app.py

import gradio as gr
import torch

import torch.nn.functional as F
from transformers import BertTokenizer, BertForSequenceClassification


tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
model = BertForSequenceClassification.from_pretrained("bert-base-uncased")

model.load_state_dict(torch.load('model_after_train.pt', map_location=torch.device('cpu')), strict=False)
model.eval()
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')


def preprocess_text(text):
    parts = []

    text_len = len(text.split(' '))
    delta = 300
    max_parts = 5
    nb_cuts = int(text_len / delta)
    nb_cuts = min(nb_cuts, max_parts)
    
    
    for i in range(nb_cuts + 1):
        text_part = ' '.join(text.split(' ')[i * delta: (i + 1) * delta])
        parts.append(tokenizer.encode(text_part, return_tensors="pt", max_length=500).to(device))

    return parts

def test(text):
    text_parts = preprocess_text(text)
    overall_output = torch.zeros((1,2)).to(device)
    try:
        for part in text_parts:
            if len(part) > 0:
                overall_output += model(part.reshape(1, -1))[0]
    except RuntimeError:
        print("GPU out of memory, skipping this entry.")

    overall_output = F.softmax(overall_output[0], dim=-1)

    value, result = overall_output.max(0)

    term = "fake"
    if result.item() == 0:
        term = "real"

    return term + " at " + str(int(value.item()*100)) + " %"


iface = gr.Interface(fn=test, inputs="text", outputs="text")
iface.launch()