Spaces:
Sleeping
Sleeping
| """Multilingual zero-shot classification: python zeroshot_app.py.""" | |
| from pathlib import Path | |
| from threading import Lock | |
| from string import Formatter | |
| from time import perf_counter | |
| import json | |
| import math | |
| import tempfile | |
| import gradio as gr | |
| import pandas as pd | |
| MODEL_ID = 'MoritzLaurer/mDeBERTa-v3-base-mnli-xnli' | |
| LOCK = Lock() | |
| CLASSIFIER = None | |
| def validate(text, labels, template): | |
| text = (text or '').strip() | |
| names = [line.strip() for line in (labels or '').splitlines() if line.strip()] | |
| if not text or len(text) > 10000: | |
| raise gr.Error('Enter text between 1 and 10,000 characters.') | |
| if not 2 <= len(names) <= 20 or any(len(n)>100 for n in names): | |
| raise gr.Error('Enter 2–20 labels, one per line, each at most 100 characters.') | |
| if len({n.casefold() for n in names}) != len(names): | |
| raise gr.Error('Labels must be distinct, ignoring case.') | |
| try: | |
| fields = [(field,spec,conversion) for _,field,spec,conversion in Formatter().parse(template) if field is not None] | |
| if fields != [('', '', None)]: | |
| raise ValueError() | |
| template.format(names[0]) | |
| except (ValueError, KeyError, IndexError, TypeError): | |
| raise gr.Error('Template must contain exactly one plain {} placeholder.') | |
| return text, names | |
| def classify(text, labels, mode, threshold, template): | |
| global CLASSIFIER | |
| text, names = validate(text, labels, template) | |
| start = perf_counter() | |
| try: | |
| with LOCK: | |
| import torch | |
| if CLASSIFIER is None: | |
| from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline | |
| device = 'cuda' if torch.cuda.is_available() else 'cpu' | |
| tokenizer = AutoTokenizer.from_pretrained(MODEL_ID) | |
| model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID, torch_dtype=torch.float32).to(device).eval() | |
| CLASSIFIER = pipeline('zero-shot-classification', model=model, tokenizer=tokenizer, device=device) | |
| loading = perf_counter()-start | |
| for name in names: | |
| size = len(CLASSIFIER.tokenizer(text, template.format(name), truncation=False)['input_ids']) | |
| if size > 512: | |
| raise gr.Error(f'Text plus hypothesis uses {size} tokens; maximum is 512. Shorten the text, label, or template.') | |
| start = perf_counter() | |
| with torch.inference_mode(): | |
| output = CLASSIFIER(text, candidate_labels=names, hypothesis_template=template, multi_label=mode=='Multiple labels') | |
| elapsed = perf_counter()-start | |
| device = str(CLASSIFIER.device) | |
| revision = getattr(CLASSIFIER.model.config, '_commit_hash', None) | |
| ranked = list(zip(output['labels'],map(float,output['scores']))) | |
| if len(ranked)!=len(names) or set(output['labels'])!=set(names) or any(not math.isfinite(s) or not 0<=s<=1 for _,s in ranked): | |
| raise ValueError('Model returned invalid scores.') | |
| ranked.sort(key=lambda row:row[1],reverse=True) | |
| if mode == 'Single label' and not math.isclose(sum(s for _,s in ranked),1,abs_tol=1e-4): | |
| raise ValueError('Single-label scores are not normalized.') | |
| selected = [ranked[0][0]] if mode=='Single label' else [n for n,s in ranked if s>=threshold] | |
| frame = pd.DataFrame([{'Label':n,'Score':s,'Selected':n in selected} for n,s in ranked]) | |
| report = dict(text=text, candidate_labels=names, template=template, mode=mode, threshold=threshold, scores=dict(ranked), selected_labels=selected, model=MODEL_ID, revision=revision, device=device, loading_seconds=loading, inference_seconds=elapsed) | |
| with tempfile.NamedTemporaryFile(mode='w',suffix='.json',delete=False,encoding='utf-8') as file: | |
