Transfer Learning.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Transfer Learning:"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;O objetivo deste notebook é apresentar o tema transfer learning, assim como mostrar exemplos prático de como fazer e quais as vantagens relacionadas ao seu uso. As informações e exemplos são livremente adaptados deste excelente post no [Medium](https://towardsdatascience.com/a-comprehensive-hands-on-guide-to-transfer-learning-with-real-world-applications-in-deep-learning-212bf3b2f27a)."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Introdução"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Redes neurais, assim como *deep learning*, são temas em alta quando o assunto é ***machine learning*** no mundo atual. Na realidade *machine learning* é um dos temas mais tratados em técnologia, sendo *Transfer Learning*, talvez, uma das motrizes mais importantes em impulsionar o desenvolvimento de *deep learning*, e para falar sobre um tema tão importante e complexo, primeiro precisamos elucidar um pouco os temas citados anteriormente.\n",
    "\n",
    "&emsp;Redes neurais são algoritmos de aprendizado por máquina que buscam mimetizar o comportamento do cérebro humano, afinal com eles buscamos fazer com que o computador aprenda uma determinada tarefa. Neste âmbito faz sentido falar em fazer o computador aprender à semelhança do que ocorre no processo de aprendizado realizado por nosso sistema nervoso constantemente, desde antes de nascermos. Esta é a função dos algoritmos de redes neurais, que hoje já evoluiram enormemente, estando disponíveis diversas arquiteturas diferentes: Convolucionais, recorrentes, memórias de longo e curto prazo, etc.\n",
    "\n",
    "&emsp;O termo *deep learning* é utilizado para se referir à estruturas de redes neurais profundas, como o próprio nome diz. Redes profundas são aquelas dotadas de muitas camadas e, mais simplificadamente, dotadas de muitos parâmetros treináveis. Estes parâmetros são efetivamente o que é desejado que uma rede aprenda, parâmetros corretos de uma rede neural podem levá-la a níveis de generalização que uma pessoa não consegue. O grande problema de termos muitos destes parâmetros está no fato que quanto mais profunda a rede, maior é a demanda computacional de processamento e de quantidade de dados na base para realizar o treinamento. Em poucas palavras, uma rede profunda tem capacidade maior de aprendizado, mas em troca custa muito, é dificimente acessível com os computadores que temos em casa atualmente e precisa de quantidades realmente grandes de dados. \n",
    "\n",
    "&emsp;Dentro deste contexto de dificuldade associada ao treino de redes muito complexas que existem redes deste tipo treinadas por gigantes corporações que detem recursos financeiros, tecnológicos e de informação, por exemplo a Google, Microsoft e Facebook, para realizar o dispendioso processo de treinamento. Estas redes estão disponíveis online para quem quiser testá-las e utilizá-las para seus próprios projetos. Surgem a partir destas a ideia de utilizar elas como parte do aprendizado de outras redes que desempenham tarefas diferentes, mas para explicar isso melhor vamos a um exemplo.\n",
    "\n",
    "&emsp;O nosso sistema nervoso sabe transferir o aprendizado entre tarefas, afinal de contas não aprendemos sempre do zero, tudo que aprendemos em nossa vida é utilizado como base para a construção de novos aprendizados. Vamos pensar em andar de bicicleta, é importante ter noções básicas de equilíbrio que nos ajudam a aprender como se manter em cima de uma, sem cair, podendo finalmente abrir mão das rodinhas. Imagine agora este mesmo indivíduo que aos 8 anos aprendeu a andar de bicicleta, aos 20 resolve tirar carta e aprender a andar de moto, o cérebro dele não irá começar do zero a aprender como se equilibrar em cima de moto, certo? Ele já sabe andar de bicicleta e, salvas devidas proporções, muito do que ele aprendeu aos 8 anos será utilizado agora para que ele aprenda a andar em uma motocicleta. \n",
    "\n",
    "&emsp;Esse é o objetivo do *trasnfer learning*, proporcionar que um projeto nosso com tarefa ou base de dados parecida com uma rede já treinada possa se aproveitar dos parâmetros já definidos desta. Assumimos vários destes parâmetros como nossos em nossa rede. Agora é possível realizar esse treinamento em casa, sem deter grande quantidade de poder computacional ou de uma base de dados enorme, simplesmente pelo fato que daquela grande quantidade de parâmetros treinaveis que um modelo de deeplearning possui, uma boa parte, será importada desta outra rede, deixando para nosso computador e nossa base de dados treinar apenas uma pequena parte da rede. Este processo, além de tornar tudo muito mais veloz, é capaz de aumentar significativamente o desempenho do modelo, como veremos mais à frente.\n",
    "\n",
    "&emsp;Qual parte da rede especificamente será discutido mais adiante, mas é importante compreender estes conceitos, para tanto seguem aqui algumas imagens que ilustram o que foi explorado na explicação acima.\n",
    "\n",
    "<img src=\"https://miro.medium.com/max/1838/1*9GTEzcO8KxxrfutmtsPs3Q.png\">\n",
    "\n",
    "<img src = \"https://miro.medium.com/max/1400/1*Ww3AMxZeoiB84GVSRBr4Bw.png\">\n",
    "\n",
    "Fonte: [Medium](https://towardsdatascience.com/a-comprehensive-hands-on-guide-to-transfer-learning-with-real-world-applications-in-deep-learning-212bf3b2f27a)\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Definições formais"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vamos definir formalmente o conceito de *transfer learning*, se você não se adaptar bem à esta parte, por conta das definições um pouco mais matemáticas, recomendo que se esforçe e busque pesquisar e seguir os links, pois é importante para a compreenção adiante do tema.\n",
    "\n",
    "&emsp;Pois bem, esta definição é dada por Pan e Yang no seguinte [Paper](https://ieeexplore.ieee.org/document/5288526). Baseada altamente em conceitos de domínio, tarefa e probabilidades marginais da matemática, tarefa é um pouco mais intuitivo, mas se o leitor não estiver familiarizado com os outros conceitos, temos alguns links que podem ajudar a exclarecer um pouco, respectivamente: [domínio](https://pt.wikipedia.org/wiki/Dom%C3%ADnio_(matem%C3%A1tica)) e [probabilidade marginal](https://pt.wikipedia.org/wiki/Distribui%C3%A7%C3%A3o_marginal). Dados os *disclaimers* vamos às definições:\n",
    ">Um domínio **D** definido como uma tupla de dois elementos, consistindo de um espaço de *features* **$\\chi$** e uma probabilidade marginal **$P(X)$**, onde **X** é uma amostra dos dados. Assim vamos definir o a seguinte relação: **$D = \\{\\chi, P(X)\\}$**.\n",
    ">>Lembrando que: **$P(X), X = \\{x_1,x_2,...,x_n\\}, x_i \\in \\chi$**\n",
    "\n",
    ">Uma tarefa **$T$** é definida também como uma tupla, mas do espaço de *labels*, **$\\gamma$**, e de uma função objetiva **$\\eta$**. Esta função será definida probabilisticamente como: **$\\eta = P(Y \\mid X)$**. Assim temos que a tarefa **$T = \\{\\gamma, \\eta\\} = \\{\\gamma, P(Y \\mid X)\\}$**.\n",
    ">>Neste caso: **$Y = \\{y_1,y_2,...,y_n\\}, y_i \\in \\gamma$**\n",
    "\n",
    "&emsp;Nesta definição acho que vale uma breve explicação, a função objetiva é preditiva, probabilidade de **Y** dado **X**, que aprendeu a partir dos pares correspondentes de *features* e *labels*, $(x_i,y_i),\\; x_i \\in \\chi \\; ; y_i \\in \\gamma$.\n",
    "\n",
    "&emsp;Assim podemos finalmente definir *transfer learning*, a partir da definição e do completíssimo *post* de Sebastian Ruder aqui:[\n",
    "Transfer Learning - Machine Learning's Next Frontier](https://ruder.io/transfer-learning/), então bora lá:\n",
    "\n",
    ">Dado um domínio fonte: **$D_f$** e uma tarefa fonte associada à este domínio, **$T_f$**. E dado um domínio alvo **$D_a$**, com uma tarefa alvo associada, **$T_a$**. O objetivo do *transfer learning* é habilitar o aprendizado da função objetiva alvo, **$P(X_a \\mid Y_a)$**, por meio das informações obtidas a partir de **$D_f$** e **$T_f$**, claro que **$D_f \\ne D_a$** e **$T_f \\ne T_a$**.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Cenários de *Transfer Learning*"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Agora que o leitor já tem completo conhecimento da ideia, além da definição formal do tema, vamos falar um pouco mais sobre os possíveis cenários que você pode encontrar ao lidar com uma abordagem de *TL* no seu dia-a-dia. Dadas as definições acima, elas refletem imposição de diferenças entre os elementos das duas tuplas definidas anteriormente.\n",
    "\n",
    ">**$\\chi_f \\ne \\chi_a$**: O espaço de *features* é diferente. Um exemplo bem exagerado para este cenário, no âmbito da classificação de imagens, é quando temos uma domínio fonte que classifica cães, então ele busca *features* como patas, orelhas, fucinho, etc, e o domínio alvo é classificativo de carros, temos então características como rodas, vidros, portas, entre outros. \n",
    "\n",
    ">**$P(X_f) \\ne P(X_a)$**: A probabilidade marginal entre os domínios é diferente, assim quando as redes são treinadas para cenários distintos em termos de quantidades de features a serem extraidas para classificação. Geralmente conhecido como *domain adaptation*.\n",
    "\n",
    ">**$\\gamma_f \\ne \\gamma_a$**: Aqui o espaço de *labels* da fonte e alvo são diferentes, este é o típico cenário que dificilmente ocorre isoladamente, geralmente quando este acontece, tambem acontece o próximo, visto que é raro que dois modelos apresentem diferentes apresentem probabilidades condicionais semelhantes enquanto os espaços de *labels* são distintos.\n",
    "\n",
    ">**$P(X_f \\mid Y_f) \\ne P(X_a \\mid Y_a)$**: O último cenário é o já spoilado no item anterior, quando as probabilidades condiçionais são diferentes entre domínios, isto é, alvo e fonte possuem base de dados desbalanceadas internamente. Essa ideia faz referência à casos bastante comuns na prática como *undersampling* ou *oversampling*. \n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### ***Confusing Domain***"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Antes de tratarmos de conceitos mais práticos e aplicaveis relacionados ao tema, acho interessante introduzir brevemente um conceito mais avançado de *transfer learning* com o objetivo de fundamentar melhor a argumentação que esta é uma área com o potencial de avançar o desenvolvimento de aprendizagem de máquina, que vai muito além de dar um *import* no *TensorFlow* ou no *PyTorch*."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;***Confusing Domain*** é um dos conceitos mais interessantes desenvolvidos em TL, ao meu ver, e consiste em fazer o processo de selecionar uma rede com um domínio fonte e um domínio alvo que deseja-se realizar uma tarefa de classificação. A ideia do método é similarizar ambos domínios, de forma que a rede não consiga distinguir bem entre ambos, ou seja, fazer a rede confundir os domínios. Efetivamente isto significa adicionar um objetivo novo à rede, sem apagar o anterior para o qual ela já foi treinada.\n",
    "\n",
    "&emsp;Assumindo um pouco de conhecimento de funcionamento de redes neurais, que caso o leitor não esteja familiarizado bem, pode buscar este conhecimento aqui: [Redes Neurais | Teoria](https://medium.com/turing-talks/turing-talks-19-modelos-de-predi%C3%A7%C3%A3o-redes-neurais-1f165583a927). Como temos dois domínios, dois objetivos, é assumível também que teremos duas funções de custo e dois gradientes sendo propagados de volta na rede, o novo gradiente associado ao domínio inserido, porém, será propagado com valor invertido, da seguinte forma:\n",
    "\n",
    "<img src = \"https://ruder.io/content/images/2017/03/confusing_domains_with_gradient_reversal_layer.png\">\n",
    "\n",
    "Fonte: [Sebastian Ruder](https://ruder.io/transfer-learning/)\n",
    "\n",
    "&emsp;Efetivamente o que acontece na imagem é que ao inverter o valor do gradiente teremos uma rede que ao invés de minimizar a função de custo, irá maximizá-la. Na prática ao fazer este processo o modelo aprende novas representações que diminui as tarefas fonte, sem permitir que o modelo consiga diferenciar entre ambas, que é benéfico para o processo de transferência de aprendizado. Vamos à um exemplo que ilustra este procedimento:\n",
    "\n",
