app.py

import datetime
import gradio as gr
import fasttext, torch, clip
from sentence_transformers import SentenceTransformer, util
 
model_en, _ = clip.load("ViT-B/32")
model_multi = SentenceTransformer("sentence-transformers/clip-ViT-B-32-multilingual-v1")

def prep_examples():
    example_text1 = "Coronavirus disease (COVID-19) is an infectious disease caused by the SARS-CoV-2 virus. Most \
    people who fall sick with COVID-19 will experience mild to moderate symptoms and recover without special treatment. \
    However, some will become seriously ill and require medical attention."
    example_labels1 = "business;;health related;;politics;;climate change"

    example_text2 = "Elephants are"
    example_labels2 = "big;;small;;strong;;fast;;carnivorous"

    example_text3 = "Elephants"
    example_labels3 = "are big;;can be very small;;generally not strong enough;;are faster than you think"

    example_text4 = "Dogs are man's best friend"
    example_labels4 = "positive;;negative;;neutral"

    example_text5 = "Şampiyonlar Ligi’nde 5. hafta oynanan karşılaşmaların ardından sona erdi. Real Madrid, \
    Inter ve Sporting oynadıkları mücadeleler sonrasında Son 16 turuna yükselmeyi başardı. \
    Gecenin dev mücadelesinde ise Manchester City, PSG’yi yenerek liderliği garantiledi."
    example_labels5 = "dünya;;ekonomi;;kültür;;siyaset;;spor;;teknoloji"

    example_text6 = "Letzte Woche gab es einen Selbstmord in einer nahe gelegenen kolonie"
    example_labels6 = "verbrechen;;tragödie;;stehlen"

    example_text7 = "El autor se perfila, a los 50 años de su muerte, como uno de los grandes de su siglo"
    example_labels7 = "cultura;;sociedad;;economia;;salud;;deportes"

    example_text8 = "Россия в среду заявила, что военные учения в аннексированном Москвой Крыму закончились \
    и что солдаты возвращаются в свои гарнизоны, на следующий день после того, как она объявила о первом выводе \
    войск от границ Украины."
    example_labels8 = "новости;;комедия"

    example_text9 = "I quattro registi - Federico Fellini, Pier Paolo Pasolini, Bernardo Bertolucci e Vittorio De Sica - \
    hanno utilizzato stili di ripresa diversi, ma hanno fortemente influenzato le giovani generazioni di registi."
    example_labels9 = "cinema;;politica;;cibo"

    example_text10 = "Ja, vi elsker dette landet,\
    som det stiger frem,\
    furet, værbitt over vannet,\
    med de tusen hjem.\
    Og som fedres kamp har hevet\
    det av nød til seir"
    example_labels10 = "helse;;sport;;religion;;mat;;patriotisme og nasjonalisme"

    example_text11 = "Amar sonar bangla ami tomay bhalobasi"
    example_labels11 = "bhalo;;kharap"

    examples = [
        [example_text1, example_labels1],
        [example_text2, example_labels2],
        [example_text3, example_labels3],
        [example_text4, example_labels4],
        [example_text5, example_labels5],
        [example_text6, example_labels6],
        [example_text7, example_labels7],
        [example_text8, example_labels8],
        [example_text9, example_labels9],
        [example_text10, example_labels10],
        [example_text11, example_labels11]]

    return examples

def detect_lang(text):
    seq_lang = 'en'
    
    text = text.replace('\n', ' ')

    try:
        seq_lang = fasttext_model.predict(text, k=1)[0][0].split("__label__")[1]
    except:
        print("Language detection failed!",
              "Date:{}, Sequence:{}".format(
			  str(datetime.datetime.now()),
			  text))

    return seq_lang

def sequence_to_classify(text, labels):
	lang = detect_lang(text)
	if lang == 'en':
		model = model_en
		hypothesis_template = "This example is {}."
	else:
		model = model_multi
		hypothesis_template = "{}."
	
	labels = [hypothesis_template.format(label) for label in labels.split(";;")]
	
	if str(type(model)) == "<class 'clip.model.CLIP'>":
		text_tokens = clip.tokenize(text)
		text_features = model.encode_text(text_tokens)
		
		label_tokens = clip.tokenize(labels)
		labels_features = model.encode_text(label_tokens)
	else:    
		text_features = torch.tensor(model.encode(text))
		labels_features = torch.tensor(self.model.encode(labels))
		
	sim_scores = util.cos_sim(text_features, labels_features)
	preds = []
	for textlet, sim_score in zip([text], sim_scores):
		out = []
		pred = {}
		for raw_score in sim_score:
			out.append(raw_score.item() * 100)
		probs = torch.tensor([out])
		probs = probs.softmax(dim=-1).cpu().numpy()
		scores = list(probs.flatten())

		sorted_sl = sorted(zip(scores, labels), key=lambda t:t[0], reverse=True)  

		pred["text"] = textlet
		pred["scores"], pred["labels"] = zip(*sorted_sl)
		preds.append(pred)
	
	print(preds)
	predicted_labels = preds['labels']
	predicted_scores = preds['scores']
	clean_output = {idx: float(predicted_scores.pop(0)) for idx in predicted_labels}
	print("Date:{}, Sequence:{}, Labels: {}".format(
		str(datetime.datetime.now()),
		text,
		predicted_labels))

	return clean_output

iface = gr.Interface(
    title="Alternate Zero-shot Multi-label Multilingual NLP Classifier",
    description="Work in progress.",
    fn=sequence_to_classify,
    inputs=[gr.inputs.Textbox(lines=10,
        label="Please enter the text you would like to classify...",
        placeholder="Text here..."),
        gr.inputs.Textbox(lines=2,
        label="Please enter the candidate labels (separated by 2 consecutive semicolons)...",
        placeholder="Labels here separated by ;;")],
    outputs=gr.outputs.Label(num_top_classes=5),
    #interpretation="default",
    examples=prep_examples())

iface.launch()