Cnn completo

Exemplo completo juntando tudo que já vimos e mais um pouco...¶

Vamos usar o conjunto de dados CIFAR-10, que é um conjunto de dados de 60.000 imagens de 10 classes, com 6.000 imagens por classe.

As imagens são de tamanho 32x32 pixels com três canais de cores (RGB).

Ao longo do notebook vamos relembrar alguns conceitos e conhecer novos.

LEMBRETE IMPORTANTE: Lembre-se de setar o colab para usar a GPU.

Configurando o google drive¶

Vamos setar o google drive para salvar o modelo durante o treinamento

In [5]:

from google.colab import drive
drive.mount('/content/drive')

# Vamos definir o caminho onde o modelo será salvo no Google Drive
model_save_path = '/content/drive/MyDrive/checkpoints/cifar10_best_model.h5'

from google.colab import drive drive.mount('/content/drive') # Vamos definir o caminho onde o modelo será salvo no Google Drive model_save_path = '/content/drive/MyDrive/checkpoints/cifar10_best_model.h5'

Drive already mounted at /content/drive; to attempt to forcibly remount, call drive.mount("/content/drive", force_remount=True).

Importando e Preparando os Dados CIFAR-10¶

Primeiro, vamos importar e preparar os dados CIFAR-10:

In [6]:

import tensorflow as tf
from tensorflow.keras.datasets import cifar10
from tensorflow.keras.utils import to_categorical

# Carregando o conjunto de dados CIFAR-10
(x_train, y_train), (x_val, y_val) = cifar10.load_data()

# Normalizando os valores dos pixels para o intervalo [0, 1]
x_train = x_train.astype('float32') / 255.0
x_val = x_val.astype('float32') / 255.0

# Convertendo os rótulos para vetores one-hot
y_train = to_categorical(y_train, 10)
y_val = to_categorical(y_val, 10)

import tensorflow as tf from tensorflow.keras.datasets import cifar10 from tensorflow.keras.utils import to_categorical # Carregando o conjunto de dados CIFAR-10 (x_train, y_train), (x_val, y_val) = cifar10.load_data() # Normalizando os valores dos pixels para o intervalo [0, 1] x_train = x_train.astype('float32') / 255.0 x_val = x_val.astype('float32') / 255.0 # Convertendo os rótulos para vetores one-hot y_train = to_categorical(y_train, 10) y_val = to_categorical(y_val, 10)

Definindo o Modelo¶

Vamos definir o modelo para o dataset:

In [7]:

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout, BatchNormalization
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping, ReduceLROnPlateau
from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Definindo o modelo
model = Sequential()

# Camadas convolucionais com número crescente de filtros
model.add(Conv2D(32, (3, 3), activation='relu', input_shape=(32, 32, 3)))
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(128, (3, 3), activation='relu'))
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(2, 2)))

# Camada Flatten para converter a saída das camadas convolucionais em um vetor 1D
model.add(Flatten())

# Camadas totalmente conectadas
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))

model.add(Dense(64, activation='relu'))
model.add(Dropout(0.5))

# Última camada totalmente conectada com 10 saídas (10 classes de categoria de imagem)
model.add(Dense(10, activation='softmax'))

# Compilando o modelo
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

# Resumo do modelo
model.summary()

from tensorflow.keras.models import Sequential from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout, BatchNormalization from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping, ReduceLROnPlateau from tensorflow.keras.preprocessing.image import ImageDataGenerator # Definindo o modelo model = Sequential() # Camadas convolucionais com número crescente de filtros model.add(Conv2D(32, (3, 3), activation='relu', input_shape=(32, 32, 3))) model.add(BatchNormalization()) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Conv2D(64, (3, 3), activation='relu')) model.add(BatchNormalization()) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Conv2D(128, (3, 3), activation='relu')) model.add(BatchNormalization()) model.add(MaxPooling2D(pool_size=(2, 2))) # Camada Flatten para converter a saída das camadas convolucionais em um vetor 1D model.add(Flatten()) # Camadas totalmente conectadas model.add(Dense(128, activation='relu')) model.add(Dropout(0.5)) model.add(Dense(64, activation='relu')) model.add(Dropout(0.5)) # Última camada totalmente conectada com 10 saídas (10 classes de categoria de imagem) model.add(Dense(10, activation='softmax')) # Compilando o modelo model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy']) # Resumo do modelo model.summary()

