main_dl.py

from fastapi import FastAPI, HTTPException, Header, Depends, Request
from fastapi.responses import JSONResponse
from fastapi.security import HTTPBasic, HTTPBasicCredentials
from fastapi.exceptions import RequestValidationError
from typing import Optional, List
from pydantic import BaseModel, ValidationError
import pandas as pd
import numpy as np
import os
from filesplit.merge import Merge
import tensorflow as tf
import string
import re
from tensorflow import keras
from keras_nlp.layers import TransformerEncoder
from tensorflow.keras import layers
from tensorflow.keras.utils import plot_model

api = FastAPI()
dataPath = "data"

# ===== Keras ====
strip_chars = string.punctuation + "¿"
strip_chars = strip_chars.replace("[", "")
strip_chars = strip_chars.replace("]", "")

def custom_standardization(input_string):
    lowercase = tf.strings.lower(input_string)
    lowercase=tf.strings.regex_replace(lowercase, "[à]", "a")
    return tf.strings.regex_replace(
        lowercase, f"[{re.escape(strip_chars)}]", "")

def load_vocab(file_path):
    with open(file_path, "r",  encoding="utf-8") as file:
        return file.read().split('\n')[:-1]


def decode_sequence_rnn(input_sentence, src, tgt):
    global translation_model

    vocab_size = 15000
    sequence_length = 50

    source_vectorization = layers.TextVectorization(
        max_tokens=vocab_size,
        output_mode="int",
        output_sequence_length=sequence_length,
        standardize=custom_standardization,
        vocabulary = load_vocab(dataPath+"/vocab_"+src+".txt"),
    )

    target_vectorization = layers.TextVectorization(
        max_tokens=vocab_size,
        output_mode="int",
        output_sequence_length=sequence_length + 1,
        standardize=custom_standardization,
        vocabulary = load_vocab(dataPath+"/vocab_"+tgt+".txt"),
    )

    tgt_vocab = target_vectorization.get_vocabulary()
    tgt_index_lookup = dict(zip(range(len(tgt_vocab)), tgt_vocab))
    max_decoded_sentence_length = 50
    tokenized_input_sentence = source_vectorization([input_sentence])
    decoded_sentence = "[start]"
    for i in range(max_decoded_sentence_length):
        tokenized_target_sentence = target_vectorization([decoded_sentence])
        next_token_predictions = translation_model.predict(
            [tokenized_input_sentence, tokenized_target_sentence], verbose=0)
        sampled_token_index = np.argmax(next_token_predictions[0, i, :])
        sampled_token = tgt_index_lookup[sampled_token_index]
        decoded_sentence += " " + sampled_token
        if sampled_token == "[end]":
            break
    return decoded_sentence[8:-6]

# ===== Enf of Keras ====

# ===== Transformer section ====

class TransformerDecoder(layers.Layer):
    def __init__(self, embed_dim, dense_dim, num_heads, **kwargs):
        super().__init__(**kwargs)
        self.embed_dim = embed_dim
        self.dense_dim = dense_dim
        self.num_heads = num_heads
        self.attention_1 = layers.MultiHeadAttention(
            num_heads=num_heads, key_dim=embed_dim)
        self.attention_2 = layers.MultiHeadAttention(
            num_heads=num_heads, key_dim=embed_dim)
        self.dense_proj = keras.Sequential(
            [layers.Dense(dense_dim, activation="relu"),
             layers.Dense(embed_dim),]
        )
        self.layernorm_1 = layers.LayerNormalization()
        self.layernorm_2 = layers.LayerNormalization()
        self.layernorm_3 = layers.LayerNormalization()
        self.supports_masking = True

    def get_config(self):
        config = super().get_config()
        config.update({
            "embed_dim": self.embed_dim,
            "num_heads": self.num_heads,
            "dense_dim": self.dense_dim,
        })
        return config

    def get_causal_attention_mask(self, inputs):
        input_shape = tf.shape(inputs)
        batch_size, sequence_length = input_shape[0], input_shape[1]
        i = tf.range(sequence_length)[:, tf.newaxis]
        j = tf.range(sequence_length)
        mask = tf.cast(i >= j, dtype="int32")
        mask = tf.reshape(mask, (1, input_shape[1], input_shape[1]))
        mult = tf.concat(
            [tf.expand_dims(batch_size, -1),
             tf.constant([1, 1], dtype=tf.int32)], axis=0)
        return tf.tile(mask, mult)