    "<img src = \"https://ruder.io/content/images/2017/03/domain_confusion_tzeng.png\">\n",
    "\n",
    "Fonte: [Sebastian Ruder](https://ruder.io/transfer-learning/)\n",
    "\n",
    "&emsp;Claramente no primeiro caso a rede é capaz de separar ambas as tarefas, já no segundo gráfico não é possível fazer a diferenciação devido à sobreposição de ambas."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Casos de *Trasfer Learning*"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Já falei muito sobre TL, dei definições formais, cenários diferentes, mas vamos para uma parte mais prática agora, como eu faço isso de fato? Pra cada caso elencado acima, como eu faço? O que eu deveria deixar fixo? O que eu deveria treinar? Bem vamos responder a todas essas perguntar agora. Primeiramente segue uma imagem que sumariza tudo que vou dizer:\n",
    "\n",
    "<img src = \"https://miro.medium.com/max/1222/1*mEHO0-LifV7MgwXSpY9wyQ.png\">\n",
    "\n",
    "Fonte: [Medium](https://towardsdatascience.com/a-comprehensive-hands-on-guide-to-transfer-learning-with-real-world-applications-in-deep-learning-212bf3b2f27a)\n",
    "\n",
    "Vamos elencar os casos de transferência de aprendizado com base na imagem e nas definições:\n",
    "\n",
    ">***Transfer Learning Indutivo***: Neste primeiro caso temos que os domínios fonte e alvo são semelhantes, ou pelo menos análogos, enquanto as tarefas são diferentes. Assim o modelo tenta se aproveitar dos vieses presentes na fonte. Um exemplo pode ser de uma rede especializada em identificar cachorros em uma foto pode ser usada para classificar cachorros a partir de sua raça, o domínio é o mesmo, ou semelhante, cachorro, mas a tarefa de classificação não é. Pela imagem é possível ver que há mais subclassificações, dependendo se o domínio possui ou não *labels*, *multitask* e *self-taught learning*.\n",
    "\n",
    ">***Transfer Learning Transdutivo***: No segundo caso temos o exato oposto ao primeiro, aqui as tarefas fonte e alvo são iguais, porém os respectivos domínios são diferentes. Este problema se aplica constantemente em desafios de reconhecimento de objetos, por exemplo, imagine uma rede treinada em reconhecer cães, por outro lado nossa tarefa alvo é o reconhecimento de gatos, reconhecimento é o mesmo em ambas as tarefas, mas o domínio é diferente, já que: $Cão \\ne Gato$ , apesar de que...\n",
    "<img src = https://miro.medium.com/max/875/1*mHTXSibh4teRnxnNatmXmQ.png>\n",
    "Fonte: [Medium 2](https://medium.com/voice-tech-podcast/how-intelligent-is-ai-7ed574b02c2b)\n",
    "\n",
    ">***Transfer Learning Não Supervisionado***: Aqui temos um terceiro caso que difere dos outros dois, mas é mais semelhante ao primeiro que ao segundo. Este é o caso quando os *labels* não estão disponíveis tanto para os domínios alvo e fonte. Mas é semelhante ao primeiro no sentido que aqui também os domínios fonte e alvo são semelhantes, enquanto as tarefas são diferentes.\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Estratégias de abordagem"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;A ideia desta seção é mostrar algumas das abordagens mais utilizadas para treinar seus modelos a partir da transferência de aprendizado. Estas estratégias são baseadas altamente em modelos de redes neurais convolucionais, além de estarem no contexto de *deep learning*.\n",
    "\n",
    "&emsp;Uma rede neural convolucional, CNN, geralmente tem uma estrutura padrão de montagem, claro que existem variações, mas ao abordar um novo problema é comum o desenvolvimento de uma mesma estrutura, que consiste em: \n",
    "- Camadas convolucionais, alternadas com camadas de *pooling* e ativação\n",
    "- Camadas totalmente conectadas, com ativação.\n",
    "- É possível haver camadas de regularização no meio, mas vamos manter as coisas mais simples por enquanto.\n",
    "\n",
    "&emsp;Se o que eu acabei de falar está em grego para você, leitor, por favor procure dar uma lida neste [artigo](https://medium.com/turing-talks/turing-talks-23-modelos-de-predi%C3%A7%C3%A3o-redes-neurais-convolucionais-d364654a34de) antes de prosseguir. Uma das estratégias consiste em cortar as últimas camadas do modelo, geralmente as *fully-connected*, todas ou parte delas, de forma a utilizar a arquitetura restante como parte do seu modelo alvo, e treinar novas camadas classificativas. Iremos fazer isso na parte prática mais a frente, mas por enquanto veja a ilustração a seguir:\n",
    "\n",
    "<img src = \"https://miro.medium.com/max/1400/1*qfQ3hmHLwApXZBN-A85r8g.png\">\n",
    "\n",
    "Fonte: [Medium](https://towardsdatascience.com/a-comprehensive-hands-on-guide-to-transfer-learning-with-real-world-applications-in-deep-learning-212bf3b2f27a)\n",
    "\n",
    "&emsp;Esta é uma estratégia que faz muito sentido para o primeiro caso exposto anteriormente, ou seja, de *Transfer learning* indutivo, visto que se os domínios são semelhantes, as *features* identificadas pela rede fonte são de interesse para o modelo alvo, enquanto que como as tarefas são distintas, a forma de classificação da rede fonte não é algo interessante para a outra. A partir deste ponto podemos criar novas camadas finais e treiná-las com nossos dados para ajustar a rede à tarefa alvo.\n",
    "\n",
    "&emsp;Esta estratégia é conhecida como \"*Off the shelf features*\", pelo simples motivo que pegamos as features disponíveis na rede fonte e utilizamos elas diretamente em nossa rede alvo, um questionamento bastante pertinente seria o quão bem tal método consegue performar por si só? Bem existem várias tarefas diferentes que podem ser analisadas e discutidas, então eu vou simplesmente colocar este gráfico aqui que sumariza bastante tudo que eu tenho para falar:\n",
    "\n",
    "<img src = \"https://miro.medium.com/max/1400/1*d__1oSbHOSEq6m4bFflE5g.png\">\n",
    "\n",
    "Fonte:[Medium](https://towardsdatascience.com/a-comprehensive-hands-on-guide-to-transfer-learning-with-real-world-applications-in-deep-learning-212bf3b2f27a)\n",
    "\n",
    "&emsp;Bem impressionante não é? Em muitas tarefas o desempenho é próximo, se não igual, à redes treinadas especialmente para aquela *task*. De cara da pra perceber e ressaltar a importância deste método atualmente para o progresso em *machine learning*, estamos falando de desempenhos altíssimos com redes muito simples de serem montadas. Mas já temos um bom resultado nesta estratégia, vamos falar sobre outra. Legal, mas e se quisessemos melhor o desempenho deste modelo ainda mais?\n",
    "\n",
    "&emsp; Vamos abordar um outro tema importante, pode ser considerado uma nova estratégia ou apenas uma continuação da anterior, fica a gosto do freguês, enfim, vamos falar sobre ***Fine Tuning*** ou ajuste fino em tradução livre. A ideia aqui é não só se aproveitar de um modelo fonte, extrair suas camadas finais, mas também treinar alguns de seus parâmetros anteriores, para que eles possam se adaptar um pouco melhor à sua tarefa alvo. Uma abordagem seria \"congelar\" alguns parâmetros mais genéricos, como os referentes às primeiras camadas convolucionais em uma CNN, e ajustar os parâmetros mais específicos, referentes às camadas mais internas. Sem dúvida é uma abordagem que gera um pouco mais de custo e requer um poder computacional maior, mas ainda nada próximo à treinar uma rede inteira com pesos aleatórios. Necessita de mais dados, mas se estes estão disponíveis, pode valer bastante a pena.\n",
    "\n",
    "&emsp;Só uma consideração do que significa \"congelar\" a rede, basicamente á atribuir *learning rates* diferentes para cada camdada da rede, quando congelamos as primeiras camadas, damos à elas $lr = 0$, enquanto que para camadas mais avançadas, $lr > 0$. Mas novamente essa parte \"operacional\" vai ficar mais clara no código."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Redes estado da arte"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;O leitor talvez esteja se perguntando neste exato momento, mas e quais modelos eu deveria utilizar para fazer o *transfer*? Bem a resposta é simples, você deveria utilizar modelos eficientes, com altos indices de acurácia e outras métricas em variados datasets. Modelos deste tipo geralmente recebem o nome de **Estado da arte** por serem muito bons ou revolucionários em suas tarefas. A ideia por trás de utilizar estes modelos é que por serem tão eficientes, eles possuem uma grande capacidade de generalização. \n",
    "\n",
    "&emsp;Existem várias das redes estado da arte, para diferentes campos de *machine learning*, vou elencar algumas das mais populares para cada área:\n",
    ">Em visão computacional: \n",
    "- [VGG-16](https://www.kaggle.com/keras/vgg16/home)\n",
    "- [VGG-19](https://www.kaggle.com/keras/vgg19/home)\n",
    "- [ResNet-50](https://www.kaggle.com/keras/resnet50/home)\n",
    "\n",
    ">Em Processamento de linguagem natural:\n",
    "- [Word2Vec](https://en.wikipedia.org/wiki/Word2vec)\n",
    "- [Glove](https://nlp.stanford.edu/projects/glove/)\n",
    "\n",
    "&emsp;Estas somente para elencar algumas, existem outras dentro destas áreas e fora. Vou contar um segredo, a parte teórica de tipos de *transfer learning* muitas vezes pode ser ignorada ao fazer um transfer com alguma destas redes, isso porque elas são muito eficientes, treinadas com muitos dados em diversas categorias e por isso sua capacidade de generalização é muito grande, logo pode ser um incremento mesmo em *off the shelf features* mesmo que os domnínios sejam diferentes, por exemplo. Nestes casos o *fine-tuning* gera melhoras consideráveis também."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Hands-On"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Finalmente vamos colocar em prática tudo que foi exposto acima, para tal vamos utilizar primeiramente o dataset: [Flowers Recognition](https://www.kaggle.com/alxmamaev/flowers-recognition). Este consiste de um dataset desbalanceado, com 5 classes diferentes, correspondente à diferentes espécies de flores: Rosa, Tulipa, Girassol, Margarida e Dente-de-leão. Algumas das dificuldades de associadas ao uso deste dataset estão na composição das imagens, algumas vezes não ficam muito evidentes as flores, mas no geral é um dataset que consegue-se um desempenho razoável.\n",
    "\n",
    "&emsp;É por essas e outras que vamos utilizar este dataset, ele vem separado em pastas, cada qual com uma espécie apenas. Para começar você vai precisar baixar o dataset no link acima, e de preferência extraí-lo na mesma pasta em que este notebook estiver presente, para que não seja necessário modificar nenhuma linha de código, mas se precisar é onde estiverem os caminhos para os arquivos e pastas."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Organizando os arquivos"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vamos começar importando as bibliotecas necessárias:\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np #para trabalhar com vetores\n",
    "import glob\n",
    "import os \n",
    "import shutil\n",
    "import pathlib\n",
    "import os\n",
    "#Estas últimas apenas para tratar de arquivos "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Em seguida vamos tentar facilitar nosso processo de gerar um dataset com base na ideia de dar nome às imagens com base nos labels:\n",
    "\n",
    "Ps: Se você já rodou esse código uma vez não precisa rodar novamente, vai acabar conflitando os arquivos."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset_path = '/home/fill/Documentos/Turing/Árvore/Transfer Learning/flowers-recognition'\n",
    "os.chdir(dataset_path)   #Aqui entramos no diretório do dataset\n",
    "class_folders = os.listdir('.') #Pegamos os strings referente aos diretórios referentes a cada classe\n",
    "for class_folder in class_folders: #Loop para varrer cada diretório\n",
    "    i = 1 #Variável auxiliar para nomear as imagens\n",
    "    os.chdir(class_folder) #Entramos no diretório de uma classe\n",
    "    images = os.listdir('.') #Pegamos os strings de cada arquivo de imagem\n",
    "    for image in images: #Loop para varrer as imagens\n",
    "        os.rename(image, class_folder+'.{}'.format(i)+'.jpg') #Renomeamos cada imagem com o nome no formato que queremos\n",
    "        i = i+1 #Atualizamos a variável\n",
