app.py

import gradio as gr
import yaml
from joeynmt.prediction import load_params_for_prediction,translate
from huggingface_hub import hf_hub_download



language_map = {'English':'en','Swahili':'sw','Fon':'fon','Igbo':'ig',
                'Arabic':'ar','Shona':'sn','Ẹ̀dó':'bin','Hausa':'ha',
                'Efik':'efi','Twi':'twi','Afrikaans':'af','Yoruba':'yo'}  
                
#List of available languages I worked on.
available_language_pairs =['en-sw','en-af','en-ar','efi-en','en-ha','en-ig','en-fon','en-twi','sn-en','sw-en','yo-en']


def load_config(path="configs/default.yaml") -> dict:
    """
    CODE ADAPTED FROM: https://github.com/joeynmt/joeynmt
    Loads and parses a YAML configuration file.

    :param path: path to YAML configuration file
    :return: configuration dictionary
    """
    with open(path, 'r', encoding="utf-8") as ymlfile:
      
        cfg = yaml.safe_load(ymlfile)
    return cfg
  
def load_model(source_language,target_language):  
    #source_language = language_map[source_language_]
    #target_language = language_map[target_language_]
    
    #source_language = 'en'
    #target_language = 'sw'
    translation_dir = 'main'

    try:
      file_yaml = hf_hub_download("chrisjay/masakhane_benchmarks", filename=f"{source_language}-{target_language}/{translation_dir}/config.yaml",force_filename='config.yaml')
      src_vocab = hf_hub_download("chrisjay/masakhane_benchmarks", filename=f"{source_language}-{target_language}/{translation_dir}/src_vocab.txt")
      trg_vocab  = hf_hub_download("chrisjay/masakhane_benchmarks", filename=f"{source_language}-{target_language}/{translation_dir}/trg_vocab.txt")
      best_ckpt = hf_hub_download("chrisjay/masakhane_benchmarks", filename=f"{source_language}-{target_language}/{translation_dir}/best.ckpt")
    except Exception:
      raise Exception(f'It seems we do not have a working configuration repo yet for {source_language} -> {target_language}. \n You could help us by creating it here: https://huggingface.co/chrisjay/masakhane_benchmarks/tree/main')


    parsed_yaml_file = load_config(file_yaml)
    parsed_yaml_file['data']['src_vocab']=src_vocab
    parsed_yaml_file['data']['trg_vocab']=trg_vocab

    params = load_params_for_prediction(parsed_yaml_file,best_ckpt)
    return params

#Load models of all available language pairs
examples_available_models=[]
model_mapping = {} 
for availabe_lang in available_language_pairs:
    try:
        model_mapping.update({availabe_lang:load_model(availabe_lang.split('-')[0],availabe_lang.split('-')[1])})
        examples_available_models.append([f"{list(language_map.keys())[list(language_map.values()).index(availabe_lang.split('-')[0])]}",f"{list(language_map.keys())[list(language_map.values()).index(availabe_lang.split('-')[1])]}"]) #idea to extract key from value got from https://stackoverflow.com/questions/8023306/get-key-by-value-in-dictionary
    except Exception:
        continue

if examples_available_models==[]:
    raise Exception(f'Available models for Space cannot be empty!')


def get_translation(source_language,target_language,source_sentence=None,source_file=None):
    '''
    This takes a sentence and gets the translation.
    type_=2 tells joeynmt translate that it should expect a sentence. 
    '''
    
    source_language_ = language_map[source_language]
    target_language_ = language_map[target_language]
   
    
    source = source_sentence
    type_=2
    if source_file!=None:
        type_=1
        source = source_file.name
    try:
       
        #params = load_model(source_language,target_language)
        params = model_mapping[f'{source_language_}-{target_language_}']
        pred = translate(params,source,type_)
    except Exception:
        return f'There was an issue loading the translation model for {source_language} -> {target_language}. Try another pair please'
    
    return pred[0] if source_file==None else pred
 



title = "Interact with Masakhane Benchmark Models"
description = "This enables you to interact with some of the Masakhane Benchmark Models and keep up with their improvement. Some of these models undergo finetuning on a regular basis. This way, you can easily use the best model with no hassles."

iface = gr.Interface(fn=get_translation, 
  inputs=[gr.inputs.Dropdown(choices = available_languages,default='English'),
  gr.inputs.Dropdown(choices = available_languages,default='Swahili'),
  gr.inputs.Textbox(label="Input"),
  gr.inputs.File(file_count="single", type="file", label='Or upload txt file containing sentences', optional=True)],
  outputs=gr.outputs.Textbox(type="auto", label='Translation'),
  title=title,
  description=description,
  examples=examples_available_models,
  enable_queue=True,
  theme='huggingface')
iface.launch()