!pip install -U -q datasets transformers[torch] timm wandb torchmetrics
# Tested with datasets==2.21.0, transformers==4.44.2 timm==1.0.9, wandb==0.17.9 torchmetrics==1.4.1

Fashionpedia is a dataset which consists of two parts: (1) an ontology built by fashion experts containing 27 main apparel categories, 19 apparel parts, 294 fine-grained attributes and their relationships; (2) a dataset with 48k everyday and celebrity event fashion images annotated with segmentation masks and their associated per-mask fine-grained attributes, built upon the Fashionpedia ontology.

from datasets import load_dataset

dataset = load_dataset("detection-datasets/fashionpedia")

dataset

dataset["train"][0]

train_dataset = dataset["train"]
test_dataset = dataset["val"]

"""
def create_sample(dataset, sample_fraction=0.01, seed=42):
    sample_size = int(sample_fraction * len(dataset))
    sampled_dataset = dataset.shuffle(seed=seed).select(range(sample_size))
    print(f"Original size: {len(dataset)}")
    print(f"Sample size: {len(sampled_dataset)}")
    return sampled_dataset

# Apply function to both splits
train_dataset = create_sample(train_dataset)
test_dataset = create_sample(test_dataset)
"""

import numpy as np
from PIL import Image, ImageDraw


id2label = {
    0: "shirt, blouse",
    1: "top, t-shirt, sweatshirt",
    2: "sweater",
    3: "cardigan",
    4: "jacket",
    5: "vest",
    6: "pants",
    7: "shorts",
    8: "skirt",
    9: "coat",
    10: "dress",
    11: "jumpsuit",
    12: "cape",
    13: "glasses",
    14: "hat",
    15: "headband, head covering, hair accessory",
    16: "tie",
    17: "glove",
    18: "watch",
    19: "belt",
    20: "leg warmer",
    21: "tights, stockings",
    22: "sock",
    23: "shoe",
    24: "bag, wallet",
    25: "scarf",
    26: "umbrella",
    27: "hood",
    28: "collar",
    29: "lapel",
    30: "epaulette",
    31: "sleeve",
    32: "pocket",
    33: "neckline",
    34: "buckle",
    35: "zipper",
    36: "applique",
    37: "bead",
    38: "bow",
    39: "flower",
    40: "fringe",
    41: "ribbon",
    42: "rivet",
    43: "ruffle",
    44: "sequin",
    45: "tassel",
}


label2id = {v: k for k, v in id2label.items()}

>>> def draw_image_from_idx(dataset, idx):
...     sample = dataset[idx]
...     image = sample["image"]
...     annotations = sample["objects"]
...     draw = ImageDraw.Draw(image)
...     width, height = sample["width"], sample["height"]

...     print(annotations)

...     for i in range(len(annotations["bbox_id"])):
...         box = annotations["bbox"][i]
...         x1, y1, x2, y2 = tuple(box)
...         draw.rectangle((x1, y1, x2, y2), outline="red", width=3)
...         draw.text((x1, y1), id2label[annotations["category"][i]], fill="green")

...     return image


>>> draw_image_from_idx(dataset=train_dataset, idx=10)  # You can test changing this id

>>> import matplotlib.pyplot as plt


>>> def plot_images(dataset, indices):
...     """
...     Plot images and their annotations.
...     """
...     num_cols = 3
...     num_rows = int(np.ceil(len(indices) / num_cols))
...     fig, axes = plt.subplots(num_rows, num_cols, figsize=(15, 10))

...     for i, idx in enumerate(indices):
...         row = i // num_cols
...         col = i % num_cols

...         image = draw_image_from_idx(dataset, idx)

...         axes[row, col].imshow(image)
...         axes[row, col].axis("off")

...     for j in range(i + 1, num_rows * num_cols):
...         fig.delaxes(axes.flatten()[j])

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


>>> plot_images(train_dataset, range(9))

>>> from datasets import Dataset


>>> def filter_invalid_bboxes(example):
...     valid_bboxes = []
...     valid_bbox_ids = []
...     valid_categories = []
...     valid_areas = []