    "    os.chdir('..') #Após renomear todas as imagens em um diretório, saimos dele para ir ao próximo"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vamos agora criar as listas com as imagens de cada classe, a função glob um string com o nome dos arquivos na pasta destino:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "ename": "NameError",
     "evalue": "name 'glob' is not defined",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mNameError\u001b[0m                                 Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-1-43eada2143a9>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[0;32m----> 1\u001b[0;31m \u001b[0mdaisy_files\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mglob\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mglob\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'/home/fill/Documentos/Turing/Árvore/Transfer Learning/flowers-recognition/daisy/*'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m      2\u001b[0m \u001b[0mdandelion_files\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mglob\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mglob\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'/home/fill/Documentos/Turing/Árvore/Transfer Learning/flowers-recognition/dandelion/*'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      3\u001b[0m \u001b[0mrose_files\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mglob\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mglob\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'/home/fill/Documentos/Turing/Árvore/Transfer Learning/flowers-recognition/rose/*'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      4\u001b[0m \u001b[0msunflower_files\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mglob\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mglob\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'/home/fill/Documentos/Turing/Árvore/Transfer Learning/flowers-recognition/sunflower/*'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      5\u001b[0m \u001b[0mtulip_files\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mglob\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mglob\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'/home/fill/Documentos/Turing/Árvore/Transfer Learning/flowers-recognition/tulip/*'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;31mNameError\u001b[0m: name 'glob' is not defined"
     ]
    }
   ],
   "source": [
    "daisy_files = glob.glob('/home/fill/Documentos/Turing/Árvore/Transfer Learning/flowers-recognition/daisy/*')\n",
    "dandelion_files = glob.glob('/home/fill/Documentos/Turing/Árvore/Transfer Learning/flowers-recognition/dandelion/*')\n",
    "rose_files = glob.glob('/home/fill/Documentos/Turing/Árvore/Transfer Learning/flowers-recognition/rose/*')\n",
    "sunflower_files = glob.glob('/home/fill/Documentos/Turing/Árvore/Transfer Learning/flowers-recognition/sunflower/*')\n",
    "tulip_files = glob.glob('/home/fill/Documentos/Turing/Árvore/Transfer Learning/flowers-recognition/tulip/*')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vamos checar quantas imagens temos em cada espécie:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(769, 1052, 784, 734, 984)"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(daisy_files), len(dandelion_files), len(rose_files), len(sunflower_files), len(tulip_files)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Trata-se de um dataset claramente desbalanceado, é mais uma das dificuldades associadas à este. De qualquer forma, vamos separar as imagens em bases de treino, validação e teste. Usaremos para tal aproximadamente 70% de cada espécie como treino, 15% como validação e 15% como teste. Primeiro vamos apenas separar o treino, lembrando que na função choice é importante definir o replace como **False** para evitar que uma mesma imagem seja escolhida mais de uma vez. Depois precisamos apagar os elementos selecionados da lista geral:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "daisy_train = np.random.choice(daisy_files, size = 538, replace = False) #Seleciona sem sobrepor\n",
    "daisy_files = list(set(daisy_files) - set(daisy_train))  #Apaga os elementos já selecionados\n",
    "\n",
    "dandelion_train = np.random.choice(dandelion_files, size = 738, replace = False)\n",
    "dandelion_files = list(set(dandelion_files) - set(dandelion_train))\n",
    "\n",
    "rose_train = np.random.choice(rose_files, size = 549, replace = False)\n",
    "rose_files = list(set(rose_files) - set(rose_train))\n",
    "\n",
    "sunflower_train = np.random.choice(sunflower_files, size = 514, replace = False)\n",
    "sunflower_files = list(set(sunflower_files) - set(sunflower_train))\n",
    "\n",
    "tulip_train = np.random.choice(tulip_files, size = 689, replace = False)\n",
    "tulip_files = list(set(tulip_files) - set(tulip_train))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Verificando então as quantidades restantes:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(231, 314, 235, 220, 295)"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(daisy_files), len(dandelion_files), len(rose_files), len(sunflower_files), len(tulip_files)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Agora podemos selecionar metade para validação, metade para teste, para cada categoria que temos:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "daisy_val = np.random.choice(daisy_files, size = 115, replace = False)\n",
    "daisy_test = list(set(daisy_files) - set(daisy_val))\n",
    "\n",
    "dandelion_val = np.random.choice(dandelion_files, size = 158, replace = False)\n",
    "dandelion_test = list(set(dandelion_files) - set(dandelion_val))\n",
    "\n",
    "rose_val = np.random.choice(rose_files, size = 117, replace = False)\n",
    "rose_test = list(set(rose_files) - set(rose_val))\n",
    "\n",
    "sunflower_val = np.random.choice(sunflower_files, size = 110, replace = False)\n",
    "sunflower_test = list(set(sunflower_files) - set(sunflower_val))\n",
    "\n",
    "tulip_val = np.random.choice(tulip_files, size = 147, replace = False)\n",
    "tulip_test = list(set(tulip_files) - set(tulip_val))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vamos analisar como ficou composto os datasets referentes a cada uma das classes:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "daisy datasets: 538, 115, 116 imagens\n",
      "dandelion datasets: 738, 158, 156 imagens\n",
      "rose datasets: 549, 117, 118 imagens\n",
      "sunflower datasets: 514, 110, 110 imagens\n",
      "tulip datasets: 689, 147, 148 imagens\n"
     ]
    }
   ],
   "source": [
    "print('daisy datasets: {}, {}, {} imagens'.format(len(daisy_train), len(daisy_val), len(daisy_test)))\n",
    "print('dandelion datasets: {}, {}, {} imagens'.format(len(dandelion_train), len(dandelion_val), len(dandelion_test)))\n",
    "print('rose datasets: {}, {}, {} imagens'.format(len(rose_train), len(rose_val), len(rose_test)))\n",
    "print('sunflower datasets: {}, {}, {} imagens'.format(len(sunflower_train), len(sunflower_val), len(sunflower_test)))\n",
    "print('tulip datasets: {}, {}, {} imagens'.format(len(tulip_train), len(tulip_val), len(tulip_test)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Agora que temos uma definição boa de quais imagens estão onde, vamos tratar de carregar o dataset propriamente, afinal strings não são imagens, não pretendemos passar strings para nossa CNN."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Montando o dataset"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vamos trabalhar com imagens no tamanho 100x100 pixels, pelo simples fato de não querer desenvolver e calcular muitos parâmetros treináveis, principalemente em modelos futuros, para ficar mais rápido o processamento. Mas vale dizer que essa decisão não mantém o *aspect ratio* do dataset. Vamos concatenar as listas que criamos anteriormente também, e embaralhar esses vetores, pois concatenar não é embaralhar:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "img_shape = (100, 100) #Definindo o tamanho das imagens\n",
    "\n",
    "train_files = np.concatenate([daisy_train, dandelion_train, rose_train, sunflower_train, tulip_train])\n",
    "val_files = np.concatenate([daisy_val, dandelion_val, rose_val, sunflower_val, tulip_val])\n",
    "test_files = np.concatenate([daisy_test, dandelion_test, rose_test, sunflower_test, tulip_test])\n",
    "\n",
    "np.random.shuffle(train_files) #Embaralhar os arquivos de treino\n",
    "np.random.shuffle(val_files) #Embaralhar os arquivos de validação\n",
    "np.random.shuffle(test_files) #Embaralhar os arquivos de teste"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vamos importar algumas bibliotecas importantes para montar o dataset. Trata-se de funções do keras, biblioteca famosa de *deeplearning*, mas poderiamos utilizar métodos de outras bibliotecas:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt #Biblioteca para criar os gráficos de análise\n",
    "from keras.preprocessing.image import ImageDataGenerator, load_img, img_to_array, array_to_img\n",
    "#Funções de processamento de imagens do keras\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;A seguir precisamos definir algo extremamente importante, os *labels*, o gabarito, sem isso a rede não aprende qual imagem corresponde à qual espécie. Com a manipulação feita anteriormente, em que colocamos o nome das espécies como nome dos arquivos *.jpg*, assim sabemos exatamente à qual classe corresponde cada imagem, vamos ler a imagem, tranformar ela em um array para o computador processar, e vamos ler o nome da mesma, extraíndo a informação do nome da classe:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train dataset shape: (3028, 100, 100, 3)\n",
      "Validation dataset shape: (647, 100, 100, 3)\n"
     ]
    }
   ],
   "source": [
    "train_imgs = [img_to_array(load_img(img, target_size=img_shape)) for img in train_files]\n",
    "#Aqui para cada imagem na lista *train_files* salvamos a matriz correspondente usando a função\n",
    "#img_to_array do keras.\n",
    "train_labels = [nome.split('/')[8].strip() for nome in train_files]\n",
    "#Aqui vamos usar o método split() de strings, selecionar o oitavo elemento da lista que ele \n",
    "#retorna, acredite em mim este elemento é o nome da classe, e vamos salvá-lo na lista de labels.\n",
    "\n",
    "#Repetimos o processo para o teste e a validação:\n",
    "val_imgs = [img_to_array(load_img(img, target_size=img_shape)) for img in val_files]\n",
    "val_labels = [nome.split('/')[8].strip() for nome in val_files]\n",
    "\n",
    "train_imgs = np.array(train_imgs)\n",
    "val_imgs = np.array(val_imgs)\n",
    "\n",
    "print('Train dataset shape:', train_imgs.shape)\n",
    "print('Validation dataset shape:', val_imgs.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Temos um dataset de treino com modestas 3028 imagnes, de tamanho 100x100 pixels e composta de 3 canais, já que vamos trabalhar com imagens RGB. A seguir precisamos normalizar elas, já que as redes convolucionais trabalham melhor com números entre 0 e 1:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(100, 100, 3)\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<PIL.Image.Image image mode=RGB size=100x100 at 0x7FDE13E2CD90>"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train_imgs_scaled = train_imgs.astype('float32')\n",
    "val_imgs_scaled  = val_imgs.astype('float32')\n",
    "train_imgs_scaled /= 255\n",
    "val_imgs_scaled /= 255\n",
    "\n",
    "#Vamos printar uma amostra:\n",
    "print(train_imgs[2].shape)\n",
    "array_to_img(train_imgs[2])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Montando uma CNN básica"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Agora que temos um dataset bem definido, com imagens prontas para serem passadas à uma rede, vamos começar a organizar nosso primeiro modelo de análise: uma CNN pura, sem regularização ou augumentation. Para isso vamos definir o batch_size, o número de classes, as epochs e o formato do imput. Outra coisa que precisamos fazer é converter os labels para um formato que a rede compreenda, ela não entende strings, vamos usar o labelencoder, que vai atribuir numeros para cada classe:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "['daisy', 'sunflower', 'tulip', 'rose', 'tulip', 'rose', 'sunflower', 'sunflower', 'daisy', 'daisy'] [0 3 4 2 4 2 3 3 0 0]\n"
     ]
    }
   ],
   "source": [