Model: "sequential_1"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d_3 (Conv2D)           (None, 30, 30, 32)        896       
                                                                 
 batch_normalization_3 (Bat  (None, 30, 30, 32)        128       
 chNormalization)                                                
                                                                 
 max_pooling2d_3 (MaxPoolin  (None, 15, 15, 32)        0         
 g2D)                                                            
                                                                 
 conv2d_4 (Conv2D)           (None, 13, 13, 64)        18496     
                                                                 
 batch_normalization_4 (Bat  (None, 13, 13, 64)        256       
 chNormalization)                                                
                                                                 
 max_pooling2d_4 (MaxPoolin  (None, 6, 6, 64)          0         
 g2D)                                                            
                                                                 
 conv2d_5 (Conv2D)           (None, 4, 4, 128)         73856     
                                                                 
 batch_normalization_5 (Bat  (None, 4, 4, 128)         512       
 chNormalization)                                                
                                                                 
 max_pooling2d_5 (MaxPoolin  (None, 2, 2, 128)         0         
 g2D)                                                            
                                                                 
 flatten_1 (Flatten)         (None, 512)               0         
                                                                 
 dense_3 (Dense)             (None, 128)               65664     
                                                                 
 dropout_2 (Dropout)         (None, 128)               0         
                                                                 
 dense_4 (Dense)             (None, 64)                8256      
                                                                 
 dropout_3 (Dropout)         (None, 64)                0         
                                                                 
 dense_5 (Dense)             (None, 10)                650       
                                                                 
=================================================================
Total params: 168714 (659.04 KB)
Trainable params: 168266 (657.29 KB)
Non-trainable params: 448 (1.75 KB)
_________________________________________________________________

Configurando Callbacks¶

Vamos configurar as callbacks: ModelCheckpoint, EarlyStopping e ReduceLROnPlateau:

In [8]:

# Configurando a callback ModelCheckpoint para salvar o modelo no google drive
checkpoint = ModelCheckpoint(model_save_path,
                             monitor='val_accuracy',
                             save_best_only=True,
                             mode='max',
                             verbose=1)

# Configurando a callback EarlyStopping
early_stopping = EarlyStopping(monitor='val_loss',
                               patience=10,
                               verbose=1,
                               restore_best_weights=True)

# Configurando a callback ReduceLROnPlateau
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
                              factor=0.2,
                              patience=5,
                              min_lr=0.001,
                              verbose=1)

# Configurando a callback ModelCheckpoint para salvar o modelo no google drive checkpoint = ModelCheckpoint(model_save_path, monitor='val_accuracy', save_best_only=True, mode='max', verbose=1) # Configurando a callback EarlyStopping early_stopping = EarlyStopping(monitor='val_loss', patience=10, verbose=1, restore_best_weights=True) # Configurando a callback ReduceLROnPlateau reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=5, min_lr=0.001, verbose=1)

Configurando Data Augmentation¶

Vamos configurar a data augmentation usando ImageDataGenerator:

In [9]:

# Configurando o data augmentation
datagen = ImageDataGenerator(
    rotation_range=20,
    width_shift_range=0.2,
    height_shift_range=0.2,
    horizontal_flip=True,
    zoom_range=0.2
)

# Gerando dados de treinamento augmentados
train_generator = datagen.flow(x_train, y_train, batch_size=32)

# Configurando o data augmentation datagen = ImageDataGenerator( rotation_range=20, width_shift_range=0.2, height_shift_range=0.2, horizontal_flip=True, zoom_range=0.2 ) # Gerando dados de treinamento augmentados train_generator = datagen.flow(x_train, y_train, batch_size=32)