...     for i, bbox in enumerate(example["objects"]["bbox"]):
...         x_min, y_min, x_max, y_max = bbox[:4]
...         if x_min < x_max and y_min < y_max:
...             valid_bboxes.append(bbox)
...             valid_bbox_ids.append(example["objects"]["bbox_id"][i])
...             valid_categories.append(example["objects"]["category"][i])
...             valid_areas.append(example["objects"]["area"][i])
...         else:
...             print(
...                 f"Image with invalid bbox: {example['image_id']} Invalid bbox detected and discarded: {bbox} - bbox_id: {example['objects']['bbox_id'][i]} - category: {example['objects']['category'][i]}"
...             )

...     example["objects"]["bbox"] = valid_bboxes
...     example["objects"]["bbox_id"] = valid_bbox_ids
...     example["objects"]["category"] = valid_categories
...     example["objects"]["area"] = valid_areas

...     return example


>>> train_dataset = train_dataset.map(filter_invalid_bboxes)
>>> test_dataset = test_dataset.map(filter_invalid_bboxes)

>>> print(train_dataset)
>>> print(test_dataset)

id_list = []
category_examples = {}
for example in train_dataset:
    id_list += example["objects"]["bbox_id"]
    for category in example["objects"]["category"]:
        if id2label[category] not in category_examples:
            category_examples[id2label[category]] = 1
        else:
            category_examples[id2label[category]] += 1

id_list.sort()

>>> import matplotlib.pyplot as plt

>>> categories = list(category_examples.keys())
>>> values = list(category_examples.values())

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

>>> bars = ax.bar(categories, values, color="skyblue")

>>> ax.set_xlabel("Categories", fontsize=14)
>>> ax.set_ylabel("Number of Occurrences", fontsize=14)
>>> ax.set_title("Number of Occurrences by Category", fontsize=16)

>>> ax.set_xticklabels(categories, rotation=90, ha="right")
>>> ax.grid(axis="y", linestyle="--", alpha=0.7)

>>> for bar in bars:
...     height = bar.get_height()
...     ax.text(bar.get_x() + bar.get_width() / 2.0, height, f"{height}", ha="center", va="bottom", fontsize=10)

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

import albumentations as A

train_transform = A.Compose(
    [
        A.LongestMaxSize(500),
        A.PadIfNeeded(500, 500, border_mode=0, value=(0, 0, 0)),
        A.HorizontalFlip(p=0.5),
        A.RandomBrightnessContrast(p=0.5),
        A.HueSaturationValue(p=0.5),
        A.Rotate(limit=10, p=0.5),
        A.RandomScale(scale_limit=0.2, p=0.5),
        A.GaussianBlur(p=0.5),
        A.GaussNoise(p=0.5),
    ],
    bbox_params=A.BboxParams(format="pascal_voc", label_fields=["category"]),
)

val_transform = A.Compose(
    [
        A.LongestMaxSize(500),
        A.PadIfNeeded(500, 500, border_mode=0, value=(0, 0, 0)),
    ],
    bbox_params=A.BboxParams(format="pascal_voc", label_fields=["category"]),
)

from transformers import AutoImageProcessor

checkpoint = "facebook/detr-resnet-50-dc5"
image_processor = AutoImageProcessor.from_pretrained(checkpoint)

def formatted_anns(image_id, category, area, bbox):
    annotations = []
    for i in range(0, len(category)):
        new_ann = {
            "image_id": image_id,
            "category_id": category[i],
            "isCrowd": 0,
            "area": area[i],
            "bbox": list(bbox[i]),
        }
        annotations.append(new_ann)

    return annotations


def convert_voc_to_coco(bbox):
    xmin, ymin, xmax, ymax = bbox
    width = xmax - xmin
    height = ymax - ymin
    return [xmin, ymin, width, height]


def transform_aug_ann(examples, transform):
    image_ids = examples["image_id"]
    images, bboxes, area, categories = [], [], [], []
    for image, objects in zip(examples["image"], examples["objects"]):
        image = np.array(image.convert("RGB"))[:, :, ::-1]
        out = transform(image=image, bboxes=objects["bbox"], category=objects["category"])

        area.append(objects["area"])
        images.append(out["image"])

