app.py

import ast
import json
import base64
import spaces
import requests
import numpy as np
import gradio as gr
from PIL import Image
from io import BytesIO
import face_recognition
from turtle import title
from openai import OpenAI
from collections import Counter
from transformers import pipeline

import urllib.request
from transformers import YolosImageProcessor, YolosForObjectDetection
import torch
import matplotlib.pyplot as plt
from torchvision.transforms import ToTensor, ToPILImage

    
client = OpenAI()

pipe = pipeline("zero-shot-image-classification", model="patrickjohncyh/fashion-clip")

color_file_path = 'color_config.json'
attributes_file_path = 'attributes_config.json'
import os
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")


# Open and read the COLOR JSON file
with open(color_file_path, 'r') as file:
    color_data = json.load(file)

# Open and read the ATTRIBUTES JSON file
with open(attributes_file_path, 'r') as file:
    attributes_data = json.load(file)

COLOURS_DICT = color_data['color_mapping']
ATTRIBUTES_DICT = attributes_data['attribute_mapping']


def shot(input, category, level):
    output_dict = {}
    if level = 'variant':
        subColour,mainColour,score = get_colour(ast.literal_eval(str(input)),category)    
        openai_parsed_response = get_openAI_tags(ast.literal_eval(str(input)))
        face_embeddings = get_face_embeddings(ast.literal_eval(str(input)))
        cropped_images = get_cropped_images(ast.literal_eval(str(input)),category)
        output_dict['colors'] = {
            "main":mainColour,
            "sub":subColour,
            "score":score
        }
        output_dict['image_mapping'] = openai_parsed_response
        output_dict['face_embeddings'] = face_embeddings
        output_dict['cropped_images'] = cropped_images


    if level = 'product':
        common_result = get_predicted_attributes(ast.literal_eval(str(input)),category)    
        output_dict['attributes'] = common_result
        output_dict['subcategory'] = category
    
    return output_dict



@spaces.GPU  
def get_colour(image_urls, category):
    colourLabels = list(COLOURS_DICT.keys())
    for i in range(len(colourLabels)):
        colourLabels[i] = colourLabels[i] + " clothing: " + category

    responses = pipe(image_urls, candidate_labels=colourLabels)
    # Get the most common colour
    mainColour = responses[0][0]['label'].split(" clothing:")[0]


    if mainColour not in COLOURS_DICT:
        return None, None, None

    # Add category to the end of each label
    labels = COLOURS_DICT[mainColour]
    for i in range(len(labels)):
        labels[i] = labels[i] + " clothing: " + category

    # Run pipeline in one go
    responses = pipe(image_urls, candidate_labels=labels)
    subColour = responses[0][0]['label'].split(" clothing:")[0]

    return subColour, mainColour, responses[0][0]['score']

@spaces.GPU  
def get_predicted_attributes(image_urls, category):
    # Get the predicted attributes for the image
    # attributes = get_category_attributes(category)
    
    attributes = list(ATTRIBUTES_DICT.get(category,{}).keys())
    
    # Mapping of possible values per attribute
    common_result = []
    for attribute in attributes:
        # values = get_attribute_values(attribute, category)
        values = ATTRIBUTES_DICT.get(category,{}).get(attribute,[])

        if len(values) == 0:
            continue

        # Adjust labels for the pipeline to be in format: "{attr}: {value}, clothing: {category}"
        attribute = attribute.replace("colartype", "collar").replace("sleevelength", "sleeve length").replace("fabricstyle", "fabric")
        values = [f"{attribute}: {value}, clothing: {category}" for value in values]

        # Get the predicted values for the attribute
        responses = pipe(image_urls, candidate_labels=values)
        result = [response[0]['label'].split(", clothing:")[0] for response in responses]

        # If attribute is details, then get the top 2 most common labels
        if attribute == "details":
            result += [response[1]['label'].split(", clothing:")[0] for response in responses]
            common_result.append(Counter(result).most_common(2))
        else:
            common_result.append(Counter(result).most_common(1))