Treinando o Modelo¶

Finalmente chegou o momento de treinar o modelo com as callbacks e data augmentation:

In [10]:

# Treinando o modelo com as callbacks e data augmentation
history = model.fit(train_generator,
                    epochs=100,
                    validation_data=(x_val, y_val),
                    callbacks=[checkpoint, early_stopping, reduce_lr])

# Treinando o modelo com as callbacks e data augmentation history = model.fit(train_generator, epochs=100, validation_data=(x_val, y_val), callbacks=[checkpoint, early_stopping, reduce_lr])

Epoch 1/100
1562/1563 [============================>.] - ETA: 0s - loss: 2.0423 - accuracy: 0.2451
Epoch 1: val_accuracy improved from -inf to 0.28590, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 43s 23ms/step - loss: 2.0421 - accuracy: 0.2451 - val_loss: 2.0514 - val_accuracy: 0.2859 - lr: 0.0010
Epoch 2/100
   1/1563 [..............................] - ETA: 52s - loss: 1.8313 - accuracy: 0.1562

/usr/local/lib/python3.10/dist-packages/keras/src/engine/training.py:3103: UserWarning: You are saving your model as an HDF5 file via `model.save()`. This file format is considered legacy. We recommend using instead the native Keras format, e.g. `model.save('my_model.keras')`.
  saving_api.save_model(

