{"cells":[{"metadata":{},"cell_type":"markdown","source":"# Introdução","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"**Resumo:**\n\nO exemplo aqui desenvolvido tem como objetivo apresentar conceitos iniciais de implementação de redes neurais com python e tensorflow/keras. Esse modelo apresenta um modelo de Rede Neural Convolucional (CNN) básico que pode ser expandido mudando o número de neurônios e camadas. Em adaptações mais avançadas, pode-se estudar possibilidade de otimização de hyperparâmetros e outras técnincas como aumento de dados.\n\n**Não é objetivo nosso desenvolver e otimizar o modelo de classificação**. O exemplo tem objetivo meramente didático.\n\n---\n\n**Para saber mais:**\n* [Palestras e cursos do Ocean](http://www.oceanbrasil.com/)\n* François Chollet. Deep Learning with Python. Manning Publications, 2017.\n* Ian Goodfellow and Yoshua Bengio and Aaron Courville. [Deep Learning](https://www.deeplearningbook.org/). MIT Press, 2016.","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"# Bibliotecas e Dados","execution_count":null},{"metadata":{"ExecuteTime":{"end_time":"2019-01-16T20:39:41.500258Z","start_time":"2019-01-16T20:39:39.261856Z"},"trusted":true},"cell_type":"code","source":"# Bibliotecas necessárias\n# Manipulação de dados\nimport pandas as pd\n# Redes Neurais\nfrom tensorflow import keras\nfrom tensorflow.keras.datasets import mnist\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense, Dropout, Conv2D, MaxPooling2D, Flatten\nfrom tensorflow.keras.optimizers import RMSprop, Adam\n# Plot\nimport matplotlib.pyplot as plt\n%matplotlib inline\n# Avaliação\nfrom sklearn.metrics import classification_report, confusion_matrix\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Lendo o dataset Kaggle\ntrain = pd.read_csv(\"../input/digit-recognizer/train.csv\")\n\n#Alternativa ler do próprio keras\n#(x_train, y_train), (x_test, y_test) = mnist.load_data()\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Analisando o dataset\nprint(\"Quantidade de elementos de treino: {}\". format(len(train)))\nprint(train.head())","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Separando x_train e y_train\nY = train[\"label\"]\nX = train.drop(labels = [\"label\"],axis = 1)\nprint(X.head())\n# Em formato numpy array de imagens 28 x 28\n#x = X.values.reshape(-1,28,28,1)\n#print(x[0])","execution_count":null,"outputs":[]},{"metadata":{"ExecuteTime":{"end_time":"2019-01-08T20:49:47.144492Z","start_time":"2019-01-08T20:49:47.016025Z"},"trusted":true},"cell_type":"code","source":"# Numtendi nada!\n# Bora ver com matplotlib\nplt.imshow(X.values[0].reshape(28,28), cmap=plt.cm.binary)\nplt.show()\nprint('Label: {}'.format(Y[0]))","execution_count":null,"outputs":[]},{"metadata":{"ExecuteTime":{"end_time":"2019-01-08T20:49:57.425537Z","start_time":"2019-01-08T20:49:57.272377Z"},"trusted":true},"cell_type":"code","source":"# Transformando a imagem 2d em um numpy array (imagem 28*28)\nx = X.values.reshape(42000, 28, 28, 1)\n\n#Normalizando para valores entre 0 e 1\nx = x.astype('float32')\nx /= 255\n\n#print(x[0])","execution_count":null,"outputs":[]},{"metadata":{"ExecuteTime":{"end_time":"2019-01-08T20:50:48.007428Z","start_time":"2019-01-08T20:50:47.986527Z"},"trusted":true},"cell_type":"code","source":"# Vamos ajustar o formato da saida\nnum_classes = 10\n\n# Convertendo para um vetor de saida com 10 dimensoes\n# ex. 8 => [0,0,0,0,0,0,0,0,1,0]\ny = keras.utils.to_categorical(Y, num_classes)\nprint(y[0])","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Separando uma parte para treino (90%) e outra para validação (10%)\nfrom