app.py

import json
import random
import spaces

import gradio as gr
import matplotlib.pyplot as plt
import numpy as np
import onnxruntime
import torch
import torchvision.transforms.functional as F
from huggingface_hub import hf_hub_download
from PIL import Image, ImageColor
from torchvision.io import read_image
from torchvision.models.detection import MaskRCNN_ResNet50_FPN_Weights
from torchvision.utils import draw_bounding_boxes, draw_segmentation_masks

# Load pre-trained model transformations.
weights = MaskRCNN_ResNet50_FPN_Weights.DEFAULT
transforms = weights.transforms()


def fix_category_id(cat_ids: list):
    # Define the excluded category ids and the remaining ones
    excluded_indices = {2, 12, 16, 19, 20}
    remaining_categories = list(set(range(27)) - excluded_indices)

    # Create a dictionary that maps new IDs to old(original) IDs
    new_id_to_org_id = dict(zip(range(len(remaining_categories)), remaining_categories))

    return [new_id_to_org_id[i-1]+1 for i in cat_ids]


def process_categories() -> tuple:
    """
    Load and process category information from a JSON file.
    Returns a tuple containing two dictionaries: `category_id_to_name` maps category IDs to their names, and
    `category_id_to_color` maps category IDs to a randomly sampled RGB color.

    Returns:
        tuple: A tuple containing two dictionaries:
            - `category_id_to_name`: a dictionary mapping category IDs to their names.
            - `category_id_to_color`: a dictionary mapping category IDs to a randomly sampled RGB color.
    """
    # Load raw categories from JSON file
    with open("categories.json") as fp:
        categories = json.load(fp)

    # Map category IDs to names
    category_id_to_name = {d["id"]: d["name"] for d in categories}

    # Set the seed for the random sampling operation
    random.seed(42)

    # Get a list of all the color names in the PIL colormap
    color_names = list(ImageColor.colormap.keys())

    # Sample 46 unique colors from the list of color names
    sampled_colors = random.sample(color_names, 46)

    # Convert the color names to RGB values
    rgb_colors = [ImageColor.getrgb(color_name) for color_name in sampled_colors]

    # Map category IDs to colors
    category_id_to_color = {
        category["id"]: color for category, color in zip(categories, rgb_colors)
    }

    return category_id_to_name, category_id_to_color


def draw_predictions(
        boxes, labels, scores, masks, img, model_name, score_threshold, proba_threshold
):
    """
    Draw predictions on the input image based on the provided boxes, labels, scores, and masks. Only predictions
    with scores above the `score_threshold` will be included, and masks with probabilities exceeding the
    `proba_threshold` will be displayed.

    Args:
    - boxes: numpy.ndarray - an array of bounding box coordinates.
    - labels: numpy.ndarray - an array of integers representing the predicted class for each bounding box.
    - scores: numpy.ndarray - an array of confidence scores for each bounding box.
    - masks: numpy.ndarray - an array of binary masks for each bounding box.
    - img: PIL.Image.Image - the input image.
    - model_name: str - name of the model given by the dropdown menu, either "facere" or "facere+".
    - score_threshold: float - a confidence score threshold for filtering out low-scoring bbox predictions.
    - proba_threshold: float - a threshold for filtering out low-probability (pixel-wise) mask predictions.

    Returns:
    - A list of strings, each representing the path to an image file containing the input image with a different
      set of predictions drawn (masks, bounding boxes, masks with bounding box labels and scores).
    """
    imgs_list = []

    # Map label IDs to names and colors
    label_id_to_name, label_id_to_color = process_categories()

    # Filter out predictions using thresholds
    labels_id = labels[scores > score_threshold].tolist()
    if model_name == "facere+":
        labels_id = fix_category_id(labels_id)
    # models output is in range: [1,class_id+1], hence re-map to: [0,class_id]
    labels = [label_id_to_name[int(i) - 1] for i in labels_id]
    masks = (masks[scores > score_threshold] > proba_threshold).astype(np.uint8)
    boxes = boxes[scores > score_threshold]

    # Draw masks to input image and save
    img_masks = draw_segmentation_masks(
        image=img,
        masks=torch.from_numpy(masks.squeeze(1).astype(bool)),
        alpha=0.9,
        colors=[label_id_to_color[int(i) - 1] for i in labels_id],
    )
    img_masks = F.to_pil_image(img_masks)
    img_masks.save("img_masks.png")
    imgs_list.append("img_masks.png")