| json.dump(report,file,ensure_ascii=False,indent=2) | |
| destination=file.name | |
| return ('Selected: '+(', '.join(selected) if selected else 'No labels meet the threshold.')+f'\n\n{device} · Loading {loading:.2f}s · Inference {elapsed:.2f}s',frame,frame,destination) | |
| except gr.Error: | |
| raise | |
| except Exception as exc: | |
| raise gr.Error(f'Classification failed. Install requirements-zeroshot.txt, check download access and available memory: {exc}') from exc | |
| with gr.Blocks(title='Zero-Shot Text Science Lab',delete_cache=(3600,86400)) as demo: | |
| gr.Markdown('# Zero-Shot Text Science Lab\nMultilingual classification with user-defined labels.',elem_id='lab-header') | |
| gr.Markdown('No task-specific training examples are needed. Scores are not calibrated measures of correctness. Language performance varies.',elem_classes=['lab-note']) | |
| with gr.Row(): | |
| with gr.Column(elem_id='lab-controls'): | |
| text=gr.Textbox(label='Text to classify',lines=8) | |
| labels=gr.Textbox(label='Candidate labels — one per line',lines=6,value='biology\nchemistry\nphysics') | |
| mode=gr.Radio(['Single label','Multiple labels'],value='Single label',label='Classification mode') | |
| threshold=gr.Slider(0,1,value=.5,step=.01,label='Multi-label threshold',interactive=False) | |
| with gr.Accordion('Hypothesis template',open=False): | |
| template=gr.Textbox(value='This example is {}.',label='Exactly one {} placeholder; translated templates are supported') | |
| button=gr.Button('Classify text',variant='primary') | |
| with gr.Column(elem_id='lab-results'): | |
| status=gr.Markdown('Enter text and candidate labels.') | |
| table=gr.Dataframe(label='Ranked label scores',interactive=False) | |
| chart=gr.BarPlot(x='Score',y='Label',title='Label scores',x_lim=[0,1]) | |
| download=gr.File(label='Download full results as JSON') | |
| gr.Examples([ | |
| ['We measured how temperature changes the rate of a chemical reaction.','biology\nchemistry\nphysics'], | |
| ['Wir untersuchen, wie Pflanzen auf Trockenheit reagieren.','Biologie\nChemie\nPhysik'], | |
| ],inputs=[text,labels]) | |
| gr.Markdown('Single-label scores sum to one across your candidates. Multi-label scores are independent and need not sum to one.\n\n### Try it\nChange label wording, add a competing label, or compare single-label and multi-label modes.\n\n[Model card and language limitations](https://huggingface.co/'+MODEL_ID+') · MIT model license. Weights download on first use; inference runs on this server.') | |
| inputs=[text,labels,mode,threshold,template] | |
| outputs=[status,table,chart,download] | |
| for control in inputs: | |
| control.change(lambda:('Inputs changed. Classify again.',None,None,None),outputs=outputs) | |
| mode.change(lambda value:gr.Slider(interactive=value=='Multiple labels'),mode,threshold) | |
| # Keep controls locked until completion, including failures, so stale runs cannot overwrite edits. | |
| event=button.click(lambda:[gr.update(interactive=False) for _ in range(6)],outputs=[*inputs,button],queue=False) | |
| event=event.then(classify,inputs,outputs,concurrency_id='text-model',concurrency_limit=1) | |
| event.then(lambda value:[gr.update(interactive=True),gr.update(interactive=True),gr.update(interactive=True),gr.update(interactive=value=='Multiple labels'),gr.update(interactive=True),gr.update(interactive=True)],mode,[*inputs,button]) | |
| demo.queue(max_size=20,default_concurrency_limit=1) | |
| if __name__=='__main__': | |
| demo.launch(theme=gr.themes.Base(primary_hue='blue',neutral_hue='slate'),css_paths=[Path(__file__).with_name('styles.css')]) | |