app.py

import torch
from torch import nn
from transformers import AutoModel, AutoTokenizer
import gradio as gr


device = torch.device("cpu")


class RaceClassifier(nn.Module):

    def __init__(self, n_classes):
        super(RaceClassifier, self).__init__()
        self.bert = AutoModel.from_pretrained("vinai/bertweet-base")
        self.drop = nn.Dropout(p=0.3)  # can be changed in future
        self.out = nn.Linear(self.bert.config.hidden_size,
                             n_classes)  # linear layer for the output with the number of classes

    def forward(self, input_ids, attention_mask):
        bert_output = self.bert(
            input_ids=input_ids,
            attention_mask=attention_mask
        )
        last_hidden_state = bert_output[0]
        pooled_output = last_hidden_state[:, 0]
        output = self.drop(pooled_output)
        return self.out(output)


race_labels = {
    0: "African American",
    1: "Asian",
    2: "Latin",
    3: "White"
}

age_labels = {
    0: "Adult",
    1: "Elderly",
    2: "Young"
}

education_labels = {
    0: "Educated",
    1: "Uneducated"
}

gender_labels = {
    0: "Female",
    1: "Male",
    2: "Non-Binary",
    3: "Transgender"
}

orientation_labels = {
    0: "Heterosexual",
    1: "LGBTQ"
}

model_race = RaceClassifier(n_classes=4)
model_race.to(device)
model_race.load_state_dict(torch.load('best_model_race.pt', map_location=torch.device('cpu')))

model_age = RaceClassifier(n_classes=3)
model_age.to(device)
model_age.load_state_dict(torch.load('best_model_age_last.pt', map_location=torch.device('cpu')))

model_education = RaceClassifier(n_classes=2)
model_education.to(device)
model_education.load_state_dict(torch.load('best_model_education_last.pt', map_location=torch.device('cpu')))

model_gender = RaceClassifier(n_classes=4)
model_gender.to(device)
model_gender.load_state_dict(torch.load('best_model_gender_last.pt', map_location=torch.device('cpu')))

model_orientation = RaceClassifier(n_classes=2)
model_orientation.to(device)
model_orientation.load_state_dict(torch.load('best_model_orientation_last.pt', map_location=torch.device('cpu')))


def evaluate(model, input, mask):
    model.eval()
    with torch.no_grad():
        outputs = model(input, mask)
        probs = torch.nn.functional.softmax(outputs, dim=1)
        predictions = torch.argmax(outputs, dim=1)
        predictions = predictions.cpu().numpy()
    return probs, predictions


def write_output(probs, predictions, title, labels):
    output_string = f"{title.upper()}\n Probabilities:\n"
    for i, prob in enumerate(probs[0]):
        print(f"{labels[i]} = {round(prob.item() * 100, 2)}%")
        output_string += f"{labels[i]} = {round(prob.item() * 100, 2)}%\n"

    output_string += f"Predicted as: {labels[predictions[0]]}\n"

    return output_string


def predict(*text):
    tweets = [tweet for tweet in text if tweet]
    print(tweets)
    sentences = tweets

    tokenizer = AutoTokenizer.from_pretrained("vinai/bertweet-base", normalization=True)

    encoded_sentences = tokenizer(
        sentences,
        padding=True,
        truncation=True,
        return_tensors='pt',
        max_length=128,
    )

    input_ids = encoded_sentences["input_ids"].to(device)
    attention_mask = encoded_sentences["attention_mask"].to(device)

    race_probs, race_predictions = evaluate(model_race, input_ids, attention_mask)
    age_probs, age_predictions = evaluate(model_age, input_ids, attention_mask)
    education_probs, education_predictions = evaluate(model_education, input_ids, attention_mask)
    gender_probs, gender_predictions = evaluate(model_gender, input_ids, attention_mask)
    orientation_probs, orientation_predictions = evaluate(model_orientation, input_ids, attention_mask)

    final_output = str()
    final_output += write_output(race_probs, race_predictions, "race", race_labels)
    final_output += "\n"
    final_output += write_output(age_probs, age_predictions,"age",age_labels)
    final_output += "\n"
    final_output += write_output(education_probs,education_predictions,"education", education_labels)
    final_output += "\n"
    final_output += write_output(gender_probs, gender_predictions, "gender", gender_labels)
    final_output += "\n"
    final_output += write_output(orientation_probs, orientation_predictions, "sexual orientation", orientation_labels)

    return final_output


max_textboxes = 20


def update_textboxes(k):
    components = []
    if k is None:
        k = 0
    for i in range(max_textboxes):
        if i < k:
            components.append(gr.update(visible=True))
        else:
            components.append(gr.update(visible=False))
    return components


def clear_textboxes():
    return [gr.update(value='') for _ in range(max_textboxes)]

def clear_output_box():
    return gr.update(value='')

with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column(scale=1):
            s = gr.Slider(1, max_textboxes, value=1, step=1, label="How many tweets do you want to enter:")
            textboxes = [gr.Textbox(label=f"Tweet {i + 1}", visible=(i == 0)) for i in range(max_textboxes)]
            s.change(fn=update_textboxes, inputs=s, outputs=textboxes)
            btn = gr.Button("Predict")
            btn_clear = gr.Button("Clear")
        with gr.Column(scale=1):
            output = gr.Textbox(label="Profile of User")

    btn.click(fn=predict, inputs=textboxes, outputs=output)
    btn_clear.click(fn=clear_textboxes, outputs=textboxes)
    btn_clear.click(fn=clear_output_box, outputs=output)


demo.launch()