app.py

import streamlit as st

if not hasattr(st, "cache_resource"):
    st.cache_resource = st.experimental_singleton

import torch
import torch.nn as nn
import torch.nn.functional as F
import transformers
import pandas as pd

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

from transformers import MarianMTModel, MarianTokenizer

model_options = [
    'Helsinki-NLP/opus-mt-roa-en',
    'Helsinki-NLP/opus-mt-en-roa',
]

col1, col2 = st.columns(2)

with col1:
    model_name = st.selectbox("Select a model", model_options + ['other'])

    if model_name == 'other':
        model_name = st.text_input("Enter model name", model_options[0])

@st.cache_resource
def get_tokenizer(model_name):
    return MarianTokenizer.from_pretrained(model_name)

@st.cache_resource
def get_model(model_name):
    model = MarianMTModel.from_pretrained(model_name).to(device)
    print(f"Loaded model, {model.num_parameters():,d} parameters.")
    return model

tokenizer = get_tokenizer(model_name)
model = get_model(model_name)

if tokenizer.supported_language_codes:
    lang_code = st.selectbox("Select a language", tokenizer.supported_language_codes)
else:
    lang_code = None


with col2:
    input_text = st.text_input("Enter text to translate", "Hola, mi nombre es Juan")

input_text = input_text.strip()
if not input_text:
    st.stop()

# prepend the language code if necessary
if lang_code:
    input_text = f"{lang_code} {input_text}"


input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to(device)

example_generations = model.generate(
    input_ids,
    num_beams=4,
    num_return_sequences=4,
)

col1, col2 = st.columns(2)
with col1:
    st.write("Example generations:")
    st.write('\n'.join(
        '- ' + translation
        for translation in tokenizer.batch_decode(example_generations, skip_special_tokens=True)))

with col2:
    example_first_word = tokenizer.decode(example_generations[0, 1])
    output_so_far = st.text_input("Enter text translated so far", example_first_word)


# tokenize the output so far
with tokenizer.as_target_tokenizer():
    output_tokens = tokenizer.tokenize(output_so_far)
    decoder_input_ids = tokenizer.convert_tokens_to_ids(output_tokens)

# Add the start token
decoder_input_ids = [model.config.decoder_start_token_id] + decoder_input_ids

with torch.no_grad():
    model_output = model(
        input_ids = input_ids,
        decoder_input_ids = torch.tensor([decoder_input_ids]).to(device))

last_token_logits = model_output.logits[0, -1].cpu()
assert len(last_token_logits.shape) == 1
most_likely_tokens = last_token_logits.topk(k=20)

probs = last_token_logits.softmax(dim=-1)
probs_for_likely_tokens = probs[most_likely_tokens.indices]

with tokenizer.as_target_tokenizer():
    probs_table = pd.DataFrame({
        'token': [tokenizer.decode(token_id) for token_id in most_likely_tokens.indices],
        'id': most_likely_tokens.indices,
        'probability': probs_for_likely_tokens,
        'logprob': probs_for_likely_tokens.log(),
        'cumulative probability': probs_for_likely_tokens.cumsum(0)
    })

st.subheader("Most likely next tokens")
st.table(probs_table.style.hide(axis='index'))

if len(decoder_input_ids) > 1:
    st.subheader("Loss by already-generated token")
    loss_table = pd.DataFrame({
        'token': [tokenizer.decode(token_id) for token_id in decoder_input_ids[1:]],
        'loss': F.cross_entropy(model_output.logits[0, :-1], torch.tensor(decoder_input_ids[1:]).to(device), reduction='none').cpu()
    })
    st.write(loss_table)
    st.write("Total loss so far:", loss_table.loss.sum())