    "batch_size = 30\n",
    "num_classes = 5\n",
    "epochs = 10  #Vamos começar com poucas epochs pois sei que é o suficiente neste modelo, sinta-se\n",
    "#a vontade para testar com mais.\n",
    "input_shape = (100, 100, 3)\n",
    "\n",
    "#Importando a biblioteca do sklearn para lidar com as categorias\n",
    "from sklearn.preprocessing import LabelEncoder\n",
    "\n",
    "lb = LabelEncoder()  #Criando o objeto\n",
    "lb.fit(train_labels)  #Fitando ao treino\n",
    "train_labels_enc = lb.transform(train_labels)  #Transforma o treino\n",
    "val_labels_enc = lb.transform(val_labels)  #Transforma a validação\n",
    "\n",
    "#Vamos analisar alguns labels, para ver se está tudo certo:\n",
    "print(train_labels[:10], train_labels_enc[:10])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Finalmente podemos criar o modelo, vou recomendar importar do tensorflow, mas pode ser do próprio keras se quiser. Vamos definir a rede de forma mais simples possível, com 3 camadas convolucionais, com números crescentes de filtros, ativação relu, mas pode testar com sigmoid, e maxpooling. As fully-conected vamos com uma de 512 categorias, e uma com 5, para classificar. Compilar o modelo com a sparse_categorical_crossentropy:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"sequential\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d (Conv2D)              (None, 98, 98, 16)        448       \n",
      "_________________________________________________________________\n",
      "max_pooling2d (MaxPooling2D) (None, 49, 49, 16)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_1 (Conv2D)            (None, 47, 47, 64)        9280      \n",
      "_________________________________________________________________\n",
      "max_pooling2d_1 (MaxPooling2 (None, 23, 23, 64)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_2 (Conv2D)            (None, 21, 21, 128)       73856     \n",
      "_________________________________________________________________\n",
      "max_pooling2d_2 (MaxPooling2 (None, 10, 10, 128)       0         \n",
      "_________________________________________________________________\n",
      "flatten (Flatten)            (None, 12800)             0         \n",
      "_________________________________________________________________\n",
      "dense (Dense)                (None, 512)               6554112   \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 5)                 2565      \n",
      "=================================================================\n",
      "Total params: 6,640,261\n",
      "Trainable params: 6,640,261\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout\n",
    "from tensorflow.keras.models import Sequential\n",
    "from tensorflow.keras import optimizers\n",
    "\n",
    "model = Sequential()\n",
    "\n",
    "model.add(Conv2D(16, kernel_size=(3, 3), activation='relu', \n",
    "                 input_shape=input_shape))\n",
    "model.add(MaxPooling2D(pool_size=(2, 2)))\n",
    "\n",
    "model.add(Conv2D(64, kernel_size=(3, 3), activation='relu'))\n",
    "model.add(MaxPooling2D(pool_size=(2, 2)))\n",
    "\n",
    "model.add(Conv2D(128, kernel_size=(3, 3), activation='relu'))\n",
    "model.add(MaxPooling2D(pool_size=(2, 2)))\n",
    "\n",
    "model.add(Flatten())\n",
    "model.add(Dense(512, activation='relu'))\n",
    "model.add(Dense(num_classes, activation='sigmoid'))\n",
    "\n",
    "\n",
    "model.compile(loss='sparse_categorical_crossentropy',\n",
    "              optimizer=optimizers.RMSprop(),\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Tudo pronto, vamos utilizar o método fit nas imagens escalonadas que criamos,os labels codificados e a validação que temos:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 3028 samples, validate on 647 samples\n",
      "Epoch 1/10\n",
      "3028/3028 [==============================] - 22s 7ms/sample - loss: 1.3944 - accuracy: 0.3804 - val_loss: 1.1441 - val_accuracy: 0.4822\n",
      "Epoch 2/10\n",
      "3028/3028 [==============================] - 21s 7ms/sample - loss: 1.1056 - accuracy: 0.5294 - val_loss: 1.1581 - val_accuracy: 0.5533\n",
      "Epoch 3/10\n",
      "3028/3028 [==============================] - 22s 7ms/sample - loss: 0.9545 - accuracy: 0.6143 - val_loss: 0.9860 - val_accuracy: 0.6074\n",
      "Epoch 4/10\n",
      "3028/3028 [==============================] - 21s 7ms/sample - loss: 0.8281 - accuracy: 0.6711 - val_loss: 0.9590 - val_accuracy: 0.6244\n",
      "Epoch 5/10\n",
      "3028/3028 [==============================] - 21s 7ms/sample - loss: 0.6966 - accuracy: 0.7292 - val_loss: 1.0352 - val_accuracy: 0.6352\n",
      "Epoch 6/10\n",
      "3028/3028 [==============================] - 21s 7ms/sample - loss: 0.5499 - accuracy: 0.7850 - val_loss: 1.0515 - val_accuracy: 0.6337\n",
      "Epoch 7/10\n",
      "3028/3028 [==============================] - 20s 7ms/sample - loss: 0.4356 - accuracy: 0.8303 - val_loss: 1.1538 - val_accuracy: 0.6182\n",
      "Epoch 8/10\n",
      "3028/3028 [==============================] - 21s 7ms/sample - loss: 0.3204 - accuracy: 0.8818 - val_loss: 1.1815 - val_accuracy: 0.6461\n",
      "Epoch 9/10\n",
      "3028/3028 [==============================] - 22s 7ms/sample - loss: 0.2415 - accuracy: 0.9184 - val_loss: 1.8479 - val_accuracy: 0.5595\n",
      "Epoch 10/10\n",
      "3028/3028 [==============================] - 22s 7ms/sample - loss: 0.1830 - accuracy: 0.9326 - val_loss: 1.4124 - val_accuracy: 0.6445\n"
     ]
    }
   ],
   "source": [
    "history = model.fit(x=train_imgs_scaled, y=train_labels_enc,\n",
    "                    validation_data = (val_imgs_scaled, val_labels_enc),\n",
    "                    batch_size=batch_size,\n",
    "                    epochs=epochs,\n",
    "                    verbose=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Fazendo algo que poderia ter feito antes, mas definindo aqui as imagens do teste:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Test dataset shape: (648, 100, 100, 3)\n"
     ]
    }
   ],
   "source": [
    "test_imgs = [img_to_array(load_img(img, target_size=img_shape)) for img in test_files]\n",
    "test_labels = [fn.split('/')[8].strip() for fn in test_files]\n",
    "test_imgs = np.array(test_imgs)\n",
    "test_labels_enc = lb.transform(test_labels)\n",
    "test_imgs_scaled = test_imgs.astype('float32')\n",
    "test_imgs_scaled /= 255\n",
    "\n",
    "print('Test dataset shape:', test_imgs.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vamos salvar aqui este modelo para analisar posteriormente o teste com os outros modelos:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.save('Modelo_Básico_CNN.h5')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vamos plotar o gráfico com o desempenho no treino e na validação do modelo, com o seguinte código retirado do Medium usado como referência:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "f, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 4))\n",
    "t = f.suptitle('Performance com uma CNN padrão', fontsize=12)\n",
    "f.subplots_adjust(top=0.85, wspace=0.3)\n",
    "\n",
    "epoch_list = list(range(1,11))\n",
    "ax1.plot(epoch_list, history.history['accuracy'], label='Acurácia do treino')\n",
    "ax1.plot(epoch_list, history.history['val_accuracy'], label='Acurácia na validação')\n",
    "ax1.set_xticks(np.arange(0, 11, 2))\n",
    "ax1.set_ylabel('Acurácia')\n",
    "ax1.set_xlabel('Epoch')\n",
    "ax1.set_title('Acurácia')\n",
    "l1 = ax1.legend(loc=\"best\")\n",
    "\n",
    "ax2.plot(epoch_list, history.history['loss'], label='Loss Treino')\n",
    "ax2.plot(epoch_list, history.history['val_loss'], label='Loss Validação')\n",
    "ax2.set_xticks(np.arange(0, 11, 2))\n",
    "ax2.set_ylabel('Valor do Loss')\n",
    "ax2.set_xlabel('Epoch')\n",
    "ax2.set_title('Loss')\n",
    "l2 = ax2.legend(loc=\"best\")\n",
    "\n",
    "ax1.grid()\n",
    "ax2.grid()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vemos que a acurácia na validação ficou em torno de 65%, inclusive pela curva é visível que a rede foi à um *overfit* completo. Vamos tentar melhorar um pouco o overfit utilizando algumas técnicas de regularização e augumentation."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### CNN com regularização e augumentation"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Nesta nova rede vamos utilizar apenas o dropout, adicionar uma nova camada densa eintroduzir mais uma camada de convolução. No import eu até coloquei as funções de Batch Normalization e Gaussian Noise, mas considero que são desnecessários e podem levar a rede à não conseguir aprender nada, mas sinta-se livre para testar."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"sequential_1\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_3 (Conv2D)            (None, 98, 98, 16)        448       \n",
      "_________________________________________________________________\n",
      "re_lu (ReLU)                 (None, 98, 98, 16)        0         \n",
      "_________________________________________________________________\n",
      "max_pooling2d_3 (MaxPooling2 (None, 49, 49, 16)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_4 (Conv2D)            (None, 47, 47, 128)       18560     \n",
      "_________________________________________________________________\n",
      "re_lu_1 (ReLU)               (None, 47, 47, 128)       0         \n",
      "_________________________________________________________________\n",
      "max_pooling2d_4 (MaxPooling2 (None, 23, 23, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_5 (Conv2D)            (None, 21, 21, 128)       147584    \n",
      "_________________________________________________________________\n",
      "re_lu_2 (ReLU)               (None, 21, 21, 128)       0         \n",
      "_________________________________________________________________\n",
      "max_pooling2d_5 (MaxPooling2 (None, 10, 10, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_6 (Conv2D)            (None, 8, 8, 64)          73792     \n",
      "_________________________________________________________________\n",
      "re_lu_3 (ReLU)               (None, 8, 8, 64)          0         \n",
      "_________________________________________________________________\n",
      "max_pooling2d_6 (MaxPooling2 (None, 4, 4, 64)          0         \n",
      "_________________________________________________________________\n",
      "flatten_1 (Flatten)          (None, 1024)              0         \n",
      "_________________________________________________________________\n",
      "dense_2 (Dense)              (None, 512)               524800    \n",
      "_________________________________________________________________\n",
      "dropout (Dropout)            (None, 512)               0         \n",
      "_________________________________________________________________\n",
      "dense_3 (Dense)              (None, 100)               51300     \n",
      "_________________________________________________________________\n",
      "dropout_1 (Dropout)          (None, 100)               0         \n",
      "_________________________________________________________________\n",
      "dense_4 (Dense)              (None, 5)                 505       \n",
      "=================================================================\n",
      "Total params: 816,989\n",
      "Trainable params: 816,989\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "from tensorflow.keras.layers import ELU, Dropout, BatchNormalization, GaussianNoise, LeakyReLU, ReLU\n",
    "\n",
    "DROPOUT_RATE = 0.3   #Taxa de desligamento dos neurônios\n",
    "size = [64, 128, 128] #Tamanho para as novas camadas convolucionais\n",
    "\n",
    "model = Sequential()\n",
    "\n",
    "model.add(Conv2D(16, kernel_size=(3, 3), input_shape=input_shape))\n",
    "model.add(ReLU())         \n",
    "model.add(MaxPooling2D((2, 2)))\n",
    "for _ in range(3):  \n",
    "    model.add(Conv2D(size.pop(), (3, 3)))\n",
    "    model.add(ReLU())\n",
    "    model.add(MaxPooling2D((2, 2)))\n",
    "    \n",
    "model.add(Flatten())\n",
    "model.add(Dense(512, activation = 'relu'))\n",
    "model.add(Dropout(DROPOUT_RATE))\n",
    "model.add(Dense(100, activation = 'relu'))\n",
    "model.add(Dropout(DROPOUT_RATE))\n",