sklearn.model_selection import train_test_split\nx_train, x_val, y_train, y_val = train_test_split(x, y, test_size = 0.1, random_state=9)\nprint('Qtde de treino: {}'.format(len(x_train)))\nprint('Qtde de validação: {}'.format(len(x_val)))","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Criando e treinando o Modelo","execution_count":null},{"metadata":{"ExecuteTime":{"end_time":"2019-01-08T21:38:00.233865Z","start_time":"2019-01-08T21:38:00.15079Z"},"scrolled":true,"trusted":true},"cell_type":"code","source":"# Criando o modelo Sequential\n# Sequential: Modelo Keras de ir adicionando camadas (como um lego)\n# Conv2D: Camada com kernels (filtros) que percorrem a imagem extraindo caracterísitcas (mapas de caracteísticas)\n# MaxPooling2D: Camada que reduz a dimensionalidade dos mapas de características 2D\n# Flatten: Camada que transforma um mapa de características 2D num vetor para classficador final\n# Dense: Camada onde todas as entradas estão conectadas em cada neurônio (totalmente conectada)\n# Dropout: Camada usa durante treino que descarta aleatoriamente um percentual de conexões (reduz overfitting)\n\nmodel = Sequential()\nmodel.add(Conv2D(20, kernel_size=(3, 3),\n                 activation='relu',\n                 input_shape=(28,28,1)))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Conv2D(40, kernel_size=(3,3), activation='relu'))\nmodel.add(Conv2D(40, kernel_size=(3,3), activation='relu'))\nmodel.add(Conv2D(40, kernel_size=(3,3), activation='relu'))\nmodel.add(Conv2D(80, kernel_size=(3,3), activation='relu'))\nmodel.add(Conv2D(80, kernel_size=(3,3), activation='relu'))\nmodel.add(Flatten())\nmodel.add(Dense(64, activation='relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(64, activation='relu'))\nmodel.add(Dropout(0.2))\nmodel.add(Dense(num_classes, activation='softmax'))\nmodel.summary()","execution_count":null,"outputs":[]},{"metadata":{"ExecuteTime":{"end_time":"2019-01-08T21:33:17.184644Z","start_time":"2019-01-08T21:33:17.129141Z"},"trusted":true},"cell_type":"code","source":"# Compila o modelo\nmodel.compile(loss='categorical_crossentropy',\n              optimizer='adam', #optimizer='adam',\n              metrics=['accuracy'])","execution_count":null,"outputs":[]},{"metadata":{"ExecuteTime":{"end_time":"2019-01-08T21:36:15.501168Z","start_time":"2019-01-08T21:33:19.322895Z"},"trusted":true},"cell_type":"code","source":"# Treina com os parte dos dados\nbatch_size = 64\nepochs = 50\n\n#Salvar o melhor modelo\ncallbacks_list = [\n    keras.callbacks.ModelCheckpoint(\n        filepath='model.h5',\n        monitor='val_loss', save_best_only=True, verbose=1),\n    keras.callbacks.EarlyStopping(monitor='val_loss', patience=12,verbose=1)\n]\n\nhistory = model.fit(x_train, y_train,\n                    batch_size=batch_size,\n                    epochs=epochs,\n                    callbacks = callbacks_list,\n                    verbose=1,\n                    validation_data=(x_val, y_val))","execution_count":null,"outputs":[]},{"metadata":{"ExecuteTime":{"end_time":"2019-01-08T21:36:20.243921Z","start_time":"2019-01-08T21:36:20.064201Z"},"trusted":true},"cell_type":"code","source":"#Vamos ver como foi o treino?