1561/1563 [============================>.] - ETA: 0s - loss: 1.7979 - accuracy: 0.3332
Epoch 2: val_accuracy improved from 0.28590 to 0.45100, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 36s 23ms/step - loss: 1.7977 - accuracy: 0.3333 - val_loss: 1.4733 - val_accuracy: 0.4510 - lr: 0.0010
Epoch 3/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.6728 - accuracy: 0.3916
Epoch 3: val_accuracy improved from 0.45100 to 0.48650, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 37s 23ms/step - loss: 1.6729 - accuracy: 0.3917 - val_loss: 1.3924 - val_accuracy: 0.4865 - lr: 0.0010
Epoch 4/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.6007 - accuracy: 0.4219
Epoch 4: val_accuracy did not improve from 0.48650
1563/1563 [==============================] - 37s 23ms/step - loss: 1.6007 - accuracy: 0.4219 - val_loss: 1.5411 - val_accuracy: 0.4524 - lr: 0.0010
Epoch 5/100
1563/1563 [==============================] - ETA: 0s - loss: 1.5361 - accuracy: 0.4527
Epoch 5: val_accuracy improved from 0.48650 to 0.49570, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 36s 23ms/step - loss: 1.5361 - accuracy: 0.4527 - val_loss: 1.4323 - val_accuracy: 0.4957 - lr: 0.0010
Epoch 6/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.4981 - accuracy: 0.4708
Epoch 6: val_accuracy did not improve from 0.49570
1563/1563 [==============================] - 35s 22ms/step - loss: 1.4981 - accuracy: 0.4708 - val_loss: 1.4484 - val_accuracy: 0.4790 - lr: 0.0010
Epoch 7/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.4468 - accuracy: 0.4942
Epoch 7: val_accuracy improved from 0.49570 to 0.56870, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 35s 23ms/step - loss: 1.4468 - accuracy: 0.4941 - val_loss: 1.1967 - val_accuracy: 0.5687 - lr: 0.0010
Epoch 8/100
1563/1563 [==============================] - ETA: 0s - loss: 1.4102 - accuracy: 0.5118
Epoch 8: val_accuracy improved from 0.56870 to 0.60030, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 36s 23ms/step - loss: 1.4102 - accuracy: 0.5118 - val_loss: 1.1149 - val_accuracy: 0.6003 - lr: 0.0010
Epoch 9/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.3760 - accuracy: 0.5249
Epoch 9: val_accuracy did not improve from 0.60030
1563/1563 [==============================] - 34s 22ms/step - loss: 1.3763 - accuracy: 0.5249 - val_loss: 1.2569 - val_accuracy: 0.5386 - lr: 0.0010
Epoch 10/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.3509 - accuracy: 0.5376
Epoch 10: val_accuracy improved from 0.60030 to 0.63180, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 35s 22ms/step - loss: 1.3510 - accuracy: 0.5376 - val_loss: 1.0326 - val_accuracy: 0.6318 - lr: 0.0010
Epoch 11/100
1563/1563 [==============================] - ETA: 0s - loss: 1.3254 - accuracy: 0.5463
Epoch 11: val_accuracy did not improve from 0.63180
1563/1563 [==============================] - 36s 23ms/step - loss: 1.3254 - accuracy: 0.5463 - val_loss: 1.0348 - val_accuracy: 0.6258 - lr: 0.0010
Epoch 12/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.3049 - accuracy: 0.5532
Epoch 12: val_accuracy did not improve from 0.63180
1563/1563 [==============================] - 35s 23ms/step - loss: 1.3051 - accuracy: 0.5532 - val_loss: 1.1351 - val_accuracy: 0.6115 - lr: 0.0010
Epoch 13/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.2862 - accuracy: 0.5637
Epoch 13: val_accuracy improved from 0.63180 to 0.63310, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 35s 22ms/step - loss: 1.2863 - accuracy: 0.5636 - val_loss: 1.0346 - val_accuracy: 0.6331 - lr: 0.0010
Epoch 14/100
1563/1563 [==============================] - ETA: 0s - loss: 1.2669 - accuracy: 0.5699
Epoch 14: val_accuracy did not improve from 0.63310
1563/1563 [==============================] - 35s 23ms/step - loss: 1.2669 - accuracy: 0.5699 - val_loss: 1.2640 - val_accuracy: 0.5506 - lr: 0.0010