    "model.add(Dense(num_classes, activation='sigmoid'))\n",
    "\n",
    "\n",
    "model.compile(loss='sparse_categorical_crossentropy',\n",
    "              optimizer=optimizers.RMSprop(),\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vamos utilizar um método de Augumentation em tempo real, ou seja, o nosso modelo vai ser alimentado com diferentes imagens toda epoch, e para isso usamos o ImageDataGenerator do tensorflow, alguns parâmetros estão definidos abaixo:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Preprocessamento em tempo real de augumentation:\n",
    "datagen_tr = ImageDataGenerator(\n",
    "    rotation_range=40,  #Rotaciona a imagem de forma randomica (graus, 0 a 180)\n",
    "    # Aleatórimente faz o shitf da imagem em fraçao, horizontalmente \n",
    "    width_shift_range=0.2,\n",
    "    # Aleatórimente faz o shitf da imagem em fraçao, verticalmente \n",
    "    height_shift_range=0.2,\n",
    "    # Adiciona um zoom na imagem\n",
    "    zoom_range=0.2\n",
    ")\n",
    "\n",
    "datagen_tr.fit(train_imgs_scaled)\n",
    "\n",
    "datagen_val = ImageDataGenerator()\n",
    "\n",
    "datagen_val.fit(val_imgs_scaled)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Como prentede-se reduzir o overfit, dificultando o aprendizado da rede, é necessário que o treinamento seja mais longo, para que ela tenha tempo de aprender as características do dataset, por isso vamos utilizar aqui 50 epochs."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "epochs = 50"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Fazendo o fit do modelo com o gerador de imagens:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:tensorflow:sample_weight modes were coerced from\n",
      "  ...\n",
      "    to  \n",
      "  ['...']\n",
      "WARNING:tensorflow:sample_weight modes were coerced from\n",
      "  ...\n",
      "    to  \n",
      "  ['...']\n",
      "Train for 101 steps, validate for 33 steps\n",
      "Epoch 1/50\n",
      "101/101 [==============================] - 30s 292ms/step - loss: 1.4063 - accuracy: 0.3649 - val_loss: 1.2699 - val_accuracy: 0.4699\n",
      "Epoch 2/50\n",
      "101/101 [==============================] - 32s 321ms/step - loss: 1.2363 - accuracy: 0.4617 - val_loss: 1.1307 - val_accuracy: 0.5425\n",
      "Epoch 3/50\n",
      "101/101 [==============================] - 32s 322ms/step - loss: 1.1558 - accuracy: 0.5188 - val_loss: 1.1822 - val_accuracy: 0.5301\n",
      "Epoch 4/50\n",
      "101/101 [==============================] - 33s 324ms/step - loss: 1.0794 - accuracy: 0.5528 - val_loss: 0.9732 - val_accuracy: 0.6291\n",
      "Epoch 5/50\n",
      "101/101 [==============================] - 31s 310ms/step - loss: 1.0391 - accuracy: 0.5898 - val_loss: 1.0699 - val_accuracy: 0.6105\n",
      "Epoch 6/50\n",
      "101/101 [==============================] - 31s 308ms/step - loss: 0.9996 - accuracy: 0.6212 - val_loss: 0.9414 - val_accuracy: 0.6491\n",
      "Epoch 7/50\n",
      "101/101 [==============================] - 31s 311ms/step - loss: 0.9483 - accuracy: 0.6298 - val_loss: 0.8677 - val_accuracy: 0.6816\n",
      "Epoch 8/50\n",
      "101/101 [==============================] - 30s 301ms/step - loss: 0.9314 - accuracy: 0.6417 - val_loss: 0.9028 - val_accuracy: 0.6383\n",
      "Epoch 9/50\n",
      "101/101 [==============================] - 31s 310ms/step - loss: 0.9152 - accuracy: 0.6569 - val_loss: 0.8241 - val_accuracy: 0.6924\n",
      "Epoch 10/50\n",
      "101/101 [==============================] - 32s 312ms/step - loss: 0.8814 - accuracy: 0.6694 - val_loss: 0.8845 - val_accuracy: 0.6615\n",
      "Epoch 11/50\n",
      "101/101 [==============================] - 32s 318ms/step - loss: 0.8499 - accuracy: 0.6793 - val_loss: 0.9684 - val_accuracy: 0.6507\n",
      "Epoch 12/50\n",
      "101/101 [==============================] - 32s 316ms/step - loss: 0.8261 - accuracy: 0.6859 - val_loss: 0.8597 - val_accuracy: 0.6847\n",
      "Epoch 13/50\n",
      "101/101 [==============================] - 32s 320ms/step - loss: 0.8255 - accuracy: 0.6919 - val_loss: 0.8187 - val_accuracy: 0.6971\n",
      "Epoch 14/50\n",
      "101/101 [==============================] - 32s 316ms/step - loss: 0.7909 - accuracy: 0.6995 - val_loss: 0.8148 - val_accuracy: 0.7017\n",
      "Epoch 15/50\n",
      "101/101 [==============================] - 32s 312ms/step - loss: 0.7800 - accuracy: 0.7074 - val_loss: 0.8475 - val_accuracy: 0.7048\n",
      "Epoch 16/50\n",
      "101/101 [==============================] - 31s 303ms/step - loss: 0.7834 - accuracy: 0.7104 - val_loss: 0.7611 - val_accuracy: 0.7295\n",
      "Epoch 17/50\n",
      "101/101 [==============================] - 31s 306ms/step - loss: 0.7476 - accuracy: 0.7289 - val_loss: 0.9658 - val_accuracy: 0.6352\n",
      "Epoch 18/50\n",
      "101/101 [==============================] - 32s 317ms/step - loss: 0.7490 - accuracy: 0.7242 - val_loss: 0.8491 - val_accuracy: 0.7172\n",
      "Epoch 19/50\n",
      "101/101 [==============================] - 31s 309ms/step - loss: 0.7275 - accuracy: 0.7365 - val_loss: 0.8478 - val_accuracy: 0.7156\n",
      "Epoch 20/50\n",
      "101/101 [==============================] - 31s 311ms/step - loss: 0.7292 - accuracy: 0.7308 - val_loss: 0.8237 - val_accuracy: 0.7079\n",
      "Epoch 21/50\n",
      "101/101 [==============================] - 32s 320ms/step - loss: 0.7247 - accuracy: 0.7401 - val_loss: 0.8501 - val_accuracy: 0.7311\n",
      "Epoch 22/50\n",
      "101/101 [==============================] - 32s 312ms/step - loss: 0.7211 - accuracy: 0.7391 - val_loss: 0.9748 - val_accuracy: 0.6847\n",
      "Epoch 23/50\n",
      "101/101 [==============================] - 32s 313ms/step - loss: 0.6960 - accuracy: 0.7483 - val_loss: 0.9473 - val_accuracy: 0.7342\n",
      "Epoch 24/50\n",
      "101/101 [==============================] - 32s 315ms/step - loss: 0.7129 - accuracy: 0.7384 - val_loss: 0.7685 - val_accuracy: 0.7589\n",
      "Epoch 25/50\n",
      "101/101 [==============================] - 32s 320ms/step - loss: 0.6885 - accuracy: 0.7477 - val_loss: 0.7816 - val_accuracy: 0.7527\n",
      "Epoch 26/50\n",
      "101/101 [==============================] - 32s 312ms/step - loss: 0.6790 - accuracy: 0.7556 - val_loss: 0.6475 - val_accuracy: 0.7558\n",
      "Epoch 27/50\n",
      "101/101 [==============================] - 31s 309ms/step - loss: 0.7005 - accuracy: 0.7553 - val_loss: 0.9660 - val_accuracy: 0.7372\n",
      "Epoch 28/50\n",
      "101/101 [==============================] - 32s 317ms/step - loss: 0.6868 - accuracy: 0.7635 - val_loss: 0.7677 - val_accuracy: 0.7543\n",
      "Epoch 29/50\n",
      "101/101 [==============================] - 33s 325ms/step - loss: 0.7009 - accuracy: 0.7517 - val_loss: 0.7397 - val_accuracy: 0.7357\n",
      "Epoch 30/50\n",
      "101/101 [==============================] - 32s 314ms/step - loss: 0.6764 - accuracy: 0.7573 - val_loss: 0.7874 - val_accuracy: 0.7342\n",
      "Epoch 31/50\n",
      "101/101 [==============================] - 32s 313ms/step - loss: 0.6785 - accuracy: 0.7602 - val_loss: 0.6763 - val_accuracy: 0.7929\n",
      "Epoch 32/50\n",
      "101/101 [==============================] - 33s 329ms/step - loss: 0.6975 - accuracy: 0.7480 - val_loss: 0.8329 - val_accuracy: 0.7372\n",
      "Epoch 33/50\n",
      "101/101 [==============================] - 32s 320ms/step - loss: 0.7049 - accuracy: 0.7457 - val_loss: 0.8223 - val_accuracy: 0.7512\n",
      "Epoch 34/50\n",
      "101/101 [==============================] - 34s 333ms/step - loss: 0.7131 - accuracy: 0.7493 - val_loss: 0.8322 - val_accuracy: 0.7682\n",
      "Epoch 35/50\n",
      "101/101 [==============================] - 31s 303ms/step - loss: 0.6828 - accuracy: 0.7602 - val_loss: 0.8607 - val_accuracy: 0.7017\n",
      "Epoch 36/50\n",
      "101/101 [==============================] - 32s 314ms/step - loss: 0.7283 - accuracy: 0.7507 - val_loss: 0.8794 - val_accuracy: 0.7465\n",
      "Epoch 37/50\n",
      "101/101 [==============================] - 33s 324ms/step - loss: 0.6937 - accuracy: 0.7483 - val_loss: 0.7474 - val_accuracy: 0.7450\n",
      "Epoch 38/50\n",
      "101/101 [==============================] - 32s 314ms/step - loss: 0.7019 - accuracy: 0.7668 - val_loss: 0.9821 - val_accuracy: 0.7512\n",
      "Epoch 39/50\n",
      "101/101 [==============================] - 32s 318ms/step - loss: 0.6840 - accuracy: 0.7556 - val_loss: 1.2427 - val_accuracy: 0.7713\n",
      "Epoch 40/50\n",
      "101/101 [==============================] - 32s 318ms/step - loss: 0.7570 - accuracy: 0.7477 - val_loss: 0.8026 - val_accuracy: 0.7512\n",
      "Epoch 41/50\n",
      "101/101 [==============================] - 32s 318ms/step - loss: 0.7662 - accuracy: 0.7520 - val_loss: 0.7983 - val_accuracy: 0.7774\n",
      "Epoch 42/50\n",
      "101/101 [==============================] - 32s 312ms/step - loss: 0.7679 - accuracy: 0.7348 - val_loss: 0.7848 - val_accuracy: 0.7481\n",
      "Epoch 43/50\n",
      "101/101 [==============================] - 31s 305ms/step - loss: 0.7461 - accuracy: 0.7421 - val_loss: 0.9928 - val_accuracy: 0.6692\n",
      "Epoch 44/50\n",
      "101/101 [==============================] - 31s 306ms/step - loss: 0.7476 - accuracy: 0.7361 - val_loss: 0.8838 - val_accuracy: 0.7635\n",
      "Epoch 45/50\n",
      "101/101 [==============================] - 30s 299ms/step - loss: 0.7948 - accuracy: 0.7381 - val_loss: 0.8401 - val_accuracy: 0.7048\n",
      "Epoch 46/50\n",
      "101/101 [==============================] - 31s 307ms/step - loss: 0.8005 - accuracy: 0.7381 - val_loss: 1.2921 - val_accuracy: 0.6600\n",
      "Epoch 47/50\n",
      "101/101 [==============================] - 31s 308ms/step - loss: 0.8006 - accuracy: 0.7421 - val_loss: 0.8687 - val_accuracy: 0.7264\n",
      "Epoch 48/50\n",
      "101/101 [==============================] - 30s 301ms/step - loss: 0.8213 - accuracy: 0.7348 - val_loss: 0.8233 - val_accuracy: 0.7280\n",
      "Epoch 49/50\n",
      "101/101 [==============================] - 31s 308ms/step - loss: 0.8007 - accuracy: 0.7414 - val_loss: 0.9944 - val_accuracy: 0.7218\n",
      "Epoch 50/50\n",
      "101/101 [==============================] - 32s 314ms/step - loss: 0.7965 - accuracy: 0.7401 - val_loss: 0.8511 - val_accuracy: 0.7388\n"
     ]
    }
   ],
   "source": [
    "history = model.fit(datagen_tr.flow(train_imgs_scaled, train_labels_enc, batch_size=30),\n",
    "                    validation_data=datagen_val.flow(val_imgs_scaled, val_labels_enc, batch_size=20),\n",
    "                    epochs=epochs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.save('Modelo_Reg_Aug_CNN.h5') #Salvando o modelo"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "f, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 4))\n",
    "t = f.suptitle('Performance com uma CNN com regularização e augumentation', fontsize=12)\n",
    "f.subplots_adjust(top=0.85, wspace=0.3)\n",
    "\n",
    "epoch_list = list(range(1, 51))\n",
    "ax1.plot(epoch_list, history.history['accuracy'], label='Acurácia do treino')\n",
    "ax1.plot(epoch_list, history.history['val_accuracy'], label='Acurácia na validação')\n",
    "ax1.set_xticks(np.arange(0, 51, 5))\n",
    "ax1.set_ylabel('Acurácia')\n",
    "ax1.set_xlabel('Epoch')\n",
    "ax1.set_title('Acurácia')\n",
    "l1 = ax1.legend(loc=\"best\")\n",
    "\n",
    "ax2.plot(epoch_list, history.history['loss'], label='Loss Treino')\n",
    "ax2.plot(epoch_list, history.history['val_loss'], label='Loss Validação')\n",
    "ax2.set_xticks(np.arange(0, 51, 5))\n",
    "ax2.set_ylabel('Valor do Loss')\n",
    "ax2.set_xlabel('Epoch')\n",
    "ax2.set_title('Loss')\n",
    "l2 = ax2.legend(loc=\"best\")\n",
    "\n",
    "ax1.grid()\n",
    "ax2.grid()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Aqui vemos como a valição e o treino se acompanham, de fato conseguimos evitar o overfit e elevamos bastante a acurácia da validação, aumentando a capacidade de generalização do modelo. A loss varia bastante mas isso é normal e uma learning rate menor pode suavizar o processo. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Transfer Learning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Por fim chegamos ao tema em sí deste notebook, vamos tratar do transfer learning. Primeiramente vamos importar a rede que usaremos como transfer: a VGG-16 treinada com o imagenet. Para maiores informações sobre ela, o leitor pode voltar na parte que passei os links sobre as redes estado da arte. Aqui vamos importá-la sem as camadas densas:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tensor(\"block5_pool_1/Identity:0\", shape=(None, 3, 3, 512), dtype=float32)\n"