    def call(self, inputs, encoder_outputs, mask=None):
        causal_mask = self.get_causal_attention_mask(inputs)
        if mask is not None:
            padding_mask = tf.cast(
                mask[:, tf.newaxis, :], dtype="int32")
            padding_mask = tf.minimum(padding_mask, causal_mask)
        else:
            padding_mask = mask
        attention_output_1 = self.attention_1(
            query=inputs,
            value=inputs,
            key=inputs,
            attention_mask=causal_mask)
        attention_output_1 = self.layernorm_1(inputs + attention_output_1)
        attention_output_2 = self.attention_2(
            query=attention_output_1,
            value=encoder_outputs,
            key=encoder_outputs,
            attention_mask=padding_mask,
        )
        attention_output_2 = self.layernorm_2(
            attention_output_1 + attention_output_2)
        proj_output = self.dense_proj(attention_output_2)
        return self.layernorm_3(attention_output_2 + proj_output)
    
class PositionalEmbedding(layers.Layer):
    def __init__(self, sequence_length, input_dim, output_dim, **kwargs):
        super().__init__(**kwargs)
        self.token_embeddings = layers.Embedding(
            input_dim=input_dim, output_dim=output_dim)
        self.position_embeddings = layers.Embedding(
            input_dim=sequence_length, output_dim=output_dim)
        self.sequence_length = sequence_length
        self.input_dim = input_dim
        self.output_dim = output_dim

    def call(self, inputs):
        length = tf.shape(inputs)[-1]
        positions = tf.range(start=0, limit=length, delta=1)
        embedded_tokens = self.token_embeddings(inputs)
        embedded_positions = self.position_embeddings(positions)
        return embedded_tokens + embedded_positions

    def compute_mask(self, inputs, mask=None):
        return tf.math.not_equal(inputs, 0)

    def get_config(self):
        config = super(PositionalEmbedding, self).get_config()
        config.update({
            "output_dim": self.output_dim,
            "sequence_length": self.sequence_length,
            "input_dim": self.input_dim,
        })
        return config
    
def decode_sequence_tranf(input_sentence, src, tgt):
    global translation_model

    vocab_size = 15000
    sequence_length = 50

    source_vectorization = layers.TextVectorization(
        max_tokens=vocab_size,
        output_mode="int",
        output_sequence_length=sequence_length,
        standardize=custom_standardization,
        vocabulary = load_vocab(dataPath+"/vocab_"+src+".txt"),
    )

    target_vectorization = layers.TextVectorization(
        max_tokens=vocab_size,
        output_mode="int",
        output_sequence_length=sequence_length + 1,
        standardize=custom_standardization,
        vocabulary = load_vocab(dataPath+"/vocab_"+tgt+".txt"),
    )

    tgt_vocab = target_vectorization.get_vocabulary()
    tgt_index_lookup = dict(zip(range(len(tgt_vocab)), tgt_vocab))
    max_decoded_sentence_length = 50
    tokenized_input_sentence = source_vectorization([input_sentence])
    decoded_sentence = "[start]"
    for i in range(max_decoded_sentence_length):
        tokenized_target_sentence = target_vectorization(
            [decoded_sentence])[:, :-1]
        predictions = translation_model(
            [tokenized_input_sentence, tokenized_target_sentence])
        sampled_token_index = np.argmax(predictions[0, i, :])
        sampled_token = tgt_index_lookup[sampled_token_index]
        decoded_sentence += " " + sampled_token
        if sampled_token == "[end]":
            break
    return decoded_sentence[8:-6]

