app.py

from transformers import BertTokenizer, BertForSequenceClassification
import torch
import gradio as gr
from transformers import TextClassificationPipeline

model_name='indolem/indobert-base-uncased'
label_dict={'Pasal 112 UU RI No. 35 Thn 2009': 0,
 'Pasal 114 UU RI No. 35 Thn 2009': 1,
 'Pasal 111 UU RI No. 35 Thn 2009': 2,
 'Pasal 127 UU RI No. 35 Thn 2009': 3,
 'Pasal 363 KUHP': 4,
 'Pasal 365 KUHP': 5,
 'Pasal 362 KUHP': 6,
 'Pasal 338 KUHP': 7,
 'Pasal 340 KUHP': 8,
 'Pasal 374 KUHP': 9,
 'Pasal 372 KUHP': 10,
 'Pasal 378 KUHP': 11,
 'Pasal 351 KUHP': 12,
 'Pasal 303 KUHP': 13}

tokenizer = BertTokenizer.from_pretrained(model_name)
model = BertForSequenceClassification.from_pretrained(model_name,
                                                      num_labels=len(label_dict),
                                                      output_attentions=False,
                                                      output_hidden_states=False)

torch_model = torch.load('FineTune_IndoLEM_BERT_H_Mean_Pooling_LR1E-5_BS2_epoch_9.model')
torch_model['classifier.weight'] = torch_model.pop('out.weight')
torch_model['classifier.bias'] = torch_model.pop('out.bias')
model.load_state_dict(torch_model)

pipe = TextClassificationPipeline(model=model, tokenizer=tokenizer, return_all_scores=True)

def get_nth_key(dictionary, n=0):
    if n < 0:
        n += len(dictionary)
    for i, key in enumerate(dictionary.keys()):
        if i == n:
            return key
    raise IndexError("dictionary index out of range")

def predict(text):
    predictions = pipe(text)[0]
    max = 0
    idx = -1
    for i in range(len(predictions)):
        if max < predictions[i]['score']:
            max = predictions[i]['score']
            idx = i
    return get_nth_key(label_dict, idx)

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