app.py

import os
import pprint
import tempfile
from typing import Dict, Text
import numpy as np
import tensorflow as tf
import tensorflow_recommenders as tfrs
import os
import unidecode
from nltk import word_tokenize
import re
import pandas as pd
from nltk.util import ngrams
import base64
import hashlib
import gradio as gr
import scann



df=pd.read_csv("/home/user/app/Qatar_translated_best_2500.csv",sep=",",header=0)
df=df.drop_duplicates()
df=df.dropna()
df["nome_vaga"]=df["nome_vaga"].map(lambda x: x.lower().title())
df["requisito"]=df["requisito"].map(lambda x: x[0:1000].lower())
my_dict=dict(df.iloc[0:int(df.shape[0]*0.9),:])
my_dict_cego=dict(df.iloc[int(df.shape[0]*0.9):,:])
ratings = tf.data.Dataset.from_tensor_slices(my_dict).map(lambda x: {
    "code": x["code"],
    "nome_vaga": x["nome_vaga"],
    "requisito": tf.strings.split(x["requisito"],maxsplit=106)
})
movies = tf.data.Dataset.from_tensor_slices(dict(df)).map(lambda x: {
    "code": x["code"],
    "nome_vaga": x["nome_vaga"]
})
movies = movies.map(lambda x: x["code"])
ratings_cego = tf.data.Dataset.from_tensor_slices(my_dict_cego).map(lambda x: {
    "code": x["code"],
    "requisito": tf.strings.split(x["requisito"],maxsplit=106)
})
tf.random.set_seed(42)
shuffled = ratings.shuffle(int(df.shape[0]*0.9), seed=42, reshuffle_each_iteration=False)
shuffled2 = ratings_cego.shuffle(int(df.shape[0]*0.1), seed=42, reshuffle_each_iteration=False)
train = shuffled.take(int(df.shape[0]*0.9))
test = shuffled.take(int(df.shape[0]*0.1))
cego=shuffled2
movie_titles = movies#.map(lambda x: x["code"])
user_ids = ratings.map(lambda x: x["requisito"])
xx=[]
for x in user_ids.as_numpy_iterator():
    try:
        xx.append(x)
    except:
        pass
unique_movie_titles = np.unique(list(movie_titles.as_numpy_iterator()))
unique_user_ids = np.unique(np.concatenate(xx))
user_ids=user_ids.batch(int(df.shape[0]*0.9))
layer = tf.keras.layers.StringLookup(vocabulary=unique_user_ids)
unique_movie_titles[:10]
embedding_dimension = 768
user_model = tf.keras.Sequential([
  tf.keras.layers.StringLookup(
      vocabulary=unique_user_ids, mask_token=None),
  # We add an additional embedding to account for unknown tokens.
  tf.keras.layers.Embedding(len(unique_user_ids) + 1, embedding_dimension),
  
])
movie_model = tf.keras.Sequential([
  tf.keras.layers.StringLookup(
      vocabulary=unique_movie_titles, mask_token=None),
  tf.keras.layers.Embedding(len(unique_movie_titles) + 1, embedding_dimension)
])
metrics = tfrs.metrics.FactorizedTopK(
  candidates=movies.batch(df.shape[0]
).map(movie_model)
)
task = tfrs.tasks.Retrieval(
  metrics=metrics
)
class MovielensModel(tfrs.Model):

  def __init__(self, user_model, movie_model):
    super().__init__()
    self.movie_model: tf.keras.Model = movie_model
    self.user_model: tf.keras.Model = user_model
    self.task: tf.keras.layers.Layer = task

  def compute_loss(self, features: Dict[Text, tf.Tensor], training=False) -> tf.Tensor:
    user_embeddings = self.user_model(features["requisito"])
    positive_movie_embeddings = self.movie_model(features["code"])
    return self.task(tf.reduce_sum(user_embeddings,axis=1), positive_movie_embeddings)

class NoBaseClassMovielensModel(tf.keras.Model):