        # Convert to COCO format
        converted_bboxes = [convert_voc_to_coco(bbox) for bbox in out["bboxes"]]
        bboxes.append(converted_bboxes)

        categories.append(out["category"])

    targets = [
        {"image_id": id_, "annotations": formatted_anns(id_, cat_, ar_, box_)}
        for id_, cat_, ar_, box_ in zip(image_ids, categories, area, bboxes)
    ]

    return image_processor(images=images, annotations=targets, return_tensors="pt")


def transform_train(examples):
    return transform_aug_ann(examples, transform=train_transform)


def transform_val(examples):
    return transform_aug_ann(examples, transform=val_transform)


train_dataset_transformed = train_dataset.with_transform(transform_train)
test_dataset_transformed = test_dataset.with_transform(transform_val)

>>> # Updated draw function to accept an optional transform
>>> def draw_augmented_image_from_idx(dataset, idx, transform=None):
...     sample = dataset[idx]
...     image = sample["image"]
...     annotations = sample["objects"]

...     # Convert image to RGB and NumPy array
...     image = np.array(image.convert("RGB"))[:, :, ::-1]

...     if transform:
...         augmented = transform(image=image, bboxes=annotations["bbox"], category=annotations["category"])
...         image = augmented["image"]
...         annotations["bbox"] = augmented["bboxes"]
...         annotations["category"] = augmented["category"]

...     image = Image.fromarray(image[:, :, ::-1])  # Convert back to PIL Image
...     draw = ImageDraw.Draw(image)
...     width, height = sample["width"], sample["height"]

...     for i in range(len(annotations["bbox_id"])):
...         box = annotations["bbox"][i]
...         x1, y1, x2, y2 = tuple(box)

...         # Normalize coordinates if necessary
...         if max(box) <= 1.0:
...             x1, y1 = int(x1 * width), int(y1 * height)
...             x2, y2 = int(x2 * width), int(y2 * height)
...         else:
...             x1, y1 = int(x1), int(y1)
...             x2, y2 = int(x2), int(y2)

...         draw.rectangle((x1, y1, x2, y2), outline="red", width=3)
...         draw.text((x1, y1), id2label[annotations["category"][i]], fill="green")

...     return image


>>> # Updated plot function to include augmentation
>>> def plot_augmented_images(dataset, indices, transform=None):
...     """
...     Plot images and their annotations with optional augmentation.
...     """
...     num_rows = len(indices) // 3
...     num_cols = 3
...     fig, axes = plt.subplots(num_rows, num_cols, figsize=(15, 10))

...     for i, idx in enumerate(indices):
...         row = i // num_cols
...         col = i % num_cols

...         # Draw augmented image
...         image = draw_augmented_image_from_idx(dataset, idx, transform=transform)

...         # Display image on the corresponding subplot
...         axes[row, col].imshow(image)
...         axes[row, col].axis("off")

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


>>> # Now use the function to plot augmented images
>>> plot_augmented_images(train_dataset, range(9), transform=train_transform)

from transformers import AutoModelForObjectDetection

model = AutoModelForObjectDetection.from_pretrained(
    checkpoint,
    id2label=id2label,
    label2id=label2id,
    ignore_mismatched_sizes=True,
)

output_dir = "detr-resnet-50-dc5-fashionpedia-finetuned"  # change this

from huggingface_hub import notebook_login

notebook_login()

from transformers import TrainingArguments
from transformers import Trainer

import torch

# Define the training arguments

training_args = TrainingArguments(
    output_dir=output_dir,
    per_device_train_batch_size=4,
    per_device_eval_batch_size=4,
    max_steps=10000,
    fp16=True,
    save_steps=10,
    logging_steps=1,
    learning_rate=1e-5,
    weight_decay=1e-4,
    save_total_limit=2,
    remove_unused_columns=False,
    evaluation_strategy="steps",
    eval_steps=50,
    eval_strategy="steps",
    report_to="wandb",
    push_to_hub=True,
    batch_eval_metrics=True,
)