    # Clean up the results into one long string
    for i, result in enumerate(common_result):
        common_result[i] = ", ".join([f"{x[0]}" for x in result])
    
    result = {}

    # Iterate through the list and split each item into key and value
    for item in common_result:
        # Split by ': ' to separate the key and value
        key, value = item.split(': ', 1)
        # Add to the dictionary
        result[key] = value

    return result


def get_openAI_tags(image_urls):
    # Create list containing JSONs of each image URL
    imageList = []
    for image in image_urls:
        imageList.append({"type": "image_url", "image_url": {"url": image}})
    
    openai_response = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {
            "role": "system",
            "content": [
                {
                "type": "text",
                "text": "You're a tagging assistant, you will help label and tag product pictures for my online e-commerce platform. Your tasks will be to return which angle the product images were taken from. You will have to choose from 'full-body', 'half-body', 'side', 'back', or 'zoomed' angles. You should label each of the images with one of these labels depending on which you think fits best (ideally, every label should be used at least once, but only if there are 5 or more images), and should respond with an unformatted dictionary where the key is a string representation of the url index of the url and the value is the assigned label."
                }
            ]
            },
            {
            "role": "user",
            "content": imageList
            },
        ],
        temperature=1,
        max_tokens=500,
        top_p=1,
        frequency_penalty=0,
        presence_penalty=0
    )
    response = json.loads(openai_response.choices[0].message.content)
    return response


@spaces.GPU
def get_face_embeddings(image_urls):
    # Initialize a dictionary to store the face encodings or errors
    results = {}

    # Loop through each image URL
    for index, url in enumerate(image_urls):
        try:
            # Try to download the image from the URL
            response = requests.get(url)
            # Raise an exception if the response is not successful
            response.raise_for_status()

            # Load the image using face_recognition
            image = face_recognition.load_image_file(BytesIO(response.content))

            # Get the face encodings for all faces in the image
            face_encodings = face_recognition.face_encodings(image)

            # If no faces are detected, store an empty list
            if not face_encodings:
                results[str(index)] = []
            else:
                # Otherwise, store the first face encoding as a list
                results[str(index)] = face_encodings[0].tolist()
        except Exception as e:
            # If any error occurs during the download or processing, store the error message
            results[str(index)] = f"Error processing image: {str(e)}"

    return results

# new
ACCURACY_THRESHOLD = 0.86

def open_image_from_url(url):
    # Fetch the image from the URL
    response = requests.get(url, stream=True)
    response.raise_for_status()  # Check if the request was successful

    # Open the image using PIL
    image = Image.open(BytesIO(response.content))

    return image

# Add the main data to the session state
main = [['Product Id', 'Sku', 'Color', 'Images', 'Status', 'Category', 'Text']]

# This is the order of the categories list. NO NOT CHANGE. Just for visualization purposes
cats = ['shirt, blouse', 'top, t-shirt, sweatshirt', 'sweater', 'cardigan', 'jacket', 'vest', 'pants', 'shorts', 'skirt', 'coat', 'dress', 'jumpsuit', 'cape', 'glasses', 'hat', 'headband, head covering, hair accessory', 'tie', 'glove', 'watch', 'belt', 'leg warmer', 'tights, stockings', 'sock', 'shoe', 'bag, wallet', 'scarf', 'umbrella', 'hood', 'collar', 'lapel', 'epaulette', 'sleeve', 'pocket', 'neckline', 'buckle', 'zipper', 'applique', 'bead', 'bow', 'flower', 'fringe', 'ribbon', 'rivet', 'ruffle', 'sequin', 'tassel']

filter = ['dress', 'jumpsuit', 'cape', 'glasses', 'hat', 'headband, head covering, hair accessory', 'tie', 'glove', 'watch', 'belt', 'leg warmer', 'tights, stockings', 'sock', 'shoe', 'scarf', 'umbrella', 'hood', 'collar', 'lapel', 'epaulette', 'sleeve', 'pocket', 'neckline', 'buckle', 'zipper', 'applique', 'bead', 'bow', 'flower', 'fringe', 'ribbon', 'rivet', 'ruffle', 'sequin', 'tassel']

# 0 for full body, 1 for upper body, 2 for lower body, 3 for over body (jacket, coat, etc), 4 for accessories
yolo_mapping = {
    'shirt, blouse': 3,
    'top, t-shirt, sweatshirt' : 1,
    'sweater': 1,
    'cardigan': 1,
    'jacket': 3,
    'vest': 1,
    'pants': 2,
    'shorts': 2,
    'skirt': 2,
    'coat': 3,
    'dress': 0,
    'jumpsuit': 0,
    'bag, wallet': 4
}

# First line full body, second line upper body, third line lower body, fourth line over body, fifth line accessories
label_mapping = [
    ['women-dress-mini', 'women-dress-dress', 'women-dress-maxi', 'women-dress-midi', 'women-playsuitsjumpsuits-playsuit', 'women-playsuitsjumpsuits-jumpsuit', 'women-coords-coords', 'women-swimwear-onepieces', 'women-swimwear-bikinisets'],
    ['women-sweatersknits-cardigan', 'women-top-waistcoat', 'women-top-blouse', 'women-sweatersknits-blouse', 'women-sweatersknits-sweater', 'women-top-top', 'women-loungewear-hoodie', 'women-top-camistanks', 'women-top-tshirt', 'women-top-croptop', 'women-loungewear-sweatshirt', 'women-top-body'],
    ['women-loungewear-joggers', 'women-bottom-trousers', 'women-bottom-leggings', 'women-bottom-jeans', 'women-bottom-shorts', 'women-bottom-skirt', 'women-loungewear-activewear', 'women-bottom-joggers'],
    ['women-top-shirt', 'women-outwear-coatjacket', 'women-outwear-blazer', 'women-outwear-coatjacket', 'women-outwear-kimonos'],
    ['women-accessories-bags']
]