# ==== End Transforformer section ====

def load_all_data():
   
    merge = Merge( dataPath+"/rnn_en-fr_split",  dataPath, "seq2seq_rnn-model-en-fr.h5").merge(cleanup=False)
    merge = Merge( dataPath+"/rnn_fr-en_split",  dataPath, "seq2seq_rnn-model-fr-en.h5").merge(cleanup=False)
    rnn_en_fr = keras.models.load_model(dataPath+"/seq2seq_rnn-model-en-fr.h5", compile=False)
    rnn_fr_en = keras.models.load_model(dataPath+"/seq2seq_rnn-model-fr-en.h5", compile=False)
    rnn_en_fr.compile(optimizer="rmsprop", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
    rnn_fr_en.compile(optimizer="rmsprop", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
    
    custom_objects = {"TransformerDecoder": TransformerDecoder, "PositionalEmbedding": PositionalEmbedding}
    with keras.saving.custom_object_scope(custom_objects):
        transformer_en_fr = keras.models.load_model( "data/transformer-model-en-fr.h5")
        transformer_fr_en = keras.models.load_model( "data/transformer-model-fr-en.h5")
    merge = Merge( "data/transf_en-fr_weight_split",  "data", "transformer-model-en-fr.weights.h5").merge(cleanup=False)
    merge = Merge( "data/transf_fr-en_weight_split",  "data", "transformer-model-fr-en.weights.h5").merge(cleanup=False)
    transformer_en_fr.compile(optimizer="rmsprop", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
    transformer_fr_en.compile(optimizer="rmsprop", loss="sparse_categorical_crossentropy", metrics=["accuracy"])

    return rnn_en_fr, rnn_fr_en, transformer_en_fr, transformer_fr_en

rnn_en_fr, rnn_fr_en, transformer_en_fr, transformer_fr_en = load_all_data() 


def display_translation(n1, Lang,model_type):
    global df_data_src, df_data_tgt, placeholder
    
    placeholder = st.empty()
    with st.status(":sunglasses:", expanded=True):
        s = df_data_src.iloc[n1:n1+5][0].tolist()
        s_trad = []
        s_trad_ref = df_data_tgt.iloc[n1:n1+5][0].tolist()
        source = Lang[:2]
        target = Lang[-2:]
        for i in range(3):
            if model_type==1:
                s_trad.append(decode_sequence_rnn(s[i], source, target))
            else:
                s_trad.append(decode_sequence_tranf(s[i], source, target))
            st.write("**"+source+"   :**  :blue["+ s[i]+"]")
            st.write("**"+target+"   :**  "+s_trad[-1])
            st.write("**ref. :** "+s_trad_ref[i])
            st.write("")
    with placeholder:
        st.write("<p style='text-align:center;background-color:red; color:white')>Score Bleu = "+str(int(round(corpus_bleu(s_trad,[s_trad_ref]).score,0)))+"%</p>", \
            unsafe_allow_html=True)
        
     
def find_lang_label(lang_sel):
    global lang_tgt, label_lang
    return label_lang[lang_tgt.index(lang_sel)]

@api.get('/', name="Vérification que l'API fonctionne")
def check_api():
    load_all_data()
    return {'message': "L'API fonctionne"}

@api.get('/small_vocab/rnn', name="Traduction par RNN")
def check_api(lang_tgt:str,
              texte: str):
    global translation_model
    
    if (lang_tgt=='en'):
        translation_model = rnn_fr_en
        return decode_sequence_rnn(texte, "fr", "en")
    else:
        translation_model = rnn_en_fr
        return decode_sequence_rnn(texte, "en", "fr")
    
@api.get('/small_vocab/transformer', name="Traduction par Transformer")
def check_api(lang_tgt:str,
              texte: str):
    global translation_model
    
    if (lang_tgt=='en'):
        translation_model = rnn_fr_en
        return decode_sequence_tranf(texte, "fr", "en")
    else:
        translation_model = rnn_en_fr
        return decode_sequence_tranf(texte, "en", "fr")

