyü RGBA formatına dönüştürün
...     image_rgba = image["pixel_values"].convert("RGBA")
...     colored_mask_rgba = colored_mask_img.convert("RGBA")

...     # Maskenin şeffaflığını ayarlayın
...     alpha = 128  # Şeffaflık seviyesi (0 tamamen şeffaf, 255 tamamen opak)
...     image_2_with_alpha = Image.new("RGBA", colored_mask_rgba.size)
...     for x in range(colored_mask_rgba.width):
...         for y in range(colored_mask_rgba.height):
...             r, g, b, a = colored_mask_rgba.getpixel((x, y))
...             image_2_with_alpha.putpixel((x, y), (r, g, b, alpha))

...     superposed = Image.alpha_composite(image_rgba, image_2_with_alpha)

...     # Show the mask overlay
...     ax[2].imshow(superposed)
...     ax[2].set_title("Maske Katmanı")
...     ax[2].axis("off")

...     plt.show()

import matplotlib.pyplot as plt
import numpy as np

class_counts = np.zeros(len(id2label))

for example in train_ds:
    mask_np = np.array(example["label"])
    unique, counts = np.unique(mask_np, return_counts=True)
    for u, c in zip(unique, counts):
        class_counts[u] += c

>>> from matplotlib import pyplot as plt
>>> import numpy as np
>>> from matplotlib import patches

>>> labels = list(id2label.values())

>>> # Renkleri [0, 1] aralığına normalleştir
>>> normalized_palette = [tuple(c / 255 for c in color) for color in sidewalk_palette]

>>> # Görselleştirme
>>> fig, ax = plt.subplots(figsize=(12, 8))

>>> bars = ax.bar(range(len(labels)), class_counts, color=[normalized_palette[i] for i in range(len(labels))])

>>> ax.set_xticks(range(len(labels)))
>>> ax.set_xticklabels(labels, rotation=90, ha="right")

>>> ax.set_xlabel("Kategoriler", fontsize=14)
>>> ax.set_ylabel("Görünüm Sayısı", fontsize=14)
>>> ax.set_title("Kategoriye Göre Görünüm Sayısı", fontsize=16)

>>> ax.grid(axis="y", linestyle="--", alpha=0.7)

>>> # Y ekseni limitini ayarlayın
>>> y_max = max(class_counts)
>>> ax.set_ylim(0, y_max * 1.25)

>>> for bar in bars:
...     height = int(bar.get_height())
...     offset = 10  # Metin konumunu ayarlayın
...     ax.text(
...         bar.get_x() + bar.get_width() / 2.0,
...         height + offset,
...         f"{height}",
...         ha="center",
...         va="bottom",
...         rotation=90,
...         fontsize=10,
...         color="black",
...     )

>>> fig.legend(
...     handles=legend_patches, loc="center left", bbox_to_anchor=(1, 0.5), ncol=1, fontsize=8
... )  # ncol'u gereken durumlarda ayarlayın

>>> plt.tight_layout()
>>> plt.show()

import albumentations as A
from transformers import SegformerImageProcessor

image_processor = SegformerImageProcessor()

albumentations_transform = A.Compose(
    [
        A.HorizontalFlip(p=0.5),
        A.ShiftScaleRotate(shift_limit=0.1, scale_limit=0.1, rotate_limit=30, p=0.7),
        A.RandomResizedCrop(height=512, width=512, scale=(0.8, 1.0), ratio=(0.75, 1.33), p=0.5),
        A.RandomBrightnessContrast(brightness_limit=0.25, contrast_limit=0.25, p=0.5),
        A.HueSaturationValue(hue_shift_limit=10, sat_shift_limit=25, val_shift_limit=20, p=0.5),
        A.GaussianBlur(blur_limit=(3, 5), p=0.3),
        A.GaussNoise(var_limit=(10, 50), p=0.4),
    ]
)


def train_transforms(example_batch):
    augmented_images = [albumentations_transform(image=np.array(x))["image"] for x in example_batch["pixel_values"]]
    labels = [x for x in example_batch["label"]]
    inputs = image_processor(augmented_images, labels)
    return inputs


def val_transforms(example_batch):
    images = [x for x in example_batch["pixel_values"]]
    labels = [x for x in example_batch["label"]]
    inputs = image_processor(images, labels)
    return inputs


# Transformları ayarlayın
train_ds.set_transform(train_transforms)
test_ds.set_transform(val_transforms)

from transformers import SegformerForSemanticSegmentation

pretrained_model_name = "nvidia/mit-b0"
model = SegformerForSemanticSegmentation.from_pretrained(pretrained_model_name, id2label=id2label, label2id=label2id)

from transformers import TrainingArguments

output_dir = "segformer-b0-segments-sidewalk-finetuned"

training_args = TrainingArguments(
    output_dir=output_dir,
    learning_rate=6e-5,
    num_train_epochs=20,
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    save_total_limit=2,
    evaluation_strategy="steps",
    save_strategy="steps",
    save_steps=20,
    eval_steps=20,
    logging_steps=1,
    eval_accumulation_steps=5,
    load_best_model_at_end=True,
    push_to_hub=True,
    report_to="wandb",
)

import wandb

wandb.init(
    project="segformer-b0-segments-sidewalk-finetuned",  # bunu değiştirin
    name="segformer-b0-segments-sidewalk-finetuned",  # bunu değiştirin
    config=training_args,
)

