app.py

import streamlit as st 
import torch
from torch import nn
from transformers import BertModel, AutoTokenizer, AutoModel, pipeline
from time import time
import matplotlib.pyplot as plt
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device = 'cpu'

# dict for decoding / enclding labels
labels = {'cs.NE': 0, 'cs.CL': 1, 'cs.AI': 2, 'stat.ML': 3, 'cs.CV': 4, 'cs.LG': 5}
labels_decoder = {'cs.NE': 'Neural and Evolutionary Computing', 'cs.CL': 'Computation and Language', 'cs.AI': 'Artificial Intelligence', 
 'stat.ML': 'Machine Learning (stat)', 'cs.CV': 'Computer Vision', 'cs.LG': 'Machine Learning'}
 
model_name = 'bert-base-uncased'
tokenizer = AutoTokenizer.from_pretrained(model_name)

class BertClassifier(nn.Module):

    def __init__(self, n_classes, dropout=0.5, model_name='bert-base-uncased'):
        super(BertClassifier, self).__init__()
        self.bert = BertModel.from_pretrained(model_name)
        self.dropout = nn.Dropout(dropout)
        self.linear = nn.Linear(768, n_classes)
        self.relu = nn.ReLU()

    def forward(self, input_id, mask):
        _, pooled_output = self.bert(input_ids=input_id, attention_mask=mask,return_dict=False)
        dropout_output = self.dropout(pooled_output)
        linear_output = self.linear(dropout_output)
        final_layer = self.relu(linear_output)
        return final_layer

@st.cache(suppress_st_warning=True)
def build_model():
    model = BertClassifier(n_classes=len(labels))
    st.markdown("Model created")
    model.load_state_dict(torch.load('model_weights_1.pt', map_location=torch.device('cpu')))
    model.eval()
    st.markdown("Model weights loaded")
    return model
    
def inference(txt, mode=None):
    # infers classes for text topic based on the trained model from above
    # has separate mode 'print' for just output
    t2 = tokenizer(txt.lower().replace('\n', ''), 
       padding='max_length', max_length = 512, truncation=True,
       return_tensors="pt")

    inp2 =  t2['input_ids'].to(device)
    mask2 = t2['attention_mask'].unsqueeze(0).to(device)

    out = model(inp2, mask2)
    out = out.cpu().detach().numpy().reshape(-1)
    out = out/out.sum() * 100
    res = [(l, o) for l, o in zip (list(labels.keys()), out.tolist())]   
    return res   
    
model = build_model()
    
st.markdown("### Privet, mir!")
st.markdown("<img width=200px src='https://i.pinimg.com/736x/11/33/19/113319f0ffe91f4bb0f468914b9916da.jpg'>", unsafe_allow_html=True)

text = st.text_area("ENTER TEXT HERE")
start_time = time()
st.markdown("INFERENCE STARTS ...")
    

res = inference(text, mode=None)
res.sort(key = lambda x : - x[1])
st.markdown("<b>INFERENCE RESULT:</b>")    
for lbl, score in res:
    if score >=1:
        st.markdown(f"[        {lbl:<7}] {labels_decoder[lbl]:<35}  {score:.1f}%")

res_plot = []
total=0
for r in res:
    if total < 95:
        res_plot.append(r)
        total += r[1]
    else:
        break

fig, ax = plt.subplots(figsize=(10, len(res_plot)+1))
for r in res_plot :
    ax.barh(r[0], r[1])
st.pyplot(fig)

st.markdown(f"cycle time = {time() - start_time:.2f} s.")