'''
def run():

    global n1, df_data_src, df_data_tgt, translation_model, placeholder, model_speech
    global df_data_en, df_data_fr, lang_classifier, translation_en_fr, translation_fr_en
    global lang_tgt, label_lang

    st.write("")
    st.title(tr(title))
    #
    st.write("## **"+tr("Explications")+" :**\n")

    st.markdown(tr(
        """
        Enfin, nous avons réalisé une traduction :red[**Seq2Seq**] ("Sequence-to-Sequence") avec des :red[**réseaux neuronaux**].  
        """)
        , unsafe_allow_html=True)
    st.markdown(tr(
        """
        La traduction Seq2Seq est une méthode d'apprentissage automatique qui permet de traduire des séquences de texte d'une langue à une autre en utilisant 
        un :red[**encodeur**] pour capturer le sens du texte source, un :red[**décodeur**] pour générer la traduction, 
        avec un ou plusieurs :red[**vecteurs d'intégration**] qui relient les deux, afin de transmettre le contexte, l'attention ou la position.  
        """)
        , unsafe_allow_html=True)
    st.image("assets/deepnlp_graph1.png",use_column_width=True)
    st.markdown(tr(
        """      
        Nous avons mis en oeuvre ces techniques avec des Réseaux Neuronaux Récurrents (GRU en particulier) et des Transformers  
        Vous en trouverez :red[**5 illustrations**] ci-dessous.
        """)
    , unsafe_allow_html=True)

    # Utilisation du module translate
    lang_tgt   = ['en','fr','af','ak','sq','de','am','en','ar','hy','as','az','ba','bm','eu','bn','be','my','bs','bg','ks','ca','ny','zh','si','ko','co','ht','hr','da','dz','gd','es','eo','et','ee','fo','fj','fi','fr','fy','gl','cy','lg','ka','el','gn','gu','ha','he','hi','hu','ig','id','iu','ga','is','it','ja','kn','kk','km','ki','rw','ky','rn','ku','lo','la','lv','li','ln','lt','lb','mk','ms','ml','dv','mg','mt','mi','mr','mn','nl','ne','no','nb','nn','oc','or','ug','ur','uz','ps','pa','fa','pl','pt','ro','ru','sm','sg','sa','sc','sr','sn','sd','sk','sl','so','st','su','sv','sw','ss','tg','tl','ty','ta','tt','cs','te','th','bo','ti','to','ts','tn','tr','tk','tw','uk','vi','wo','xh','yi']
    label_lang = ['Anglais','Français','Afrikaans','Akan','Albanais','Allemand','Amharique','Anglais','Arabe','Arménien','Assamais','Azéri','Bachkir','Bambara','Basque','Bengali','Biélorusse','Birman','Bosnien','Bulgare','Cachemiri','Catalan','Chichewa','Chinois','Cingalais','Coréen','Corse','Créolehaïtien','Croate','Danois','Dzongkha','Écossais','Espagnol','Espéranto','Estonien','Ewe','Féroïen','Fidjien','Finnois','Français','Frisonoccidental','Galicien','Gallois','Ganda','Géorgien','Grecmoderne','Guarani','Gujarati','Haoussa','Hébreu','Hindi','Hongrois','Igbo','Indonésien','Inuktitut','Irlandais','Islandais','Italien','Japonais','Kannada','Kazakh','Khmer','Kikuyu','Kinyarwanda','Kirghiz','Kirundi','Kurde','Lao','Latin','Letton','Limbourgeois','Lingala','Lituanien','Luxembourgeois','Macédonien','Malais','Malayalam','Maldivien','Malgache','Maltais','MaorideNouvelle-Zélande','Marathi','Mongol','Néerlandais','Népalais','Norvégien','Norvégienbokmål','Norvégiennynorsk','Occitan','Oriya','Ouïghour','Ourdou','Ouzbek','Pachto','Pendjabi','Persan','Polonais','Portugais','Roumain','Russe','Samoan','Sango','Sanskrit','Sarde','Serbe','Shona','Sindhi','Slovaque','Slovène','Somali','SothoduSud','Soundanais','Suédois','Swahili','Swati','Tadjik','Tagalog','Tahitien','Tamoul','Tatar','Tchèque','Télougou','Thaï','Tibétain','Tigrigna','Tongien','Tsonga','Tswana','Turc','Turkmène','Twi','Ukrainien','Vietnamien','Wolof','Xhosa','Yiddish']