Epoch 15/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.2572 - accuracy: 0.5741
Epoch 15: val_accuracy did not improve from 0.63310
1563/1563 [==============================] - 34s 22ms/step - loss: 1.2573 - accuracy: 0.5740 - val_loss: 1.1206 - val_accuracy: 0.6215 - lr: 0.0010
Epoch 16/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.2447 - accuracy: 0.5805
Epoch 16: val_accuracy did not improve from 0.63310
1563/1563 [==============================] - 35s 23ms/step - loss: 1.2446 - accuracy: 0.5805 - val_loss: 1.1991 - val_accuracy: 0.6057 - lr: 0.0010
Epoch 17/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.2226 - accuracy: 0.5873
Epoch 17: val_accuracy improved from 0.63310 to 0.64720, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 35s 22ms/step - loss: 1.2227 - accuracy: 0.5873 - val_loss: 0.9827 - val_accuracy: 0.6472 - lr: 0.0010
Epoch 18/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.2050 - accuracy: 0.5965
Epoch 18: val_accuracy did not improve from 0.64720
1563/1563 [==============================] - 35s 22ms/step - loss: 1.2047 - accuracy: 0.5966 - val_loss: 1.0118 - val_accuracy: 0.6391 - lr: 0.0010
Epoch 19/100
1563/1563 [==============================] - ETA: 0s - loss: 1.1952 - accuracy: 0.5983
Epoch 19: val_accuracy improved from 0.64720 to 0.65800, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 37s 23ms/step - loss: 1.1952 - accuracy: 0.5983 - val_loss: 0.9626 - val_accuracy: 0.6580 - lr: 0.0010
Epoch 20/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.1898 - accuracy: 0.6032
Epoch 20: val_accuracy improved from 0.65800 to 0.68460, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 37s 24ms/step - loss: 1.1898 - accuracy: 0.6032 - val_loss: 0.9178 - val_accuracy: 0.6846 - lr: 0.0010
Epoch 21/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.1815 - accuracy: 0.6051
Epoch 21: val_accuracy did not improve from 0.68460
1563/1563 [==============================] - 37s 23ms/step - loss: 1.1815 - accuracy: 0.6051 - val_loss: 1.0920 - val_accuracy: 0.6279 - lr: 0.0010
Epoch 22/100
1563/1563 [==============================] - ETA: 0s - loss: 1.1674 - accuracy: 0.6085
Epoch 22: val_accuracy improved from 0.68460 to 0.69460, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 36s 23ms/step - loss: 1.1674 - accuracy: 0.6085 - val_loss: 0.8778 - val_accuracy: 0.6946 - lr: 0.0010
Epoch 23/100
1563/1563 [==============================] - ETA: 0s - loss: 1.1468 - accuracy: 0.6188
Epoch 23: val_accuracy improved from 0.69460 to 0.71950, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 36s 23ms/step - loss: 1.1468 - accuracy: 0.6188 - val_loss: 0.8091 - val_accuracy: 0.7195 - lr: 0.0010
Epoch 24/100
1563/1563 [==============================] - ETA: 0s - loss: 1.1462 - accuracy: 0.6197
Epoch 24: val_accuracy did not improve from 0.71950
1563/1563 [==============================] - 35s 23ms/step - loss: 1.1462 - accuracy: 0.6197 - val_loss: 0.9789 - val_accuracy: 0.6681 - lr: 0.0010
Epoch 25/100
1563/1563 [==============================] - ETA: 0s - loss: 1.1482 - accuracy: 0.6186
Epoch 25: val_accuracy did not improve from 0.71950
1563/1563 [==============================] - 36s 23ms/step - loss: 1.1482 - accuracy: 0.6186 - val_loss: 0.9466 - val_accuracy: 0.6623 - lr: 0.0010
Epoch 26/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.1343 - accuracy: 0.6254
Epoch 26: val_accuracy did not improve from 0.71950
1563/1563 [==============================] - 35s 23ms/step - loss: 1.1342 - accuracy: 0.6254 - val_loss: 0.8559 - val_accuracy: 0.7100 - lr: 0.0010
Epoch 27/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.1221 - accuracy: 0.6244
Epoch 27: val_accuracy did not improve from 0.71950
1563/1563 [==============================] - 36s 23ms/step - loss: 1.1218 - accuracy: 0.6245 - val_loss: 0.9000 - val_accuracy: 0.6829 - lr: 0.0010