  def __init__(self, user_model, movie_model):
    super().__init__()
    self.movie_model: tf.keras.Model = movie_model
    self.user_model: tf.keras.Model = user_model
    self.task: tf.keras.layers.Layer = task

  def train_step(self, features: Dict[Text, tf.Tensor]) -> tf.Tensor:

    with tf.GradientTape() as tape:

      user_embeddings = self.user_model(features["requisito"])
      positive_movie_embeddings = self.movie_model(features["code"])
      loss = self.task(user_embeddings, positive_movie_embeddings)

      regularization_loss = sum(self.losses)

      total_loss = loss + regularization_loss

    gradients = tape.gradient(total_loss, self.trainable_variables)
    self.optimizer.apply_gradients(zip(gradients, self.trainable_variables))

    metrics = {metric.name: metric.result() for metric in self.metrics}
    metrics["loss"] = loss
    metrics["regularization_loss"] = regularization_loss
    metrics["total_loss"] = total_loss

    return metrics

  def test_step(self, features: Dict[Text, tf.Tensor]) -> tf.Tensor:

    user_embeddings = self.user_model(features["requisito"])
    positive_movie_embeddings = self.movie_model(features["code"])
    loss = self.task(user_embeddings, positive_movie_embeddings)

    regularization_loss = sum(self.losses)

    total_loss = loss + regularization_loss

    metrics = {metric.name: metric.result() for metric in self.metrics}
    metrics["loss"] = loss
    metrics["regularization_loss"] = regularization_loss
    metrics["total_loss"] = total_loss

    return metrics

model = MovielensModel(user_model, movie_model)
model.compile(optimizer=tf.keras.optimizers.Adagrad(learning_rate=0.08))
cached_train = train.shuffle(int(df.shape[0]*0.9)).batch(int(df.shape[0]*0.9)).cache()
cached_test = test.batch(int(df.shape[0]*0.15)).cache()
path = os.path.join("/home/user/app/", "model/")
cp_callback = tf.keras.callbacks.ModelCheckpoint(
    filepath=path, 
    verbose=1, 
    save_weights_only=True,
    save_freq=2)

model.fit(cached_train, callbacks=[cp_callback],epochs=110)

index=df["code"].map(lambda x: [model.movie_model(tf.constant(x))])

indice=[]
for i in range(0,1633):
    indice.append(np.array(index)[i][0])

searcher = scann.scann_ops_pybind.builder(np.array(indice), 10, "dot_product").tree(
    num_leaves=1500, num_leaves_to_search=500, training_sample_size=df.shape[0]).score_brute_force(
    2, quantize=True).build()

import matplotlib.pyplot as plt

def predict(text):
    campos=str(text).lower()
    query=np.sum([model.user_model(tf.constant(campos.split()[i])) for i in range(0,len(campos.split()))],axis=0)
    neighbors, distances = searcher.search_batched([query])
    xx = df.iloc[neighbors[0],:].nome_vaga
    fig = plt.figure(figsize=(14,9))
    plt.bar(list(xx),distances[0]*0.8*10)
    plt.title('Degree of match')
    plt.xlabel('Labels')
    plt.xticks(rotation=270)

    plt.ylabel('Distances')
    for x, y in zip(list(range(0,10)),distances[0]*0.8*10):
      plt.text(x, y, y, ha='center', va='bottom', fontsize=12, color='black')
    return xx, fig

demo = gr.Interface(fn=predict, inputs=gr.inputs.Textbox(label='CANDIDATE COMPETENCES - Click *Clear* before adding new input'), \
    outputs=[gr.outputs.Textbox(label='SUGGESTED VACANCIES'),\
      gr.Plot()],\
        css='div {margin-left: auto; margin-right: auto; width: 100%;\
            background-image: url("https://drive.google.com/uc?export=view&id=1KNnISAUcvh2Pt08f-EJZJYCIgkrKw3PI"); repeat 0 0;}')\
              .launch(share=False)