    # Draw bboxes to input image and save
    img_bbox = draw_bounding_boxes(img, boxes=torch.from_numpy(boxes), width=4)
    img_bbox = F.to_pil_image(img_bbox)
    img_bbox.save("img_bbox.png")
    imgs_list.append("img_bbox.png")

    # Save masks with their bbox labels & bbox scores
    for col, (mask, label, score) in enumerate(zip(masks, labels, scores)):
        mask = Image.fromarray(mask.squeeze())
        plt.imshow(mask)
        plt.axis("off")
        plt.title(f"{label}: {score:.2f}", fontsize=9)
        plt.savefig(f"mask-{col}.png")
        plt.close()
        imgs_list.append(f"mask-{col}.png")

    return imgs_list


@spaces.GPU(duration=30)
def inference(image, model_name, mask_threshold, bbox_threshold):
    """
    Load the ONNX model and run inference with the provided input `image`. Visualize the predictions and save them in a
    figure, which will be shown in the Gradio app.
    """
    # Load image.
    img = read_image(image)
    # Apply original transformation to the image.
    img_transformed = transforms(img)

    # Download model
    path_onnx = hf_hub_download(
        repo_id="rizavelioglu/fashionfail",
        filename="facere_plus.onnx" if model_name == "facere+" else "facere_base.onnx"
    )
    # Create an inference session.
    ort_session = onnxruntime.InferenceSession(
        path_onnx, providers=["CUDAExecutionProvider", "CPUExecutionProvider"]
    )

    # compute ONNX Runtime output prediction
    ort_inputs = {
        ort_session.get_inputs()[0].name: img_transformed.unsqueeze(dim=0).numpy()
    }
    ort_outs = ort_session.run(None, ort_inputs)

    boxes, labels, scores, masks = ort_outs
    imgs_list = draw_predictions(boxes, labels, scores, masks, img, model_name,
                                 score_threshold=bbox_threshold, proba_threshold=mask_threshold
                                 )

    return imgs_list


title = "Facere - Demo"
description = r"""This is the demo of the paper <a href="https://arxiv.org/abs/2404.08582">FashionFail: Addressing 
Failure Cases in Fashion Object Detection and Segmentation</a>. <br>Upload your image and choose the model for inference
from the dropdown menu—either `Facere` or `Facere+` <br> Check out the <a 
href="https://rizavelioglu.github.io/fashionfail/">project page</a> for more information."""
article = r"""
Example images are sampled from the `FashionFail-test` set, which the models did not see during training.

<br>**Citation** <br>If you find our work useful in your research, please consider giving a star ⭐ and 
a citation: 
``` 
@inproceedings{velioglu2024fashionfail,
  author    = {Velioglu, Riza and Chan, Robin and Hammer, Barbara},
  title     = {FashionFail: Addressing Failure Cases in Fashion Object Detection and Segmentation},
  journal   = {IJCNN},
  eprint    = {2404.08582},
  year      = {2024},
}
``` 
"""

examples = [
    ["adi_103_6.jpg",  "facere", 0.5, 0.7],
    ["adi_103_6.jpg",  "facere+", 0.5, 0.7],
    ["adi_1201_2.jpg", "facere", 0.5, 0.7],
    ["adi_1201_2.jpg", "facere+", 0.5, 0.7],
    ["adi_2149_5.jpg", "facere", 0.5, 0.7],
    ["adi_2149_5.jpg", "facere+", 0.5, 0.7],
    ["adi_5476_3.jpg", "facere", 0.5, 0.7],
    ["adi_5476_3.jpg", "facere+", 0.5, 0.7],
    ["adi_5641_4.jpg", "facere", 0.5, 0.7],
    ["adi_5641_4.jpg", "facere+", 0.5, 0.7]
]

demo = gr.Interface(
    fn=inference,
    inputs=[
        gr.Image(type="filepath", label="input"),
        gr.Dropdown(["facere", "facere+"], value="facere", label="Models"),
        gr.Slider(value=0.5, minimum=0.0, maximum=0.9, step=0.05, label="Mask threshold", info="a threshold for "
                                                                                               "filtering out "
                                                                                               "low-probability ("
                                                                                               "pixel-wise) mask "
                                                                                               "predictions"),
        gr.Slider(value=0.7, minimum=0.0, maximum=0.9, step=0.05, label="BBox threshold", info="a threshold for "
                                                                                               "filtering out "
                                                                                               "low-scoring bbox "
                                                                                               "predictions")
    ],
    outputs=gr.Gallery(label="output", preview=True, height=500),
    title=title,
    description=description,
    article=article,
    examples=examples,
    cache_examples=True,
    examples_per_page=6
)

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