Epoch 28/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.1122 - accuracy: 0.6333
Epoch 28: val_accuracy did not improve from 0.71950
1563/1563 [==============================] - 36s 23ms/step - loss: 1.1123 - accuracy: 0.6333 - val_loss: 0.8558 - val_accuracy: 0.7058 - lr: 0.0010
Epoch 29/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.1095 - accuracy: 0.6322
Epoch 29: val_accuracy did not improve from 0.71950
1563/1563 [==============================] - 36s 23ms/step - loss: 1.1096 - accuracy: 0.6322 - val_loss: 0.9187 - val_accuracy: 0.6892 - lr: 0.0010
Epoch 30/100
1563/1563 [==============================] - ETA: 0s - loss: 1.1052 - accuracy: 0.6344
Epoch 30: val_accuracy did not improve from 0.71950
1563/1563 [==============================] - 35s 22ms/step - loss: 1.1052 - accuracy: 0.6344 - val_loss: 0.8275 - val_accuracy: 0.7135 - lr: 0.0010
Epoch 31/100
1563/1563 [==============================] - ETA: 0s - loss: 1.1002 - accuracy: 0.6357
Epoch 31: val_accuracy did not improve from 0.71950
1563/1563 [==============================] - 35s 23ms/step - loss: 1.1002 - accuracy: 0.6357 - val_loss: 1.0448 - val_accuracy: 0.6508 - lr: 0.0010
Epoch 32/100
1563/1563 [==============================] - ETA: 0s - loss: 1.0892 - accuracy: 0.6390
Epoch 32: val_accuracy did not improve from 0.71950
1563/1563 [==============================] - 36s 23ms/step - loss: 1.0892 - accuracy: 0.6390 - val_loss: 0.8500 - val_accuracy: 0.7127 - lr: 0.0010
Epoch 33/100
1563/1563 [==============================] - ETA: 0s - loss: 1.0825 - accuracy: 0.6395
Epoch 33: val_accuracy improved from 0.71950 to 0.73240, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 35s 22ms/step - loss: 1.0825 - accuracy: 0.6395 - val_loss: 0.7995 - val_accuracy: 0.7324 - lr: 0.0010
Epoch 34/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.0815 - accuracy: 0.6441
Epoch 34: val_accuracy did not improve from 0.73240
1563/1563 [==============================] - 36s 23ms/step - loss: 1.0817 - accuracy: 0.6440 - val_loss: 0.7821 - val_accuracy: 0.7324 - lr: 0.0010
Epoch 35/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.0743 - accuracy: 0.6473
Epoch 35: val_accuracy did not improve from 0.73240
1563/1563 [==============================] - 36s 23ms/step - loss: 1.0742 - accuracy: 0.6473 - val_loss: 0.7895 - val_accuracy: 0.7282 - lr: 0.0010
Epoch 36/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.0695 - accuracy: 0.6486
Epoch 36: val_accuracy did not improve from 0.73240
1563/1563 [==============================] - 37s 24ms/step - loss: 1.0693 - accuracy: 0.6487 - val_loss: 0.8054 - val_accuracy: 0.7223 - lr: 0.0010
Epoch 37/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.0633 - accuracy: 0.6510
Epoch 37: val_accuracy improved from 0.73240 to 0.73660, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 35s 23ms/step - loss: 1.0631 - accuracy: 0.6511 - val_loss: 0.7637 - val_accuracy: 0.7366 - lr: 0.0010
Epoch 38/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.0701 - accuracy: 0.6509
Epoch 38: val_accuracy did not improve from 0.73660
1563/1563 [==============================] - 37s 23ms/step - loss: 1.0701 - accuracy: 0.6509 - val_loss: 0.8457 - val_accuracy: 0.7136 - lr: 0.0010
Epoch 39/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.0467 - accuracy: 0.6584
Epoch 39: val_accuracy did not improve from 0.73660
1563/1563 [==============================] - 38s 24ms/step - loss: 1.0463 - accuracy: 0.6585 - val_loss: 0.8873 - val_accuracy: 0.7043 - lr: 0.0010
Epoch 40/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.0546 - accuracy: 0.6548
Epoch 40: val_accuracy did not improve from 0.73660
1563/1563 [==============================] - 37s 24ms/step - loss: 1.0547 - accuracy: 0.6547 - val_loss: 0.8820 - val_accuracy: 0.6985 - lr: 0.0010
Epoch 41/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.0483 - accuracy: 0.6572