import wandb

wandb.init(
    project="detr-resnet-50-dc5-fashionpedia-finetuned",  # change this
    name="detr-resnet-50-dc5-fashionpedia-finetuned",  # change this
    config=training_args,
)

from torchmetrics.detection.mean_ap import MeanAveragePrecision
from torch.nn.functional import softmax


def denormalize_boxes(boxes, width, height):
    boxes = boxes.clone()
    boxes[:, 0] *= width  # xmin
    boxes[:, 1] *= height  # ymin
    boxes[:, 2] *= width  # xmax
    boxes[:, 3] *= height  # ymax
    return boxes


batch_metrics = []


def compute_metrics(eval_pred, compute_result):
    global batch_metrics

    (loss_dict, scores, pred_boxes, last_hidden_state, encoder_last_hidden_state), labels = eval_pred

    image_sizes = []
    target = []
    for label in labels:

        image_sizes.append(label["orig_size"])
        width, height = label["orig_size"]
        denormalized_boxes = denormalize_boxes(label["boxes"], width, height)
        target.append(
            {
                "boxes": denormalized_boxes,
                "labels": label["class_labels"],
            }
        )
    predictions = []
    for score, box, target_sizes in zip(scores, pred_boxes, image_sizes):
        # Extract the bounding boxes, labels, and scores from the model's output
        pred_scores = score[:, :-1]  # Exclude the no-object class
        pred_scores = softmax(pred_scores, dim=-1)
        width, height = target_sizes
        pred_boxes = denormalize_boxes(box, width, height)
        pred_labels = torch.argmax(pred_scores, dim=-1)

        # Get the scores corresponding to the predicted labels
        pred_scores_for_labels = torch.gather(pred_scores, 1, pred_labels.unsqueeze(-1)).squeeze(-1)
        predictions.append(
            {
                "boxes": pred_boxes,
                "scores": pred_scores_for_labels,
                "labels": pred_labels,
            }
        )

    metric = MeanAveragePrecision(box_format="xywh", class_metrics=True)

    if not compute_result:
        # Accumulate batch-level metrics
        batch_metrics.append({"preds": predictions, "target": target})
        return {}
    else:
        # Compute final aggregated metrics
        # Aggregate batch-level metrics (this should be done based on your metric library's needs)
        all_preds = []
        all_targets = []
        for batch in batch_metrics:
            all_preds.extend(batch["preds"])
            all_targets.extend(batch["target"])

        # Update metric with all accumulated predictions and targets
        metric.update(preds=all_preds, target=all_targets)
        metrics = metric.compute()

        # Convert and format metrics as needed
        classes = metrics.pop("classes")
        map_per_class = metrics.pop("map_per_class")
        mar_100_per_class = metrics.pop("mar_100_per_class")

        for class_id, class_map, class_mar in zip(classes, map_per_class, mar_100_per_class):
            class_name = id2label[class_id.item()] if id2label is not None else class_id.item()
            metrics[f"map_{class_name}"] = class_map
            metrics[f"mar_100_{class_name}"] = class_mar

        # Round metrics for cleaner output
        metrics = {k: round(v.item(), 4) for k, v in metrics.items()}

        # Clear batch metrics for next evaluation
        batch_metrics = []

        return metrics

def collate_fn(batch):
    pixel_values = [item["pixel_values"] for item in batch]
    encoding = image_processor.pad(pixel_values, return_tensors="pt")
    labels = [item["labels"] for item in batch]

    batch = {}
    batch["pixel_values"] = encoding["pixel_values"]
    batch["pixel_mask"] = encoding["pixel_mask"]
    batch["labels"] = labels

    return batch

trainer = Trainer(
    model=model,
    args=training_args,
    data_collator=collate_fn,
    train_dataset=train_dataset_transformed,
    eval_dataset=test_dataset_transformed,
    tokenizer=image_processor,
    compute_metrics=compute_metrics,
)

trainer.train()

trainer.push_to_hub()

