app.py

import transformers
import torch

import streamlit as st

from transformers import BertTokenizer

st.markdown("### Из какой области статья? Введите название и аннотация научной статьи и я попробую угадать из какой она области)")
# link = 'https://www.clipartmax.com/png/middle/87-873210_akinator-with-transparent-background.png'
# st.markdown(f"<img width=200px src='{link}'>", unsafe_allow_html=True)
# st.markdown("<img width=200px src='https://rozetked.me/images/uploads/dwoilp3BVjlE.jpg'>", unsafe_allow_html=True)

# from transformers import 

# pipe = pipeline("ner", "Davlan/distilbert-base-multilingual-cased-ner-hrl")


num_classes = 8
class BERTClass(torch.nn.Module):
    def __init__(self, n_hid1 = 1024, n_out=num_classes, bert_path='bert-base-uncased'):
        super(BERTClass, self).__init__()
        self.l1 = transformers.BertModel.from_pretrained(bert_path)
        self.l2 = torch.nn.Dropout(0.3)
        self.l3 = torch.nn.Linear(768, n_hid1)
        self.l4 = torch.nn.ReLU()
        self.l5 = torch.nn.Dropout(0.2)
        self.l6 = torch.nn.Linear(n_hid1, n_out)
    
    def forward(self, ids, mask, token_type_ids):
#         _, output_1= self.l1(ids, attention_mask = mask, token_type_ids = token_type_ids)
        out = self.l1(ids, attention_mask = mask, token_type_ids = token_type_ids)
        out = self.l2(out[1])
        out = self.l3(out)
        out = self.l4(out)
        out = self.l5(out)
        out = self.l6(out)
        return out

@st.cache(suppress_st_warning=True,  allow_output_mutation=True)
def load_bert():
    model = BERTClass(bert_path='bert_pretrained')
    model.load_state_dict(torch.load('bert_pretrained.pt'))
    model.eval()
    
    tokenizer = BertTokenizer.from_pretrained('bert_tokenizer')

    return model, tokenizer


def apply_bert(text, model, tokenizer):
    """returns probabilities"""
    MAX_LEN = 200
    ins = tokenizer.encode_plus(text, None, add_special_tokens=True,
            max_length=MAX_LEN,
            pad_to_max_length=True,
            return_token_type_ids=True
        )
    ids = torch.tensor(ins['input_ids']).unsqueeze(0)
    mask = torch.tensor(ins['attention_mask']).unsqueeze(0)
    token_type_ids = torch.tensor(ins["token_type_ids"])
    out = model(ids, mask, token_type_ids)
    return torch.sigmoid(out).flatten().detach()


class TinyBERTClass(torch.nn.Module):
    def __init__(self, n_hid1 = 1024, n_out=num_classes, path='distilbert-base-uncased'):
        super(TinyBERTClass, self).__init__()
        self.l1 = transformers.DistilBertModel.from_pretrained(path)
        self.l2 = torch.nn.Dropout(0.3)
        self.l3 = torch.nn.Linear(768, n_hid1)
        self.l4 = torch.nn.ReLU()
        self.l5 = torch.nn.Dropout(0.2)
        self.l6 = torch.nn.Linear(n_hid1, n_out)
    
    def forward(self, ids, mask):
#         _, output_1= self.l1(ids, attention_mask = mask, token_type_ids = token_type_ids)
        out = self.l1(ids, attention_mask = mask)
        out = self.l2(out.last_hidden_state[:,0,:])
        out = self.l3(out)
        out = self.l4(out)
        out = self.l5(out)
        out = self.l6(out)
        return out


@st.cache(suppress_st_warning=True, allow_output_mutation=True)
def load_tiny_bert():
    model = TinyBERTClass(path = 'tiny_bert_pretrained')
    model.load_state_dict(torch.load('tiny_bert.pt'))
    model.eval()
    
    tokenizer = transformers.DistilBertTokenizer.from_pretrained('tiny_bert_tokenizer')
    
    return model, tokenizer


def apply_tiny_bert(text, model, tokenizer):
    encoded_input = tokenizer(text, return_tensors='pt')
    out = model(encoded_input['input_ids'], encoded_input['attention_mask'])
    
    return torch.sigmoid(out).flatten().detach()
    


title = st.text_area("Название статьи")
if not title.endswith('.') and title:
	title += '.'

summary = st.text_area("Аннотация статьи")

calc_button = st.button('Угадать тематику')

bert_model, bert_tokenizer = load_bert()
tiny_bert, tiny_bert_tokenizer = load_tiny_bert()

# calculate ================================================================
if calc_button:
    print('title')
    print(title)
    print('=' * 80)
    # print(text)

    if summary:
        text = title + summary
        out = apply_bert(text, bert_model, bert_tokenizer)
    else:
        out = apply_tiny_bert(title, tiny_bert, tiny_bert_tokenizer)


    RU_NAMES = ['компьютерным наукам'
               ,'экономике'
               ,'электротехнике и системотехнике'
               ,'математике'
               ,'физике'
               ,'количественной биологии'
               ,'количественным финансам'
               ,'статистике'
               ]

    def get_classes(out, bandwidth = 0.5):
        res = []
        for i in range(out.size()[0]):
            if out[i] >= bandwidth:
                res.append(i)
        
        ans = ''
        total = 0
        for i in res:
            total += out[i].item()
            if not ans:
                ans += f'\nэто статья по {RU_NAMES[i]} с вероятностью {out[i].item():.2f}'
            else:
                ans += f',\nтакже она по {RU_NAMES[i]} с вероятностью {out[i].item():.2f}'
            
        ans = 'Э' + ans[2:]
        if total >= 1.0:
            ans += '.\n(Решалась задача мультиклассификации, поэтому сумма вероятностей получилась больше 1.)'

        if ans == 'Э':
            return 'Не похоже на что-то научное, Вы уверены что это взято из статьи?'
        return ans
            
    res = get_classes(out)

    st.markdown(f"{res}")