     ]
    }
   ],
   "source": [
    "from tensorflow.keras.applications import vgg16\n",
    "from tensorflow.keras.models import Model\n",
    "import tensorflow.keras\n",
    "\n",
    "vgg = vgg16.VGG16(include_top=False,  #Parâmetro responsável por incluir ou descartar as camadas Fully-connected\n",
    "                  weights='imagenet', \n",
    "                  input_shape=input_shape)\n",
    "\n",
    "print(vgg.layers[-1].output)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Pela saída acima, como mandamos o *import* não trazer as camadas FC, temos que a última camada da VGG que importamos é a última camada de *maxpooling* do último bloco de convolução da rede, ficou confuso? A seguinte imagem vai te ajudar a compreender o que digo, prete atenção ao nome da camada acima, e os nomes na imagem:\n",
    "\n",
    "<img src = \"https://miro.medium.com/max/1210/1*E7zhhan7Sp7hats4jkKdeA.png\">\n",
    "\n",
    "Fonte:[Medium](https://towardsdatascience.com/a-comprehensive-hands-on-guide-to-transfer-learning-with-real-world-applications-in-deep-learning-212bf3b2f27a)\n",
    "\n",
    "&emsp;De forma que temos tudo que está na imagem até a última camada de *pooling*, precisamos agora formar o nosso modelo, adicionando a *flatten* e as *fully-connected*:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tensor(\"flatten_2/Identity:0\", shape=(None, 4608), dtype=float32)\n"
     ]
    }
   ],
   "source": [
    "saida = vgg.layers[-1].output           #Última camada presente no nosso modelo\n",
    "saida = tensorflow.keras.layers.Flatten()(saida)  \n",
    "vgg_model = Model(vgg.input, saida)    #Concatenando o modelo que temos com o *flatten*\n",
    "print(saida)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Com a manipulação acima conseguimos adicionar o *flatten* ao nosso novo modelo, falta adicionar as camadas FC, mas como prendemos fixar os pesos para todas as camadas até aqui, vamos fazer isso primeiro, que fica mais fácil:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [],
   "source": [
    "vgg_model.trainable = False\n",
    "for layer in vgg_model.layers:\n",
    "    layer.trainable = False"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;A seguir vamos criar as camadas *fully-connected* e ligá-las à nossa rede:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"sequential_2\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "dense_5 (Dense)              (None, 512)               2359808   \n",
      "_________________________________________________________________\n",
      "dense_6 (Dense)              (None, 512)               262656    \n",
      "_________________________________________________________________\n",
      "dense_7 (Dense)              (None, 5)                 2565      \n",
      "=================================================================\n",
      "Total params: 2,625,029\n",
      "Trainable params: 2,625,029\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "from tensorflow.keras.layers import InputLayer\n",
    "\n",
    "input_shape = vgg_model.output_shape[1]\n",
    "\n",
    "model = Sequential()\n",
    "model.add(InputLayer(input_shape=(input_shape,)))\n",
    "model.add(Dense(512, activation='relu', input_dim=input_shape))\n",
    "model.add(Dense(512, activation='relu'))\n",
    "model.add(Dense(num_classes, activation='sigmoid'))\n",
    "\n",
    "model.compile(loss='sparse_categorical_crossentropy',\n",
    "              optimizer=optimizers.RMSprop(lr=1e-4),\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Para facilitar, não vamos unir ambas as redes neste ponto, um bom motivo para isso é que demora muito tempo para passar as imagens pela arquitetura da VGG com as camadas *fully-connected* junto, por outro lado se processar as imagens pelas camadas convolucionais antes, e depois passar pela rede é muito mais e rápido, pois fazemos isto apenas uma vez, ao invés de fazer 'epoch' vezes. Além disso se algo ocorrer é melhor ter a saída deste modelo salva para que não seja necessário ficar rodando constantemente isso. No seu computador pessoal isso pode parecer besteira, mas em uma cloud ficar rodando várias vezes a mesma coisa é perder dinheiro."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Imagens de treino pela VGG: (3028, 4608)\n",
      "Imagens de validaação pela VGG: (647, 4608)\n"
     ]
    }
   ],
   "source": [
    "def passar_pela_vgg(model, input_img):  #Função responsável por passar a imagem pela VGG-16\n",
    "    saida = model.predict(input_img, verbose=0)\n",
    "    return saida\n",
    "    \n",
    "train_vgg = passar_pela_vgg(vgg_model, train_imgs_scaled)\n",
    "val_vgg = passar_pela_vgg(vgg_model, val_imgs_scaled)\n",
    "\n",
    "print('Imagens de treino pela VGG:', train_vgg.shape)\n",
    "print('Imagens de validaação pela VGG:', val_vgg.shape)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Com as imagens prontas e o modelo de camadas totalmente conectadas pronto, vamos para o método *fit*. Lembrando que aqui não vamos valer de regularização e de augumentation, primeiro vamos analisar o desempenho puro da VGG."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 3028 samples, validate on 647 samples\n",
      "Epoch 1/30\n",
      "3028/3028 [==============================] - 3s 1ms/sample - loss: 1.0785 - accuracy: 0.5938 - val_loss: 0.8167 - val_accuracy: 0.7032\n",
      "Epoch 2/30\n",
      "3028/3028 [==============================] - 3s 854us/sample - loss: 0.7243 - accuracy: 0.7328 - val_loss: 0.7102 - val_accuracy: 0.7372\n",
      "Epoch 3/30\n",
      "3028/3028 [==============================] - 3s 872us/sample - loss: 0.5895 - accuracy: 0.7906 - val_loss: 0.6934 - val_accuracy: 0.7403\n",
      "Epoch 4/30\n",
      "3028/3028 [==============================] - 3s 835us/sample - loss: 0.4956 - accuracy: 0.8243 - val_loss: 0.6441 - val_accuracy: 0.7774\n",
      "Epoch 5/30\n",
      "3028/3028 [==============================] - 3s 827us/sample - loss: 0.4268 - accuracy: 0.8550 - val_loss: 0.6606 - val_accuracy: 0.7573\n",
      "Epoch 6/30\n",
      "3028/3028 [==============================] - 3s 869us/sample - loss: 0.3642 - accuracy: 0.8752 - val_loss: 0.6309 - val_accuracy: 0.7867\n",
      "Epoch 7/30\n",
      "3028/3028 [==============================] - 3s 862us/sample - loss: 0.3060 - accuracy: 0.9019 - val_loss: 0.7527 - val_accuracy: 0.7496\n",
      "Epoch 8/30\n",
      "3028/3028 [==============================] - 3s 856us/sample - loss: 0.2539 - accuracy: 0.9191 - val_loss: 0.7476 - val_accuracy: 0.7527\n",
      "Epoch 9/30\n",
      "3028/3028 [==============================] - 3s 866us/sample - loss: 0.2179 - accuracy: 0.9346 - val_loss: 0.6912 - val_accuracy: 0.7774\n",
      "Epoch 10/30\n",
      "3028/3028 [==============================] - 3s 833us/sample - loss: 0.1841 - accuracy: 0.9389 - val_loss: 0.7087 - val_accuracy: 0.7805\n",
      "Epoch 11/30\n",
      "3028/3028 [==============================] - 2s 823us/sample - loss: 0.1461 - accuracy: 0.9584 - val_loss: 0.7258 - val_accuracy: 0.7743\n",
      "Epoch 12/30\n",
      "3028/3028 [==============================] - 2s 823us/sample - loss: 0.1196 - accuracy: 0.9666 - val_loss: 0.8739 - val_accuracy: 0.7450\n",
      "Epoch 13/30\n",
      "3028/3028 [==============================] - 3s 861us/sample - loss: 0.1011 - accuracy: 0.9742 - val_loss: 0.8097 - val_accuracy: 0.7790\n",
      "Epoch 14/30\n",
      "3028/3028 [==============================] - 3s 882us/sample - loss: 0.0779 - accuracy: 0.9785 - val_loss: 0.8200 - val_accuracy: 0.7759\n",
      "Epoch 15/30\n",
      "3028/3028 [==============================] - 3s 903us/sample - loss: 0.0642 - accuracy: 0.9812 - val_loss: 0.9319 - val_accuracy: 0.7512\n",
      "Epoch 16/30\n",
      "3028/3028 [==============================] - 3s 855us/sample - loss: 0.0540 - accuracy: 0.9884 - val_loss: 0.8736 - val_accuracy: 0.7790\n",
      "Epoch 17/30\n",
      "3028/3028 [==============================] - 3s 831us/sample - loss: 0.0423 - accuracy: 0.9924 - val_loss: 0.9496 - val_accuracy: 0.7651\n",
      "Epoch 18/30\n",
      "3028/3028 [==============================] - 3s 843us/sample - loss: 0.0359 - accuracy: 0.9917 - val_loss: 0.9109 - val_accuracy: 0.7743\n",
      "Epoch 19/30\n",
      "3028/3028 [==============================] - 3s 854us/sample - loss: 0.0327 - accuracy: 0.9934 - val_loss: 1.0804 - val_accuracy: 0.7558\n",
      "Epoch 20/30\n",
      "3028/3028 [==============================] - 3s 833us/sample - loss: 0.0293 - accuracy: 0.9917 - val_loss: 1.0091 - val_accuracy: 0.7682\n",
      "Epoch 21/30\n",
      "3028/3028 [==============================] - 3s 886us/sample - loss: 0.0225 - accuracy: 0.9960 - val_loss: 1.0165 - val_accuracy: 0.7790\n",
      "Epoch 22/30\n",
      "3028/3028 [==============================] - 3s 837us/sample - loss: 0.0253 - accuracy: 0.9911 - val_loss: 1.0915 - val_accuracy: 0.7774\n",
      "Epoch 23/30\n",
      "3028/3028 [==============================] - 3s 863us/sample - loss: 0.0201 - accuracy: 0.9947 - val_loss: 1.0275 - val_accuracy: 0.7805\n",
      "Epoch 24/30\n",
      "3028/3028 [==============================] - 3s 871us/sample - loss: 0.0171 - accuracy: 0.9950 - val_loss: 1.0604 - val_accuracy: 0.7805\n",
      "Epoch 25/30\n",
      "3028/3028 [==============================] - 3s 863us/sample - loss: 0.0154 - accuracy: 0.9964 - val_loss: 1.5467 - val_accuracy: 0.7249\n",
      "Epoch 26/30\n",
      "3028/3028 [==============================] - 3s 864us/sample - loss: 0.0162 - accuracy: 0.9957 - val_loss: 1.1092 - val_accuracy: 0.7821\n",
      "Epoch 27/30\n",
      "3028/3028 [==============================] - 3s 852us/sample - loss: 0.0144 - accuracy: 0.9964 - val_loss: 1.1469 - val_accuracy: 0.7743\n",
      "Epoch 28/30\n",
      "3028/3028 [==============================] - 3s 853us/sample - loss: 0.0139 - accuracy: 0.9960 - val_loss: 1.2656 - val_accuracy: 0.7635\n",
      "Epoch 29/30\n",
      "3028/3028 [==============================] - 3s 840us/sample - loss: 0.0138 - accuracy: 0.9964 - val_loss: 1.7108 - val_accuracy: 0.7063\n",
      "Epoch 30/30\n",
      "3028/3028 [==============================] - 3s 890us/sample - loss: 0.0116 - accuracy: 0.9944 - val_loss: 1.5021 - val_accuracy: 0.7357\n"
     ]
    }
   ],
   "source": [
    "epochs = 30\n",
    "history = model.fit(x=train_vgg, y=train_labels_enc,\n",
    "                    validation_data=(val_vgg, val_labels_enc),\n",
    "                    batch_size=batch_size,\n",
    "                    epochs=epochs,\n",
    "                    verbose=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "f, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 4))\n",
    "t = f.suptitle('Performance com uma VGG-16 padrão', fontsize=12)\n",
    "f.subplots_adjust(top=0.85, wspace=0.3)\n",
    "\n",
    "epoch_list = list(range(1,31))\n",
    "ax1.plot(epoch_list, history.history['accuracy'], label='Acurácia do treino')\n",
    "ax1.plot(epoch_list, history.history['val_accuracy'], label='Acurácia na validação')\n",
    "ax1.set_xticks(np.arange(0, 31, 5))\n",
    "ax1.set_ylabel('Acurácia')\n",
    "ax1.set_xlabel('Epoch')\n",
    "ax1.set_title('Acurácia')\n",
    "l1 = ax1.legend(loc=\"best\")\n",
    "\n",
    "ax2.plot(epoch_list, history.history['loss'], label='Loss Treino')\n",
    "ax2.plot(epoch_list, history.history['val_loss'], label='Loss Validação')\n",
    "ax2.set_xticks(np.arange(0, 31, 5))\n",
    "ax2.set_ylabel('Valor do Loss')\n",
    "ax2.set_xlabel('Epoch')\n",
    "ax2.set_title('Loss')\n",
    "l2 = ax2.legend(loc=\"best\")\n",
    "\n",
    "ax1.grid()\n",