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

url = "https://images.unsplash.com/photo-1536243298747-ea8874136d64?q=80&w=640"

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

obj_detector = pipeline(
    "object-detection", model="sergiopaniego/detr-resnet-50-dc5-fashionpedia-finetuned"  # Change with your model name
)


results = obj_detector(image)
print(results)

from PIL import Image, ImageDraw
import numpy as np


def plot_results(image, results, threshold=0.6):
    image = Image.fromarray(np.uint8(image))
    draw = ImageDraw.Draw(image)
    width, height = image.size

    for result in results:
        score = result["score"]
        label = result["label"]
        box = list(result["box"].values())

        if score > threshold:
            x1, y1, x2, y2 = tuple(box)
            draw.rectangle((x1, y1, x2, y2), outline="red", width=3)
            draw.text((x1 + 5, y1 - 10), label, fill="white")
            draw.text((x1 + 5, y1 + 10), f"{score:.2f}", fill="green" if score > 0.7 else "red")

    return image

>>> plot_results(image, results)

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

from IPython.display import IFrame

IFrame(src="https://sergiopaniego-detr-object-detection-fashionpedia-fa0081f.hf.space", width=1000, height=800)

# app.py

import gradio as gr
import spaces
import torch

from PIL import Image
from transformers import pipeline
import matplotlib.pyplot as plt
import io

model_pipeline = pipeline("object-detection", model="sergiopaniego/detr-resnet-50-dc5-fashionpedia-finetuned")


COLORS = [
    [0.000, 0.447, 0.741],
    [0.850, 0.325, 0.098],
    [0.929, 0.694, 0.125],
    [0.494, 0.184, 0.556],
    [0.466, 0.674, 0.188],
    [0.301, 0.745, 0.933],
]


def get_output_figure(pil_img, results, threshold):
    plt.figure(figsize=(16, 10))
    plt.imshow(pil_img)
    ax = plt.gca()
    colors = COLORS * 100

    for result in results:
        score = result["score"]
        label = result["label"]
        box = list(result["box"].values())
        if score > threshold:
            c = COLORS[hash(label) % len(COLORS)]
            ax.add_patch(
                plt.Rectangle((box[0], box[1]), box[2] - box[0], box[3] - box[1], fill=False, color=c, linewidth=3)
            )
            text = f"{label}: {score:0.2f}"
            ax.text(box[0], box[1], text, fontsize=15, bbox=dict(facecolor="yellow", alpha=0.5))
    plt.axis("off")

    return plt.gcf()


@spaces.GPU
def detect(image):
    results = model_pipeline(image)
    print(results)

    output_figure = get_output_figure(image, results, threshold=0.7)

    buf = io.BytesIO()
    output_figure.savefig(buf, bbox_inches="tight")
    buf.seek(0)
    output_pil_img = Image.open(buf)

    return output_pil_img


with gr.Blocks() as demo:
    gr.Markdown("# Object detection with DETR fine tuned on detection-datasets/fashionpedia")
    gr.Markdown(
        """
        This application uses a fine tuned DETR (DEtection TRansformers) to detect objects on images.
        This version was trained using detection-datasets/fashionpedia dataset.
        You can load an image and see the predictions for the objects detected.
        """
    )

    gr.Interface(
        fn=detect,
        inputs=gr.Image(label="Input image", type="pil"),
        outputs=[gr.Image(label="Output prediction", type="pil")],
    )

demo.launch(show_error=True)

# requirements.txt

transformers
timm
torch

!pip install gradio_client

from gradio_client import Client, handle_file

client = Client("sergiopaniego/DETR_object_detection_fashionpedia-finetuned")  # change this with your Space
result = client.predict(
    image=handle_file("https://images.unsplash.com/photo-1536243298747-ea8874136d64?q=80&w=640"), api_name="/predict"
)

from PIL import Image

img = Image.open(result).convert("RGB")

>>> from IPython.display import display

>>> display(img)