RuntimeWarning: invalid value encountered in divide iou = total_area_intersect / total_area_union

import evaluate

evaluate.logging.set_verbosity_error()

import torch
from torch import nn
import multiprocessing

metric = evaluate.load("mean_iou")


def compute_metrics(eval_pred):
    with torch.no_grad():
        logits, labels = eval_pred
        logits_tensor = torch.from_numpy(logits)
        # Etiket boyutuna göre logitleri ölçeklendirin
        logits_tensor = nn.functional.interpolate(
            logits_tensor,
            size=labels.shape[-2:],
            mode="bilinear",
            align_corners=False,
        ).argmax(dim=1)

        # şimdilik compute yerine _compute kullanılıyor: https://github.com/huggingface/evaluate/pull/328#issuecomment-1286866576
        pred_labels = logits_tensor.detach().cpu().numpy()
        import warnings

        with warnings.catch_warnings():
            warnings.simplefilter("ignore", RuntimeWarning)
            metrics = metric._compute(
                predictions=pred_labels,
                references=labels,
                num_labels=len(id2label),
                ignore_index=0,
                reduce_labels=image_processor.do_reduce_labels,
            )

        # kategorilere özel metrikleri ayrı key-value çiftleri olarak ekleyin
        per_category_accuracy = metrics.pop("per_category_accuracy").tolist()
        per_category_iou = metrics.pop("per_category_iou").tolist()

        metrics.update({f"accuracy_{id2label[i]}": v for i, v in enumerate(per_category_accuracy)})
        metrics.update({f"iou_{id2label[i]}": v for i, v in enumerate(per_category_iou)})

        return metrics

from transformers import Trainer

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_ds,
    eval_dataset=test_ds,
    tokenizer=image_processor,
    compute_metrics=compute_metrics,
)

trainer.train()

import requests
from transformers import pipeline
import numpy as np
from PIL import Image, ImageDraw

url = "https://images.unsplash.com/photo-1594098742644-314fedf61fb6?q=80&w=2672&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D"

image = Image.open(requests.get(url, stream=True).raw)

image_segmentator = pipeline(
    "image-segmentation",
    model="sergiopaniego/segformer-b0-segments-sidewalk-finetuned",  # Bu model adını kendi eğittiniz model ile değiştirin
)

results = image_segmentator(image)

>>> plt.imshow(image)
>>> plt.axis("off")
>>> plt.show()

>>> image_array = np.array(image)

>>> segmentation_map = np.zeros_like(image_array)

>>> for result in results:
...     mask = np.array(result["mask"])
...     label = result["label"]

...     label_index = list(id2label.values()).index(label)

...     color = sidewalk_palette[label_index]

...     for c in range(3):
...         segmentation_map[:, :, c] = np.where(mask, color[c], segmentation_map[:, :, c])

>>> plt.figure(figsize=(10, 10))
>>> plt.imshow(image_array)
>>> plt.imshow(segmentation_map, alpha=0.5)
>>> plt.axis("off")
>>> plt.show()

metrics = trainer.evaluate(test_ds)
print(metrics)

>>> url = "https://images.unsplash.com/photo-1594098742644-314fedf61fb6?q=80&w=2672&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D"
>>> image = Image.open(requests.get(url, stream=True).raw)

>>> plt.imshow(image)
>>> plt.axis("off")
>>> plt.show()

from huggingface_hub import InferenceClient

client = InferenceClient()

response = client.image_segmentation(
    model="sergiopaniego/segformer-b0-segments-sidewalk-finetuned",  # Kendi modeliniz ile değiştirin
    image="https://images.unsplash.com/photo-1594098742644-314fedf61fb6?q=80&w=2672&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D",
)

print(response)

>>> image_array = np.array(image)
>>> segmentation_map = np.zeros_like(image_array)

>>> for result in response:
...     mask = np.array(result["mask"])
...     label = result["label"]

...     label_index = list(id2label.values()).index(label)

...     color = sidewalk_palette[label_index]

...     for c in range(3):
...         segmentation_map[:, :, c] = np.where(mask, color[c], segmentation_map[:, :, c])

>>> plt.figure(figsize=(10, 10))
>>> plt.imshow(image_array)
>>> plt.imshow(segmentation_map, alpha=0.5)
>>> plt.axis("off")
>>> plt.show()

import { HfInference } from "@huggingface/inference";

const inference = new HfInference(HF_TOKEN);
await inference.imageSegmentation({
    data: await (await fetch("https://picsum.photos/300/300")).blob(),
    model: "sergiopaniego/segformer-b0-segments-sidewalk-finetuned",
});

from IPython.display import IFrame

IFrame(src="https://sergiopaniego-segformer-b0-segments-sidewalk-finetuned.hf.space", width=1000, height=800)