Epoch 41: val_accuracy did not improve from 0.73660
1563/1563 [==============================] - 37s 24ms/step - loss: 1.0483 - accuracy: 0.6572 - val_loss: 1.1795 - val_accuracy: 0.6351 - lr: 0.0010
Epoch 42/100
1563/1563 [==============================] - ETA: 0s - loss: 1.0463 - accuracy: 0.6583
Epoch 42: val_accuracy improved from 0.73660 to 0.74890, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 35s 22ms/step - loss: 1.0463 - accuracy: 0.6583 - val_loss: 0.7433 - val_accuracy: 0.7489 - lr: 0.0010
Epoch 43/100
1563/1563 [==============================] - ETA: 0s - loss: 1.0436 - accuracy: 0.6599
Epoch 43: val_accuracy did not improve from 0.74890
1563/1563 [==============================] - 38s 25ms/step - loss: 1.0436 - accuracy: 0.6599 - val_loss: 0.8552 - val_accuracy: 0.7113 - lr: 0.0010
Epoch 44/100
1563/1563 [==============================] - ETA: 0s - loss: 1.0321 - accuracy: 0.6626
Epoch 44: val_accuracy did not improve from 0.74890
1563/1563 [==============================] - 36s 23ms/step - loss: 1.0321 - accuracy: 0.6626 - val_loss: 0.7819 - val_accuracy: 0.7274 - lr: 0.0010
Epoch 45/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.0337 - accuracy: 0.6601
Epoch 45: val_accuracy did not improve from 0.74890
1563/1563 [==============================] - 36s 23ms/step - loss: 1.0336 - accuracy: 0.6602 - val_loss: 0.7743 - val_accuracy: 0.7347 - lr: 0.0010
Epoch 46/100
1563/1563 [==============================] - ETA: 0s - loss: 1.0330 - accuracy: 0.6638
Epoch 46: val_accuracy improved from 0.74890 to 0.75150, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 36s 23ms/step - loss: 1.0330 - accuracy: 0.6638 - val_loss: 0.7259 - val_accuracy: 0.7515 - lr: 0.0010
Epoch 47/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.0208 - accuracy: 0.6666
Epoch 47: val_accuracy improved from 0.75150 to 0.76160, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 35s 23ms/step - loss: 1.0209 - accuracy: 0.6665 - val_loss: 0.7046 - val_accuracy: 0.7616 - lr: 0.0010
Epoch 48/100
1562/1563 [============================>.] - ETA: 0s - loss: 1.0166 - accuracy: 0.6695
Epoch 48: val_accuracy did not improve from 0.76160
1563/1563 [==============================] - 35s 22ms/step - loss: 1.0166 - accuracy: 0.6695 - val_loss: 0.9199 - val_accuracy: 0.6951 - lr: 0.0010
Epoch 49/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.0137 - accuracy: 0.6683
Epoch 49: val_accuracy did not improve from 0.76160
1563/1563 [==============================] - 35s 23ms/step - loss: 1.0139 - accuracy: 0.6682 - val_loss: 0.7090 - val_accuracy: 0.7582 - lr: 0.0010
Epoch 50/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.0255 - accuracy: 0.6638
Epoch 50: val_accuracy did not improve from 0.76160
1563/1563 [==============================] - 35s 22ms/step - loss: 1.0254 - accuracy: 0.6638 - val_loss: 0.7449 - val_accuracy: 0.7444 - lr: 0.0010
Epoch 51/100
1563/1563 [==============================] - ETA: 0s - loss: 1.0145 - accuracy: 0.6688
Epoch 51: val_accuracy did not improve from 0.76160
1563/1563 [==============================] - 38s 24ms/step - loss: 1.0145 - accuracy: 0.6688 - val_loss: 0.8169 - val_accuracy: 0.7207 - lr: 0.0010
Epoch 52/100
1563/1563 [==============================] - ETA: 0s - loss: 1.0186 - accuracy: 0.6681
Epoch 52: val_accuracy did not improve from 0.76160
1563/1563 [==============================] - 37s 24ms/step - loss: 1.0186 - accuracy: 0.6681 - val_loss: 0.7924 - val_accuracy: 0.7355 - lr: 0.0010
Epoch 53/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.0113 - accuracy: 0.6686
Epoch 53: val_accuracy did not improve from 0.76160
1563/1563 [==============================] - 36s 23ms/step - loss: 1.0113 - accuracy: 0.6685 - val_loss: 0.7319 - val_accuracy: 0.7498 - lr: 0.0010
Epoch 54/100
1563/1563 [==============================] - ETA: 0s - loss: 1.0109 - accuracy: 0.6712