\n\nfig, ax = plt.subplots(1,2, figsize=(20,10))\nax[0].plot(history.history['loss'], color='b', label=\"Training loss\")\nax[0].plot(history.history['val_loss'], color='r', label=\"validation loss\",axes =ax[0])\nlegend = ax[0].legend(loc='best', shadow=True)\n\nax[1].plot(history.history['accuracy'], color='b', label=\"Training accuracy\")\nax[1].plot(history.history['val_accuracy'], color='r',label=\"Validation accuracy\")\nlegend = ax[1].legend(loc='best', shadow=True)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"from tensorflow.keras.models import load_model\n# Load the best saved model\nmodel = load_model('model.h5')","execution_count":null,"outputs":[]},{"metadata":{"ExecuteTime":{"end_time":"2019-01-08T21:36:39.779842Z","start_time":"2019-01-08T21:36:38.946173Z"},"trusted":true},"cell_type":"code","source":"# Testa\nscore = model.evaluate(x_val, y_val, verbose=0)\nprint('Test loss:', score[0])\nprint('Test accuracy:', score[1])","execution_count":null,"outputs":[]},{"metadata":{"ExecuteTime":{"end_time":"2019-01-08T21:06:04.363715Z","start_time":"2019-01-08T21:06:04.27849Z"},"trusted":true},"cell_type":"code","source":"# Testando uma entrada qualquer\nprint(y_train[10])\nprint(model.predict(x_train[10].reshape((1,28,28,1))))\nprint(model.predict_classes(x_train[10].reshape((1,28,28,1))))","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Avaliando o Modelo","execution_count":null},{"metadata":{"ExecuteTime":{"end_time":"2019-01-08T21:24:57.7654Z","start_time":"2019-01-08T21:24:57.220867Z"},"trusted":true},"cell_type":"code","source":"# Vendo alguns reports\n# Usando sklearn\n\n# Classificando toda base de teste\ny_pred = model.predict_classes(x_val)\n# voltando pro formato de classes\nimport numpy as np\ny_test_c = np.argmax(y_val, axis=1)\n\nprint('Confusion Matrix')\nprint(confusion_matrix(y_test_c, y_pred))\nprint('Classification Report')\ntarget_names = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']\nprint(classification_report(y_test_c, y_pred, target_names=target_names))","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Gerando Saída","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"# Gerando saída para dataset de teste\n\n#Carrega dataset de teste\ntest = pd.read_csv(\"../input/digit-recognizer/test.csv\")\nprint(\"Qtde de testes: {}\".format(len(test)))\n# Bota no formato numpy e normaliza\nx_test = test.values.reshape(len(test),28,28,1)\nx_test = x_test.astype('float32')\nx_test /= 255\n\n# Faz classificação para dataset de teste\ny_pred = model.predict_classes(x_test)\n\n# Verficando algum exemplo\ni = 0\nplt.imshow(test.values[i].reshape(28,28), cmap=plt.cm.binary)\nplt.show()\nprint('Previsto: {}'.format(y_pred[i]))\n\n# Botando no formato de saída (competição Kaggle)\nresults = pd.Series(y_pred,name=\"Label\")\nsubmission = pd.concat([pd.Series(range(1,len(y_pred)+1),name = \"ImageId\"),results],axis = 1)\nprint(submission.head(10))\n#Salvando Arquivo\nsubmission.to_csv(\"mlp_mnist_v1.csv\",index=False)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Teste Adicional: Com ruído","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"#introduzindo ruido\nimport numpy as np\nmean = 0.\nstddev = 0.2\nnoise = np.random.normal(mean, stddev, (4200, 28, 28,1))\nx_te = x_val + noise\nx_te = np.clip(x_te, 0., 1.)\n\nplt.imshow(x_te.reshape(4200, 28,28)[0], cmap=plt.cm.binary)\nplt.show()\n\n# Testa\nscore = model.evaluate(x_te, y_val, verbose=0)\nprint('Test loss:', score[0])\nprint('Test accuracy:', score[1])","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"Com ruído, percebemos que a acurácia dos exemplos de validação caiu, mas muito menos que o modelo MLP clássico (ver exemplo de MLP).\nUm modelo Convolucional (CNN) captura melhor regiões, ou padrões espaciais de pixels, tornando o modelo mais robusto.","execution_count":null},{"metadata":{"trusted":true},"cell_type":"code","source":"","execution_count":null,"outputs":[]}],"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat":4,"nbformat_minor":4}