MODEL_NAME = "valentinafeve/yolos-fashionpedia"

feature_extractor = YolosImageProcessor.from_pretrained('hustvl/yolos-small')
model = YolosForObjectDetection.from_pretrained(MODEL_NAME)

def get_category_index(category):
    # Find index of label mapping
    for i, labels in enumerate(label_mapping):
        if category in labels:
            break
    return i

def get_yolo_index(category):
    # Find index of yolo mapping
    return yolo_mapping[category]

def fix_channels(t):
    """
    Some images may have 4 channels (transparent images) or just 1 channel (black and white images), in order to let the images have only 3 channels. I am going to remove the fourth channel in transparent images and stack the single channel in back and white images.
    :param t: Tensor-like image
    :return: Tensor-like image with three channels
    """
    if len(t.shape) == 2:
        return ToPILImage()(torch.stack([t for i in (0, 0, 0)]))
    if t.shape[0] == 4:
        return ToPILImage()(t[:3])
    if t.shape[0] == 1:
        return ToPILImage()(torch.stack([t[0] for i in (0, 0, 0)]))
    return ToPILImage()(t)

def idx_to_text(i):
    return cats[i]

# Random colors used for visualization
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]]

# for output bounding box post-processing
def box_cxcywh_to_xyxy(x):
    x_c, y_c, w, h = x.unbind(1)
    b = [(x_c - 0.5 * w), (y_c - 0.5 * h),
         (x_c + 0.5 * w), (y_c + 0.5 * h)]
    return torch.stack(b, dim=1)

def rescale_bboxes(out_bbox, size):
    img_w, img_h = size
    b = box_cxcywh_to_xyxy(out_bbox)
    b = b * torch.tensor([img_w, img_h, img_w, img_h], dtype=torch.float32)

    return b

def plot_results(pil_img, prob, boxes):
    plt.figure(figsize=(16,10))
    plt.imshow(pil_img)
    ax = plt.gca()
    colors = COLORS * 100
    i = 0

    crops = []
    crop_classes = []
    for p, (xmin, ymin, xmax, ymax), c in zip(prob, boxes.tolist(), colors):
        cl = p.argmax()

        # Save each box as an image
        box_img = pil_img.crop((xmin, ymin, xmax, ymax))
        crops.append(box_img)
        crop_classes.append(idx_to_text(cl))

        ax.add_patch(plt.Rectangle((xmin, ymin), xmax - xmin, ymax - ymin,
                                   fill=False, color=c, linewidth=3))

        ax.text(xmin, ymin, idx_to_text(cl), fontsize=10,
                bbox=dict(facecolor=c, alpha=0.8))

        i += 1

    # Remove white padding all around the image
    plt.axis('off')
    plt.subplots_adjust(left=0, right=1, top=1, bottom=0)
    output_img = plt.gcf()
    plt.close()

    return output_img, crops, crop_classes


def visualize_predictions(image, outputs, threshold=0.8):
    # Keep only predictions with confidence >= threshold
    probas = outputs.logits.softmax(-1)[0, :, :-1]
    keep = probas.max(-1).values > threshold

    # Convert predicted boxes from [0; 1] to image scales
    bboxes_scaled = rescale_bboxes(outputs.pred_boxes[0, keep].cpu(), image.size)

    # Get filtered probabilities and boxes based on the filter list
    filter_set = set(filter)
    filtered_probas_boxes = [
        (proba, box) for proba, box in zip(probas[keep], bboxes_scaled)
        if idx_to_text(proba.argmax()) not in filter_set
    ]

    # If there is a jumpsuit or dress detected, remove them if there are other clothes detected
    contains_jumpsuit_or_dress = any(idx_to_text(proba.argmax()) in ["jumpsuit", "dress"] for proba, _ in filtered_probas_boxes)
    if contains_jumpsuit_or_dress and len(filtered_probas_boxes) > 1:
        filtered_probas_boxes = [
            (proba, box) for proba, box in filtered_probas_boxes
            if idx_to_text(proba.argmax()) not in ["jumpsuit", "dress"]
        ]