    "ax2.grid()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Visivelmente esse nosso transfer *off the shelf features* conseguiu atingir a mesma acurácia que o modelo com regularização e com augumentation, mas o detalhe é que este passou por um overfit também, logo estas técnicas podem ajudar a reduzir o mesmo, aumentando o desempenho geral da rede."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.save('Modelo_Transfer_Básico.h5')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Transfer Learning com Regularização e Augumentation"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;O resultado acima foi realmente muito bom, sem quase nenhum esforço já foi possível alcançar praticamente o mesmo resultado obtido pela nossa rede manual com regularização e augumentation, mas ainda não chegamos ao limite, claro, visto que o gráfico acima mostra que a rede leva ao overfitting, portanto podemos utilizar as técnicas descritas anteriormente para reduzir o mesmo e aumentar ainda mais o desempenho, vamos ver agora como fazer isso."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [],
   "source": [
    "# This will do preprocessing and realtime data augmentation:\n",
    "datagen_tr = ImageDataGenerator(\n",
    "    rotation_range=30,  # randomly rotate images in the range (degrees, 0 to 180)\n",
    "    # randomly shift images horizontally (fraction of total width)\n",
    "    width_shift_range=0.2,\n",
    "    # randomly shift images vertically (fraction of total height)\n",
    "    height_shift_range=0.2,\n",
    "    zoom_range=0.2)\n",
    "\n",
    "datagen_tr.fit(train_imgs_scaled)\n",
    "\n",
    "datagen_val = ImageDataGenerator()\n",
    "\n",
    "datagen_val.fit(val_imgs_scaled)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"sequential_3\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "model (Model)                (None, 4608)              14714688  \n",
      "_________________________________________________________________\n",
      "dense_8 (Dense)              (None, 512)               2359808   \n",
      "_________________________________________________________________\n",
      "re_lu_4 (ReLU)               (None, 512)               0         \n",
      "_________________________________________________________________\n",
      "dropout_2 (Dropout)          (None, 512)               0         \n",
      "_________________________________________________________________\n",
      "dense_9 (Dense)              (None, 512)               262656    \n",
      "_________________________________________________________________\n",
      "re_lu_5 (ReLU)               (None, 512)               0         \n",
      "_________________________________________________________________\n",
      "dropout_3 (Dropout)          (None, 512)               0         \n",
      "_________________________________________________________________\n",
      "dense_10 (Dense)             (None, 100)               51300     \n",
      "_________________________________________________________________\n",
      "re_lu_6 (ReLU)               (None, 100)               0         \n",
      "_________________________________________________________________\n",
      "dropout_4 (Dropout)          (None, 100)               0         \n",
      "_________________________________________________________________\n",
      "dense_11 (Dense)             (None, 5)                 505       \n",
      "=================================================================\n",
      "Total params: 17,388,957\n",
      "Trainable params: 2,674,269\n",
      "Non-trainable params: 14,714,688\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model = Sequential()\n",
    "\n",
    "model.add(vgg_model)\n",
    "model.add(Dense(512, input_dim=input_shape))\n",
    "model.add(ReLU())\n",
    "model.add(Dropout(0.2))\n",
    "model.add(Dense(512))\n",
    "model.add(ReLU())\n",
    "model.add(Dropout(0.2))\n",
    "model.add(Dense(100))  #Adicionar mais uma densa com 100 neurônios\n",
    "model.add(ReLU())\n",
    "model.add(Dropout(0.2))\n",
    "model.add(Dense(5, activation='sigmoid'))\n",
    "\n",
    "model.compile(loss='sparse_categorical_crossentropy',\n",
    "              optimizer=optimizers.RMSprop(lr=5e-5),\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Novamente como a intenção é reduzir o overfitting vou utilizar mais epochs, para conseguir analisar melhor o comportamento da rede:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:tensorflow:sample_weight modes were coerced from\n",
      "  ...\n",
      "    to  \n",
      "  ['...']\n",
      "WARNING:tensorflow:sample_weight modes were coerced from\n",
      "  ...\n",
      "    to  \n",
      "  ['...']\n",
      "Train for 100 steps, validate for 33 steps\n",
      "Epoch 1/40\n",
      "100/100 [==============================] - 122s 1s/step - loss: 1.4910 - accuracy: 0.3596 - val_loss: 1.2631 - val_accuracy: 0.5765\n",
      "Epoch 2/40\n",
      "100/100 [==============================] - 126s 1s/step - loss: 1.2588 - accuracy: 0.5223 - val_loss: 1.0445 - val_accuracy: 0.6198\n",
      "Epoch 3/40\n",
      "100/100 [==============================] - 126s 1s/step - loss: 1.1385 - accuracy: 0.5664 - val_loss: 0.9621 - val_accuracy: 0.6615\n",
      "Epoch 4/40\n",
      "100/100 [==============================] - 126s 1s/step - loss: 1.0619 - accuracy: 0.6001 - val_loss: 0.8601 - val_accuracy: 0.6909\n",
      "Epoch 5/40\n",
      "100/100 [==============================] - 126s 1s/step - loss: 0.9806 - accuracy: 0.6394 - val_loss: 0.8340 - val_accuracy: 0.6986\n",
      "Epoch 6/40\n",
      "100/100 [==============================] - 130s 1s/step - loss: 0.9398 - accuracy: 0.6544 - val_loss: 0.7785 - val_accuracy: 0.7187\n",
      "Epoch 7/40\n",
      "100/100 [==============================] - 122s 1s/step - loss: 0.8969 - accuracy: 0.6731 - val_loss: 0.7628 - val_accuracy: 0.7172\n",
      "Epoch 8/40\n",
      "100/100 [==============================] - 120s 1s/step - loss: 0.8676 - accuracy: 0.6815 - val_loss: 0.7619 - val_accuracy: 0.7357\n",
      "Epoch 9/40\n",
      "100/100 [==============================] - 119s 1s/step - loss: 0.8309 - accuracy: 0.6905 - val_loss: 0.7104 - val_accuracy: 0.7450\n",
      "Epoch 10/40\n",
      "100/100 [==============================] - 120s 1s/step - loss: 0.8041 - accuracy: 0.7105 - val_loss: 0.7036 - val_accuracy: 0.7342\n",
      "Epoch 11/40\n",
      "100/100 [==============================] - 122s 1s/step - loss: 0.8010 - accuracy: 0.7031 - val_loss: 0.6861 - val_accuracy: 0.7496\n",
      "Epoch 12/40\n",
      "100/100 [==============================] - 122s 1s/step - loss: 0.7721 - accuracy: 0.7098 - val_loss: 0.7138 - val_accuracy: 0.7388\n",
      "Epoch 13/40\n",
      "100/100 [==============================] - 123s 1s/step - loss: 0.7467 - accuracy: 0.7335 - val_loss: 0.6679 - val_accuracy: 0.7604\n",
      "Epoch 14/40\n",
      "100/100 [==============================] - 122s 1s/step - loss: 0.7350 - accuracy: 0.7295 - val_loss: 0.6989 - val_accuracy: 0.7589\n",
      "Epoch 15/40\n",
      "100/100 [==============================] - 126s 1s/step - loss: 0.7302 - accuracy: 0.7285 - val_loss: 0.6737 - val_accuracy: 0.7635\n",
      "Epoch 16/40\n",
      "100/100 [==============================] - 130s 1s/step - loss: 0.7042 - accuracy: 0.7365 - val_loss: 0.6533 - val_accuracy: 0.7666\n",
      "Epoch 17/40\n",
      "100/100 [==============================] - 134s 1s/step - loss: 0.6931 - accuracy: 0.7458 - val_loss: 0.6863 - val_accuracy: 0.7666\n",
      "Epoch 18/40\n",
      "100/100 [==============================] - 139s 1s/step - loss: 0.6844 - accuracy: 0.7442 - val_loss: 0.6774 - val_accuracy: 0.7573\n",
      "Epoch 19/40\n",
      "100/100 [==============================] - 128s 1s/step - loss: 0.6732 - accuracy: 0.7498 - val_loss: 0.6531 - val_accuracy: 0.7713\n",
      "Epoch 20/40\n",
      "100/100 [==============================] - 123s 1s/step - loss: 0.6611 - accuracy: 0.7538 - val_loss: 0.6629 - val_accuracy: 0.7666\n",
      "Epoch 21/40\n",
      "100/100 [==============================] - 122s 1s/step - loss: 0.6689 - accuracy: 0.7538 - val_loss: 0.6627 - val_accuracy: 0.7651\n",
      "Epoch 22/40\n",
      "100/100 [==============================] - 118s 1s/step - loss: 0.6471 - accuracy: 0.7675 - val_loss: 0.6713 - val_accuracy: 0.7651\n",
      "Epoch 23/40\n",
      "100/100 [==============================] - 118s 1s/step - loss: 0.6399 - accuracy: 0.7622 - val_loss: 0.6331 - val_accuracy: 0.7790\n",
      "Epoch 24/40\n",
      "100/100 [==============================] - 118s 1s/step - loss: 0.6160 - accuracy: 0.7775 - val_loss: 0.6155 - val_accuracy: 0.7713\n",
      "Epoch 25/40\n",
      "100/100 [==============================] - 118s 1s/step - loss: 0.6303 - accuracy: 0.7645 - val_loss: 0.6167 - val_accuracy: 0.7821\n",
      "Epoch 26/40\n",
      "100/100 [==============================] - 118s 1s/step - loss: 0.6084 - accuracy: 0.7692 - val_loss: 0.6223 - val_accuracy: 0.7805\n",
      "Epoch 27/40\n",
      "100/100 [==============================] - 117s 1s/step - loss: 0.5965 - accuracy: 0.7795 - val_loss: 0.6421 - val_accuracy: 0.7836\n",
      "Epoch 28/40\n",
      "100/100 [==============================] - 118s 1s/step - loss: 0.6034 - accuracy: 0.7718 - val_loss: 0.6393 - val_accuracy: 0.7790\n",
      "Epoch 29/40\n",
      "100/100 [==============================] - 118s 1s/step - loss: 0.6090 - accuracy: 0.7759 - val_loss: 0.6086 - val_accuracy: 0.7883\n",
      "Epoch 30/40\n",
      "100/100 [==============================] - 117s 1s/step - loss: 0.5810 - accuracy: 0.7905 - val_loss: 0.6137 - val_accuracy: 0.7929\n",
      "Epoch 31/40\n",
      "100/100 [==============================] - 116s 1s/step - loss: 0.5930 - accuracy: 0.7839 - val_loss: 0.6365 - val_accuracy: 0.7805\n",
      "Epoch 32/40\n",
      "100/100 [==============================] - 229s 2s/step - loss: 0.5612 - accuracy: 0.7949 - val_loss: 0.6641 - val_accuracy: 0.7713\n",
      "Epoch 33/40\n",
      "100/100 [==============================] - 166s 2s/step - loss: 0.5653 - accuracy: 0.7972 - val_loss: 0.6053 - val_accuracy: 0.7836\n",
      "Epoch 34/40\n",
      "100/100 [==============================] - 178s 2s/step - loss: 0.5573 - accuracy: 0.7975 - val_loss: 0.6323 - val_accuracy: 0.7805\n",
      "Epoch 35/40\n",
      "100/100 [==============================] - 180s 2s/step - loss: 0.5604 - accuracy: 0.7952 - val_loss: 0.6304 - val_accuracy: 0.7743\n",
      "Epoch 36/40\n",
      "100/100 [==============================] - 181s 2s/step - loss: 0.5525 - accuracy: 0.8005 - val_loss: 0.6191 - val_accuracy: 0.7883\n",
      "Epoch 37/40\n",
      "100/100 [==============================] - 174s 2s/step - loss: 0.5388 - accuracy: 0.8092 - val_loss: 0.5984 - val_accuracy: 0.8099\n",
      "Epoch 38/40\n",
      "100/100 [==============================] - 328s 3s/step - loss: 0.5391 - accuracy: 0.8035 - val_loss: 0.5953 - val_accuracy: 0.7898\n",
      "Epoch 39/40\n",
      "100/100 [==============================] - 633s 6s/step - loss: 0.5323 - accuracy: 0.8075 - val_loss: 0.6091 - val_accuracy: 0.7913\n",
      "Epoch 40/40\n",
      "100/100 [==============================] - 591s 6s/step - loss: 0.5094 - accuracy: 0.8179 - val_loss: 0.6157 - val_accuracy: 0.8022\n"
     ]
    }
   ],
   "source": [
    "history = model.fit(datagen_tr.flow(train_imgs_scaled, train_labels_enc, batch_size=30),\n",
    "                    validation_data=datagen_val.flow(val_imgs_scaled, val_labels_enc, batch_size=20),\n",
    "                    epochs=40,\n",
    "                    steps_per_epoch=100)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.save('Modelo_Transfer_Reg_Aug.h5') #Salvando o modelo para posterior anáise"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Transfer Learning com ajuste fino de parâmetros"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vemos acima que o modelo alcançou melhores resultados na validação que o transfer puro, além de inibir o overfitting. O ganho de desempenho foi relativamente pequeno visto que não deixei o modelo treinar por tempo suficiente, devido às limitações importas pelo meu hardware, como da pra ver pelo tempo em cada epoch, ainda assim é visível que esse modelo é melhor que o anterior.\n",
    "\n",
