import random from mtranslate import translate import streamlit as st import seaborn as sns from spacy import displacy from transformers import ( AutoConfig, AutoTokenizer, AutoModelForMaskedLM, AutoModelForSequenceClassification, AutoModelForTokenClassification, pipeline ) LOGO = "https://huggingface.co/bertin-project/bertin-roberta-base-spanish/resolve/main/images/bertin.png" WRAPPER = """
{}
""" PROMPT_LIST = [ "Fui a la librería a comprar un .", "¡Qué buen hace hoy!", "Hoy empiezan las vacaciones así que vamos a la .", "Mi color favorito es el .", "Voy a porque estoy muy cansada.", "Mañana vienen mis amigos de .", "¿Te apetece venir a conmigo?", "En verano hace mucho .", "En el bosque había .", "El ministro dijo que los impuestos.", "Si no estuviera afónica, esa canción.", "Parece que ha salido el , por eso hace tanto calor.", "Al pan, pan, y al vino, .", ] PAWS_X_PROMPT_LIST = [ "Te amo.Te adoro.", "Te amo.Te detesto.", "Te amo.Voy a caminar al campo." ] @st.cache(show_spinner=False, persist=True) def load_model(masked_text, model_url): model = AutoModelForMaskedLM.from_pretrained(model_url) tokenizer = AutoTokenizer.from_pretrained(model_url) nlp = pipeline("fill-mask", model=model, tokenizer=tokenizer) result = nlp(masked_text) return result @st.cache(show_spinner=False, persist=True) def load_model_pair_classification(text, model_url_pair_classification): model = AutoModelForSequenceClassification.from_pretrained(model_url_pair_classification) tokenizer = AutoTokenizer.from_pretrained(model_url) nlp = pipeline("text-classification", model=model, tokenizer=tokenizer) result = nlp(f"{text}") if result[0]["label"] == "entailment": return f"Entailment: {result[0]['score']:02f}" if result[0]["label"] == "neutral": return f"Neutral: {result[0]['score']:02f}" return f"Contradiction: {result[0]['score']:02f}" # Page st.set_page_config(page_title="BERTIN Demo", page_icon=LOGO) st.title("BERTIN") #Sidebar st.sidebar.markdown(f"""
# BERTIN
BERTIN is a series of BERT-based models for Spanish. The models are trained with Flax and using TPUs sponsored by Google since this is part of the [Flax/Jax Community Week](https://discuss.huggingface.co/t/open-to-the-community-community-week-using-jax-flax-for-nlp-cv/7104) organised by HuggingFace. Please read our [full report](https://huggingface.co/bertin-project/bertin-roberta-base-spanish) for more details on the methodology and metrics on downstream tasks. """, unsafe_allow_html=True) # Body st.markdown( """ All models are variations of **RoBERTa-base** trained from scratch in **Spanish** using a sample from the **mc4 dataset**. We reduced the dataset size to 50 million documents to keep training times shorter, and also to be able to bias training examples based on their perplexity. The idea is to favour examples with perplexities that are neither too small (short, repetitive texts) or too long (potentially poor quality). There are three versions of the sampling procedure (producing three different series of models): * **Random** sampling is the control baseline and simply takes documents at random with uniform probability to reduce the dataset size. * **Gaussian** rejects documents with higher probability for lower and larger perplexities, based on weighting the perplexity distribution with a Gaussian function. * **Stepwise** applies different four sampling probabilities to each of the four quartiles of the perplexity distribution. The first models have been trained (250.000 steps) on sequence length 128, and then training for Gaussian changed to sequence length 512 for the last 25.000 training steps to yield another version. Please read our [full report](https://huggingface.co/bertin-project/bertin-roberta-base-spanish) for more details on the methodology and metrics on downstream tasks. ### Masked language modeling Here you can play with the filling the mask objective of all the models. """ ) col1, col2, col3 = st.columns(3) strategy = col1.selectbox("Sampling strategy", ["Gaussian", "Stepwise", "Random"]) seq_len = col2.selectbox("Sequence length", [128, 512]) if seq_len == 128: model_url = f"bertin-project/bertin-base-{str(strategy).lower()}" else: model_url = f"bertin-project/bertin-base-{str(strategy).lower()}-exp-512seqlen" prompt = col3.selectbox("Prompt", ["Random", "Custom"]) if prompt == "Custom": prompt_box = "Enter your masked text here..." else: prompt_box = random.choice(PROMPT_LIST) text = st.text_area("Enter text", prompt_box) if st.button("Fill the mask"): with st.spinner(text="Filling the mask..."): st.subheader("Result") result = load_model(text, model_url) result_sequence = result[0]["sequence"] st.write(result_sequence) st.write("_English_ _translation:_", translate(result_sequence, "en", "es")) st.write(result) def make_color_palette(labels): color_palette = sns.color_palette(n_colors=len(labels)) color_map = {x: rgb2hex(*y) for x, y in zip(labels, color_palette)} return color_map @st.cache(allow_output_mutation=True) def get_colormap(labels): color_map = make_color_palette(labels) return color_map def add_colormap(labels): color_map = get_colormap(labels) for label in labels: if label not in color_map: rand_color = "#"+"%06x" % random.randint(0, 0xFFFFFF) color_map[label]=rand_color return color_map def load_model_ner(model_url): config = AutoConfig.from_pretrained(model_url) model = AutoModelForTokenClassification.from_pretrained( model_url, config=config ) tokenizer = AutoTokenizer.from_pretrained(model_url, use_fast=True) return pipeline( "ner", model=model, tokenizer=tokenizer, ignore_labels=[], aggregation_strategy="simple", ) def display(entities): doc = model_entities_to_displacy_format(entities, ignore_entities=["O"]) labels = list(set([ent["label"] for ent in doc["ents"]])) color_map = add_colormap(labels) html = displacy.render( doc, manual=True, style="ent", options={"colors": color_map} ) html = html.replace("\n", " ") st.write(WRAPPER.format(html), unsafe_allow_html=True) def rgb2hex(r, g, b): return "#{:02x}{:02x}{:02x}".format( int(r * 255), int(g * 255), int(b * 255) ) def model_entities_to_displacy_format(ents, ignore_entities=[]): s_ents = {} s_ents["text"] = " ".join([e["word"] for e in ents]) spacy_ents = [] start_pointer = 0 if isinstance(ents, list) and "entity_group" in ents[0]: entity_key = "entity_group" else: entity_key = "entity" for i, ent in enumerate(ents): if ent[entity_key] not in ignore_entities: spacy_ents.append({ "start": start_pointer, "end": start_pointer + len(ent["word"]), "label": ent[entity_key], }) start_pointer = start_pointer + len(ent["word"]) + 1 s_ents["ents"] = spacy_ents s_ents["title"] = None return s_ents st.markdown(""" ### Fine-tuning to CoNLL 2002 es for Named Entity Recognition (NER) Here you can play with the RoBERTa Base Gaussian Seq Len 512 model fine-tuned to conll2002-es. """) text_input = str(st.text_input( "Text", "Mi nombre es Íñigo Montoya. Viajo a Los Acantilados de la Locura " )) ner_model_url = "bertin-project/bertin-base-ner-conll2002-es" label2id = AutoConfig.from_pretrained(ner_model_url, cache=False).label2id color_map = get_colormap(list(label2id.keys())) if st.button("Recognize named entities"): with st.spinner(text="Recognizing named entities..."): ner = load_model_ner(ner_model_url) entities = ner(str(text_input)) st.write("_English_ _translation:_", translate(str(text_input), "en", "es")) if entities: if isinstance(entities, dict) and "error" in entities: st.write(entities) else: display(entities) raw_entities = [] for entity in entities: raw_entity = entity raw_entity["start"] = int(raw_entity["start"]) raw_entity["end"] = int(raw_entity["end"]) raw_entity["score"] = float(raw_entity["score"]) raw_entities.append(raw_entity) st.write(raw_entities) else: st.write("No entities found") st.markdown( """ ### Fine-tuning to XNLI Here you can play with the RoBERTa Base Gaussian Seq Len 512 model fine-tuned to XNLI. """ ) pawsx_model_url = "bertin-project/bertin-base-xnli-es" paraphrase_prompt = st.selectbox("Paraphrase Prompt", ["Random", "Custom"]) if paraphrase_prompt == "Custom": paraphrase_prompt_box = "Enter two sentences separated by here..." else: paraphrase_prompt_box = random.choice(PAWS_X_PROMPT_LIST) text = st.text_area("Enter text", paraphrase_prompt_box) if st.button("Clasify paraphrasing"): with st.spinner(text="Clasifying paraphrasing..."): st.subheader("Classification result") paraphrase_score = load_model_pair_classification(text, pawsx_model_url) st.write("_English_ _translation:_", translate(text, "en", "es")) st.write(paraphrase_score) st.markdown( """ ### Team members - Eduardo González ([edugp](https://huggingface.co/edugp)) - Javier de la Rosa ([versae](https://huggingface.co/versae)) - Manu Romero ([mrm8488](https://huggingface.co/mrm8488)) - María Grandury ([mariagrandury](https://huggingface.co/mariagrandury)) - Pablo González de Prado ([Pablogps](https://huggingface.co/Pablogps)) - Paulo Villegas ([paulo](https://huggingface.co/paulo)) ### More information You can find more information about these models [here](https://huggingface.co/bertin-project/bertin-roberta-base-spanish). """ )