    # Remove duplicates: Only keep one box per class
    unique_classes = set()
    unique_filtered_probas_boxes = []
    for proba, box in filtered_probas_boxes:
        class_text = idx_to_text(proba.argmax())
        if class_text not in unique_classes:
            unique_classes.add(class_text)
            unique_filtered_probas_boxes.append((proba, box))

    # If there are remaining filtered probabilities, plot results
    output_img = None
    crops = None
    crop_classes = None
    if unique_filtered_probas_boxes:
        final_probas, final_boxes = zip(*unique_filtered_probas_boxes)
        output_img, crops, crop_classes = plot_results(image, list(final_probas), torch.stack(final_boxes))

    # Return the classes of the detected objects
    return [proba.argmax().item() for proba, _ in unique_filtered_probas_boxes], output_img, crops, crop_classes

@spaces.GPU 
def get_objects(image, threshold=0.8):
    class_counts = {}
    image = fix_channels(ToTensor()(image))
    image = image.resize((600, 800))

    inputs = feature_extractor(images=image, return_tensors="pt")
    outputs = model(**inputs)

    detected_classes, output_img, crops, crop_classes = visualize_predictions(image, outputs, threshold=threshold)
    for cl in detected_classes:
        class_name = idx_to_text(cl)
        if class_name not in class_counts:
            class_counts[class_name] = 0
        class_counts[class_name] += 1

    if crop_classes is not None:
        crop_classes = [get_yolo_index(c) for c in crop_classes]

    return class_counts, output_img, crops, crop_classes



def get_cropped_images(images,category):
    cropped_list = []
    resultsPerCategory = {}
    for num, image in enumerate(images):
        image = open_image_from_url(image)
        class_counts, output_img, cropped_images, cropped_classes = get_objects(image, 0.37)
        print(cropped_images)
        if not class_counts:
            continue
        
        # Get the inverse category as any other mapping label except the current one corresponding category
        inverse_category = [label for i, labels in enumerate(label_mapping) for label in labels if i != get_category_index(category) and i != 0]
        
        # If category is a cardigan, we don't recommend category indices 1 and 3
        if category == 'women-sweatersknits-cardigan':
            inverse_category = [label for i, labels in enumerate(label_mapping) for label in labels if i != get_category_index(category) and i != 1 and i != 3]

        for i, image in enumerate(cropped_images):
            cropped_category = cropped_classes[i]
            print(cropped_category, cropped_classes[i], get_category_index(category))
            
            specific_category = label_mapping[cropped_category]

            if cropped_category == get_category_index(category):
                continue

            cropped_list.append(image)
    
    return cropped_list




# Define the Gradio interface with the updated components
iface = gr.Interface(
    fn=shot, 
    inputs=[
        gr.Textbox(label="Image URLs (starting with http/https) comma seperated "), 
        gr.Textbox(label="Category")
        gr.Textbox(label="Level; accepted 'variant' or 'product'")
    ], 
    outputs="text" ,
     examples=[
        [['https://d2q1sfov6ca7my.cloudfront.net/eyJidWNrZXQiOiAiaGljY3VwLWltYWdlLWhvc3RpbmciLCAia2V5IjogIlc4MDAwMDAwMTM0LU9SL1c4MDAwMDAwMTM0LU9SLTEuanBnIiwgImVkaXRzIjogeyJyZXNpemUiOiB7IndpZHRoIjogODAwLCAiaGVpZ2h0IjogMTIwMC4wLCAiZml0IjogIm91dHNpZGUifX19',
 'https://d2q1sfov6ca7my.cloudfront.net/eyJidWNrZXQiOiAiaGljY3VwLWltYWdlLWhvc3RpbmciLCAia2V5IjogIlc4MDAwMDAwMTM0LU9SL1c4MDAwMDAwMTM0LU9SLTIuanBnIiwgImVkaXRzIjogeyJyZXNpemUiOiB7IndpZHRoIjogODAwLCAiaGVpZ2h0IjogMTIwMC4wLCAiZml0IjogIm91dHNpZGUifX19',
 'https://d2q1sfov6ca7my.cloudfront.net/eyJidWNrZXQiOiAiaGljY3VwLWltYWdlLWhvc3RpbmciLCAia2V5IjogIlc4MDAwMDAwMTM0LU9SL1c4MDAwMDAwMTM0LU9SLTMuanBnIiwgImVkaXRzIjogeyJyZXNpemUiOiB7IndpZHRoIjogODAwLCAiaGVpZ2h0IjogMTIwMC4wLCAiZml0IjogIm91dHNpZGUifX19'], "women-top-shirt"]],
    description="Add an image URL (starting with http/https) or upload a picture, and provide a list of labels separated by commas.",
    title="Full product flow"
)

# Launch the interface
iface.launch()