    "&emsp;Mas será que não conseguimos melhorá-lo ainda mais? Até aqui estamos utilizando apenas a estratégia *off the shelf*, vamos tentar utilizar o *fine-tuning*, lembrando que está é mais custosa pois temos mais parâmetros treináveis. Vamos habilitar o treino apenas em uma camada de convolução do nosso modelo, a última, mas poderíamos fazer isso com mais camadas, a tendência é qua as últimas camadas de convolução busquem features de mais alto nível, de forma que o ideal é treinar estas para as particularidades do nosso dataset. A imagem a seguir ilustra as *features* que cada camada tende a procurar:\n",
    "\n",
    "<img src = https://dpzbhybb2pdcj.cloudfront.net/elgendy/v-8/Figures/01-04_img_0184.png>\n",
    "\n",
    "Fonte: [3 Convolutional Neural Networks (CNNs)](https://livebook.manning.com/book/grokking-deep-learning-for-computer-vision/chapter-3/v-8/118)\n",
    "\n",
    "&emsp;Enquanto as camadas iniciais buscam por *features* mais simples, e portanto mais genéricas, as mais profundas buscam *featuers* mais complexas e específicas. Por isso esses modelos estado da arte funcionam tão bem independente do caso de transfer que você executa, pois a capacidade delas em identificar característica de baixo nível é muito avançada, logo se treinar-mos as camadas finais a tendência é que o modelo fique ainda melhor, vamos testar:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [],
   "source": [
    "vgg_model.trainable = True  #Permitindo o treino\n",
    "\n",
    "set_trainable = False\n",
    "for layer in vgg_model.layers:\n",
    "    if layer.name in ['block5_conv1']: #Habilitamos o treino apenas para a última camada\n",
    "        set_trainable = True\n",
    "    if set_trainable:\n",
    "        layer.trainable = True\n",
    "    else:\n",
    "        layer.trainable = False"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Como o treino novamente é demorado vou voltar a utilizar apenas 10 epochs, mas será o suficiente para ver o poder do modelo, vou utilizar ainda o augumentation e o dropout:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:tensorflow:sample_weight modes were coerced from\n",
      "  ...\n",
      "    to  \n",
      "  ['...']\n",
      "WARNING:tensorflow:sample_weight modes were coerced from\n",
      "  ...\n",
      "    to  \n",
      "  ['...']\n",
      "Train for 100 steps, validate for 33 steps\n",
      "Epoch 1/10\n",
      "100/100 [==============================] - 802s 8s/step - loss: 1.0656 - accuracy: 0.5797 - val_loss: 0.6595 - val_accuracy: 0.7682\n",
      "Epoch 2/10\n",
      "100/100 [==============================] - 818s 8s/step - loss: 0.7329 - accuracy: 0.7308 - val_loss: 0.5678 - val_accuracy: 0.8006\n",
      "Epoch 3/10\n",
      "100/100 [==============================] - 801s 8s/step - loss: 0.6251 - accuracy: 0.7705 - val_loss: 0.5898 - val_accuracy: 0.7805\n",
      "Epoch 4/10\n",
      "100/100 [==============================] - 810s 8s/step - loss: 0.5396 - accuracy: 0.8065 - val_loss: 0.4754 - val_accuracy: 0.8362\n",
      "Epoch 5/10\n",
      "100/100 [==============================] - 797s 8s/step - loss: 0.4878 - accuracy: 0.8329 - val_loss: 0.4610 - val_accuracy: 0.8393\n",
      "Epoch 6/10\n",
      "100/100 [==============================] - 1964s 20s/step - loss: 0.4431 - accuracy: 0.8396 - val_loss: 0.6211 - val_accuracy: 0.8238\n",
      "Epoch 7/10\n",
      "100/100 [==============================] - 211s 2s/step - loss: 0.4181 - accuracy: 0.8466 - val_loss: 0.5968 - val_accuracy: 0.8393\n",
      "Epoch 8/10\n",
      "100/100 [==============================] - 227s 2s/step - loss: 0.4089 - accuracy: 0.8606 - val_loss: 0.4672 - val_accuracy: 0.8578\n",
      "Epoch 9/10\n",
      "100/100 [==============================] - 217s 2s/step - loss: 0.3662 - accuracy: 0.8722 - val_loss: 0.6309 - val_accuracy: 0.8408\n",
      "Epoch 10/10\n",
      "100/100 [==============================] - 231s 2s/step - loss: 0.3578 - accuracy: 0.8756 - val_loss: 0.7441 - val_accuracy: 0.8253\n"
     ]
    }
   ],
   "source": [
    "model = Sequential()\n",
    "\n",
    "model.add(vgg_model)\n",
    "model.add(Dense(512, input_dim=input_shape))\n",
    "model.add(ReLU())\n",
    "model.add(Dropout(0.2))\n",
    "model.add(Dense(512))\n",
    "model.add(ReLU())\n",
    "model.add(Dropout(0.2))\n",
    "model.add(ReLU())\n",
    "model.add(Dropout(0.2))\n",
    "model.add(Dense(5, activation='sigmoid'))\n",
    "\n",
    "model.compile(loss='sparse_categorical_crossentropy',\n",
    "              optimizer=optimizers.RMSprop(lr=1e-4),\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "history = model.fit(datagen_tr.flow(train_imgs_scaled, train_labels_enc, batch_size=30),\n",
    "                    validation_data=datagen_val.flow(val_imgs_scaled, val_labels_enc, batch_size=20),\n",
    "                    epochs=10,\n",
    "                    steps_per_epoch=100)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.save('Modelo_Transfer_Fine.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "f, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 4))\n",
    "t = f.suptitle('Performance com uma VGG-16 padrão', fontsize=12)\n",
    "f.subplots_adjust(top=0.85, wspace=0.3)\n",
    "\n",
    "epoch_list = list(range(1,11))\n",
    "ax1.plot(epoch_list, history.history['accuracy'], label='Acurácia do treino')\n",
    "ax1.plot(epoch_list, history.history['val_accuracy'], label='Acurácia na validação')\n",
    "ax1.set_xticks(np.arange(0, 11, 2))\n",
    "ax1.set_ylabel('Acurácia')\n",
    "ax1.set_xlabel('Epoch')\n",
    "ax1.set_title('Acurácia')\n",
    "l1 = ax1.legend(loc=\"best\")\n",
    "\n",
    "ax2.plot(epoch_list, history.history['loss'], label='Loss Treino')\n",
    "ax2.plot(epoch_list, history.history['val_loss'], label='Loss Validação')\n",
    "ax2.set_xticks(np.arange(0, 11, 2))\n",
    "ax2.set_ylabel('Valor do Loss')\n",
    "ax2.set_xlabel('Epoch')\n",
    "ax2.set_title('Loss')\n",
    "l2 = ax2.legend(loc=\"best\")\n",
    "\n",
    "ax1.grid()\n",
    "ax2.grid()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Incrível a evolução do modelo né? Esse é o poder de um *transfer learning* com arquiteturas estado-da-arte. Vale ver que o modelo não vai ao overfitting e que ainda poderíamos tentar obter maior acurácia, mas vamos ficar com este por enquanto, já é visível o ganho que o modelo proporciona."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Avaliando os resultados no treino"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Vamos agora analisar a eficiência dos modelos no domínio de teste, e para tal precisamos primeiro fazer o load de cada um:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [],
   "source": [
    "from tensorflow.keras.models import load_model\n",
    "\n",
    "basica = load_model('Modelo_Básico_CNN.h5')\n",
    "basica_aug_reg = load_model('Modelo_Reg_Aug_CNN.h5')\n",
    "transfer = load_model('Modelo_Transfer_Básico.h5')\n",
    "transfer_aug_red = load_model('Modelo_Transfer_Reg_Aug.h5')\n",
    "transfer_fine = load_model('Modelo_Transfer_Fine.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "648/648 [==============================] - 1s 1ms/sample - loss: 1.3061 - accuracy: 0.6775\n",
      "Acurácia no Teste: 0.6774691343307495\n",
      "Loss no Teste: 1.306059303475015\n"
     ]
    }
   ],
   "source": [
    "pred = basica.evaluate(test_imgs_scaled, test_labels_enc)\n",
    "print('Acurácia no Teste: {}'.format(pred[1]))\n",
    "print('Loss no Teste: {}'.format(pred[0]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;A rede básica de CNN nos proporcionou uma acurácia de 67.7%, bastante razoável até para um dataset complicado como este, mas abaixo do esperado, vamos ver o próximo."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "648/648 [==============================] - 1s 2ms/sample - loss: 0.8273 - accuracy: 0.7500\n",
      "Acurácia no Teste: 0.75\n",
      "Loss no Teste: 0.8272964586446314\n"
     ]
    }
   ],
   "source": [
    "pred = basica_aug_reg.evaluate(test_imgs_scaled, test_labels_enc)\n",
    "print('Acurácia no Teste: {}'.format(pred[1]))\n",
    "print('Loss no Teste: {}'.format(pred[0]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Nossa tentativa de aumentar o desempenho evitando o overfit da rede funcionou bem e nos deu um ganho de 8% no teste, que é muito bom, lembrando que quando mais alto o desempenho, mais vale 1% de aumento. Ainda assim ainda não fizemos o transfer, vamos ver como ele foi:\n",
    "Ps: Eu re-executei a vgg pois alteramos o objeto ao longo do código então para avaliar corretamente só copiei as células anteriores, nada de novo."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "648/648 [==============================] - 0s 289us/sample - loss: 1.5908 - accuracy: 0.7253\n",
      "Acurácia no Teste: 0.7253086566925049\n",
      "Loss no Teste: 1.5908256401250391\n"
     ]
    }
   ],
   "source": [
    "input_shape = (100, 100, 3)\n",
    "vgg = vgg16.VGG16(include_top=False,  #Parâmetro responsável por incluir ou descartar as camadas Fully-connected\n",
    "                  weights='imagenet', \n",
    "                  input_shape=input_shape)\n",
    "\n",
    "saida = vgg.layers[-1].output           #Última camada presente no nosso modelo\n",
    "saida = tensorflow.keras.layers.Flatten()(saida)  \n",
    "vgg_model = Model(vgg.input, saida)    #Concatenando o modelo que temos com o *flatten*\n",
    "\n",
    "vgg_model.trainable = False\n",
    "for layer in vgg_model.layers:\n",
    "    layer.trainable = False\n",
    "\n",
    "test_vgg = passar_pela_vgg(vgg_model, test_imgs_scaled)\n",
    "\n",
    "pred = transfer.evaluate(test_vgg, test_labels_enc)\n",
    "print('Acurácia no Teste: {}'.format(pred[1]))\n",
    "print('Loss no Teste: {}'.format(pred[0]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Um transfer puro, sem augumentation e *Off the shelf features* já nos deu um resultado robusto no teste de 72,5% de acurácia, lembrando que aqui também houve overfitting, então tentamos melhorar o desempenho com augumentation, nesse caso realmente não consegui treinar a rede muito por conta do meu computador, mas vamos lá:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "648/648 [==============================] - 28s 44ms/sample - loss: 0.6420 - accuracy: 0.7623\n",
      "Acurácia no Teste: 0.7623456716537476\n",
      "Loss no Teste: 0.6420142312108734\n"
     ]
    }
   ],
   "source": [
    "pred = transfer_aug_red.evaluate(test_imgs_scaled, test_labels_enc)\n",
    "print('Acurácia no Teste: {}'.format(pred[1]))\n",
    "print('Loss no Teste: {}'.format(pred[0]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Este transfer nos proporcionou um aumento até mesmo em relação ao modelo de CNN com augumentation e regularização, chegando à 76%. Mas ainda não conseguimos o melhor resultado possível, vamos ver como o *fine-tuning*, aliado às outras técnicas conseguiu performar:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "648/648 [==============================] - 33s 50ms/sample - loss: 0.7657 - accuracy: 0.8318\n",
      "Acurácia no Teste: 0.8317901492118835\n",
      "Loss no Teste: 0.7657408302212939\n"
     ]
    }
   ],
   "source": [
    "pred = transfer_fine.evaluate(test_imgs_scaled, test_labels_enc)\n",
    "print('Acurácia no Teste: {}'.format(pred[1]))\n",
    "print('Loss no Teste: {}'.format(pred[0]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Simplesmente chegamos à 83% de acurácia em apenas 10 epochs, com muito espaço pra aumentar esse resultado, tanto no tempo de treino, quanto no *learning rate* quanto treinando mais camadas de convolução. Realmente se trata de um modelo extremamente robusto, um ganho total de 26% em relação ao resultado original é muito significativo. Entendem agora o porquê o *transfer learning* é um tópico tão relevante?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "&emsp;Gostaria de agradecer Dipanjan Sakar por seu post incrivelmente completo no Medium: [A Comprehensive Hands-on Guide to Transfer Learning with Real-World Applications in Deep Learning](https://towardsdatascience.com/a-comprehensive-hands-on-guide-to-transfer-learning-with-real-world-applications-in-deep-learning-212bf3b2f27a), uma vez que todo o texto aqui exposto foi baseado no que ele escreveu e no caminho lógico que desenvolveu, bem como seu exemplo prático. Se você chegou até aqui e ainda está interessado no tema, recomendo a leitura, muitos assuntos mais secundários eu não adicionei mas ele desenvolve."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}