Webpage.py

import gradio as gr
from gradio_image_prompter import ImagePrompter
import Predict
import XGBoost_utils
import numpy as np
import cv2 as cv

GENERAL_CATEGORY = {'Potatoes / Vegetables / Fruit': 0, 'Chemical products': 1, 'Photo / Film / Optical items': 2, 'Catering industry': 3, 'Industrial products other': 4, 'Media': 5, 'Real estate': 6, 'Government': 7, 'Personnel advertisements': 8, 'Cars / Commercial vehicles': 9, 'Cleaning products': 10, 'Retail': 11, 'Fragrances': 12, 'Footwear / Leather goods': 13, 'Software / Automation': 14, 'Telecommunication equipment': 15, 'Tourism': 16, 'Transport/Communication companies': 17, 'Transport services': 18, 'Insurances': 19, 'Meat / Fish / Poultry': 20, 'Detergents': 21, 'Foods General': 22, 'Other services': 23, 'Banks and Financial Services': 24, 'Office Products': 25, 'Household Items': 26, 'Non-alcoholic beverages': 27, 'Hair, Oral and Personal Care': 28, 'Fashion and Clothing': 29, 'Other products and Services': 30, 'Paper products': 31, 'Alcohol and Other Stimulants': 32, 'Medicines': 33, 'Recreation and Leisure': 34, 'Electronics': 35, 'Home Furnishings': 36, 'Products for Business Use': 37}
CATEGORIES = list(GENERAL_CATEGORY.keys())
CATEGORIES.sort()

def calculate_areas(prompts, brand_num, pictorial_num, text_num):
    points_all = prompts["points"]
    brand_surf = 0
    for i in range(brand_num):
        x1 = points_all[i][0]; y1 = points_all[i][1]
        x2 = points_all[i][3]; y2 = points_all[i][4]
        brand_surf += np.abs((x1-x2)*(y1-y2))

    pictorial_surf = 0
    for i in range(brand_num, brand_num+pictorial_num):
        x1 = points_all[i][0]; y1 = points_all[i][1]
        x2 = points_all[i][3]; y2 = points_all[i][4]
        pictorial_surf += np.abs((x1-x2)*(y1-y2))
    
    text_surf = 0
    for i in range(brand_num+pictorial_num, brand_num+pictorial_num+text_num):
        x1 = points_all[i][0]; y1 = points_all[i][1]
        x2 = points_all[i][3]; y2 = points_all[i][4]
        text_surf += np.abs((x1-x2)*(y1-y2))

    ad_size = 0
    x1 = points_all[-2][0]; y1 = points_all[-2][1]
    x2 = points_all[-2][3]; y2 = points_all[-2][4]
    ad_size += np.abs((x1-x2)*(y1-y2))

    whole_size = 0
    x1 = points_all[-1][0]; y1 = points_all[-1][1]
    x2 = points_all[-1][3]; y2 = points_all[-1][4]
    whole_size += np.abs((x1-x2)*(y1-y2))

    return (brand_surf/whole_size*100, pictorial_surf/whole_size*100, text_surf/whole_size*100, ad_size/whole_size*100)


def attention(whole_display_prompt, ad, context, 
              brand_num, pictorial_num, text_num,
              category, ad_location, gaze_type):
    text_detection_model_path = '../XGBoost_Prediction_Model/EAST-Text-Detection/frozen_east_text_detection.pb'
    LDA_model_pth = '../XGBoost_Prediction_Model/LDA_Model_trained/lda_model_best_tot.model'
    training_ad_text_dictionary_path = '../XGBoost_Prediction_Model/LDA_Model_trained/object_word_dictionary'
    training_lang_preposition_path = '../XGBoost_Prediction_Model/LDA_Model_trained/dutch_preposition'

    caption_ad = XGBoost_utils.Caption_Generation(ad)
    caption_context = XGBoost_utils.Caption_Generation(context)
    ad_topic = XGBoost_utils.Topic_emb(caption_ad)
    ctpg_topic = XGBoost_utils.Topic_emb(caption_context)
    # ad_topic = np.random.randn(1,768)
    # ctpg_topic = np.random.randn(1,768)

    prod_group = np.zeros(38)
    prod_group[GENERAL_CATEGORY[category]] = 1

    if ad_location == 'left':
        ad_loc = 0
    elif ad_location == 'right':
        ad_loc = 1
    else:
        ad_loc = None

    brand_percent, visual_percent, text_percent, adv_size_percent = calculate_areas(whole_display_prompt, brand_num, pictorial_num, text_num)
    surfaces = [brand_percent, visual_percent, text_percent, adv_size_percent*10/100]

    ad = ad.convert('RGB')
    ad = cv.resize(np.array(ad), (640, 832))
    context = context.convert('RGB')
    context = cv.resize(np.array(context), (640, 832))


    Gaze = Predict.Ad_Gaze_Prediction(input_ad_path=ad, input_ctpg_path=context, ad_location=ad_loc,
                                    text_detection_model_path=text_detection_model_path, LDA_model_pth=LDA_model_pth, 
                                    training_ad_text_dictionary_path=training_ad_text_dictionary_path, training_lang_preposition_path=training_lang_preposition_path, training_language='dutch', 
                                    ad_embeddings=ad_topic, ctpg_embeddings=ctpg_topic,
                                    surface_sizes=surfaces, Product_Group=prod_group,
                                    obj_detection_model_pth=None, num_topic=20, Gaze_Time_Type=gaze_type)
    return np.round(Gaze,2)

def greet(name, intensity):
    return "Hello " * intensity + name + "!"

demo = gr.Interface(
    fn=attention,
    inputs=[ImagePrompter(label="Upload Entire (Ad+Context) Image, and Draw Bounding Boxes"),
            gr.Image(label="Ad Image", sources=['upload', 'webcam'], type="pil"), 
            gr.Image(label="Context Image", sources=['upload', 'webcam'], type="pil"),
            gr.Number(label="Number of brand bounding boxes drawn"),
            gr.Number(label="Number of pictorial bounding boxes drawn"),
            gr.Number(label="Number of text bounding boxes drawn"),
            gr.Dropdown(CATEGORIES, label="Product Category"),
            gr.Textbox(label="Ad Location", info="Enter left or right or full"),
            gr.Textbox(label="Gaze Type", info="Enter Ad or Brand")
            ],
    outputs=[gr.Number(label="Predicted Gaze (sec)")],
    title="Gaze Prediction",
    description="In the section right below, please first upload the entire image that contains both ad and context, then draw bounding boxes. Please draw ALL Bounding Boxes in the order of: (1) brand, (2) pictorial, (3) text elements, (4) advertisement and (5) the entire image here. NOTE: Each ad element can have more than 1 boxes.",
    theme=gr.themes.Soft()
)

demo.launch(share=True)