    lang_src = {'ar': 'arabic', 'bg': 'bulgarian', 'de': 'german', 'el':'modern greek', 'en': 'english', 'es': 'spanish', 'fr': 'french', \
                'hi': 'hindi', 'it': 'italian', 'ja': 'japanese', 'nl': 'dutch', 'pl': 'polish', 'pt': 'portuguese', 'ru': 'russian', 'sw': 'swahili', \
                'th': 'thai', 'tr': 'turkish', 'ur': 'urdu', 'vi': 'vietnamese', 'zh': 'chinese'}
    
    st.write("#### "+tr("Choisissez le type de traduction")+" :")

    chosen_id = tab_bar(data=[
        TabBarItemData(id="tab1", title="small vocab", description=tr("avec Keras et un RNN")),
        TabBarItemData(id="tab2", title="small vocab", description=tr("avec Keras et un Transformer")),
        TabBarItemData(id="tab3", title=tr("Phrase personnelle"), description=tr("à écrire")),
        TabBarItemData(id="tab4", title=tr("Phrase personnelle"), description=tr("à dicter")),
        TabBarItemData(id="tab5", title=tr("Funny translation !"), description=tr("avec le Fine Tuning"))],
        default="tab1")
    
    if (chosen_id == "tab1") or (chosen_id == "tab2") :
        if (chosen_id == "tab1"):
            st.write("<center><h5><b>"+tr("Schéma d'un Réseau de Neurones Récurrents")+"</b></h5></center>", unsafe_allow_html=True)
            st.image("assets/deepnlp_graph3.png",use_column_width=True)
        else:
            st.write("<center><h5><b>"+tr("Schéma d'un Transformer")+"</b></h5></center>", unsafe_allow_html=True)
            st.image("assets/deepnlp_graph12.png",use_column_width=True)
        st.write("## **"+tr("Paramètres")+" :**\n")
        TabContainerHolder = st.container()
        Sens = TabContainerHolder.radio(tr('Sens')+':',('Anglais -> Français','Français -> Anglais'), horizontal=True)
        Lang = ('en_fr' if Sens=='Anglais -> Français' else 'fr_en')

        if (Lang=='en_fr'):
            df_data_src = df_data_en
            df_data_tgt = df_data_fr
            if (chosen_id == "tab1"):
                translation_model = rnn_en_fr
            else:
                translation_model = transformer_en_fr
        else:
            df_data_src = df_data_fr
            df_data_tgt = df_data_en
            if (chosen_id == "tab1"):
                translation_model = rnn_fr_en
            else:
                translation_model = transformer_fr_en
        sentence1 = st.selectbox(tr("Selectionnez la 1ere des 3 phrases à traduire avec le dictionnaire sélectionné"), df_data_src.iloc[:-4],index=int(n1) )
        n1 = df_data_src[df_data_src[0]==sentence1].index.values[0]

        st.write("## **"+tr("Résultats")+" :**\n")
        if (chosen_id == "tab1"):
            display_translation(n1, Lang,1)
        else: 
            display_translation(n1, Lang,2)

        st.write("## **"+tr("Details sur la méthode")+" :**\n")
        if (chosen_id == "tab1"):
            st.markdown(tr(
                """
                Nous avons utilisé 2 Gated Recurrent Units.
                Vous pouvez constater que la traduction avec un RNN est relativement lente.
                Ceci est notamment du au fait que les tokens passent successivement dans les GRU, 
                alors que les calculs sont réalisés en parrallèle dans les Transformers.  
                Le score BLEU est bien meilleur que celui des traductions mot à mot.
                <br>
                """)
                , unsafe_allow_html=True)
        else:
            st.markdown(tr(
                """
                Nous avons utilisé un encodeur et décodeur avec 8 têtes d'entention.
                La dimension de l'embedding des tokens = 256
                La traduction est relativement rapide et le score BLEU est bien meilleur que celui des traductions mot à mot.
                <br>
                """)
                , unsafe_allow_html=True)
        st.write("<center><h5>"+tr("Architecture du modèle utilisé")+":</h5>", unsafe_allow_html=True)
        plot_model(translation_model, show_shapes=True, show_layer_names=True, show_layer_activations=True,rankdir='TB',to_file=st.session_state.ImagePath+'/model_plot.png')
        st.image(st.session_state.ImagePath+'/model_plot.png',use_column_width=True)
        st.write("</center>", unsafe_allow_html=True)

'''