Epoch 54: val_accuracy did not improve from 0.76160
1563/1563 [==============================] - 36s 23ms/step - loss: 1.0109 - accuracy: 0.6712 - val_loss: 1.0983 - val_accuracy: 0.6648 - lr: 0.0010
Epoch 55/100
1561/1563 [============================>.] - ETA: 0s - loss: 1.0129 - accuracy: 0.6713
Epoch 55: val_accuracy did not improve from 0.76160
1563/1563 [==============================] - 36s 23ms/step - loss: 1.0129 - accuracy: 0.6713 - val_loss: 0.8429 - val_accuracy: 0.7179 - lr: 0.0010
Epoch 56/100
1561/1563 [============================>.] - ETA: 0s - loss: 0.9992 - accuracy: 0.6728
Epoch 56: val_accuracy did not improve from 0.76160
1563/1563 [==============================] - 37s 23ms/step - loss: 0.9993 - accuracy: 0.6729 - val_loss: 0.8258 - val_accuracy: 0.7244 - lr: 0.0010
Epoch 57/100
1563/1563 [==============================] - ETA: 0s - loss: 0.9956 - accuracy: 0.6758
Epoch 57: val_accuracy improved from 0.76160 to 0.77110, saving model to /content/drive/MyDrive/checkpoints/cifar10_best_model.h5
1563/1563 [==============================] - 37s 24ms/step - loss: 0.9956 - accuracy: 0.6758 - val_loss: 0.6726 - val_accuracy: 0.7711 - lr: 0.0010
Epoch 58/100
1563/1563 [==============================] - ETA: 0s - loss: 1.0006 - accuracy: 0.6737
Epoch 58: val_accuracy did not improve from 0.77110
1563/1563 [==============================] - 36s 23ms/step - loss: 1.0006 - accuracy: 0.6737 - val_loss: 0.8017 - val_accuracy: 0.7251 - lr: 0.0010
Epoch 59/100
1563/1563 [==============================] - ETA: 0s - loss: 0.9981 - accuracy: 0.6732
Epoch 59: val_accuracy did not improve from 0.77110
1563/1563 [==============================] - 37s 24ms/step - loss: 0.9981 - accuracy: 0.6732 - val_loss: 0.7429 - val_accuracy: 0.7520 - lr: 0.0010
Epoch 60/100
1562/1563 [============================>.] - ETA: 0s - loss: 0.9920 - accuracy: 0.6789
Epoch 60: val_accuracy did not improve from 0.77110
1563/1563 [==============================] - 38s 24ms/step - loss: 0.9920 - accuracy: 0.6789 - val_loss: 0.7769 - val_accuracy: 0.7382 - lr: 0.0010
Epoch 61/100
1563/1563 [==============================] - ETA: 0s - loss: 0.9926 - accuracy: 0.6794
Epoch 61: val_accuracy did not improve from 0.77110
1563/1563 [==============================] - 36s 23ms/step - loss: 0.9926 - accuracy: 0.6794 - val_loss: 0.8365 - val_accuracy: 0.7146 - lr: 0.0010
Epoch 62/100
1563/1563 [==============================] - ETA: 0s - loss: 0.9928 - accuracy: 0.6751
Epoch 62: val_accuracy did not improve from 0.77110
1563/1563 [==============================] - 37s 23ms/step - loss: 0.9928 - accuracy: 0.6751 - val_loss: 0.7080 - val_accuracy: 0.7619 - lr: 0.0010
Epoch 63/100
1562/1563 [============================>.] - ETA: 0s - loss: 0.9901 - accuracy: 0.6787
Epoch 63: val_accuracy did not improve from 0.77110
1563/1563 [==============================] - 36s 23ms/step - loss: 0.9901 - accuracy: 0.6787 - val_loss: 0.8295 - val_accuracy: 0.7269 - lr: 0.0010
Epoch 64/100
1562/1563 [============================>.] - ETA: 0s - loss: 0.9890 - accuracy: 0.6796
Epoch 64: val_accuracy did not improve from 0.77110
1563/1563 [==============================] - 37s 24ms/step - loss: 0.9888 - accuracy: 0.6796 - val_loss: 0.7512 - val_accuracy: 0.7521 - lr: 0.0010
Epoch 65/100
1563/1563 [==============================] - ETA: 0s - loss: 0.9867 - accuracy: 0.6767
Epoch 65: val_accuracy did not improve from 0.77110
1563/1563 [==============================] - 35s 23ms/step - loss: 0.9867 - accuracy: 0.6767 - val_loss: 0.7518 - val_accuracy: 0.7448 - lr: 0.0010
Epoch 66/100
1562/1563 [============================>.] - ETA: 0s - loss: 0.9789 - accuracy: 0.6827
Epoch 66: val_accuracy did not improve from 0.77110
1563/1563 [==============================] - 37s 24ms/step - loss: 0.9788 - accuracy: 0.6828 - val_loss: 0.7259 - val_accuracy: 0.7563 - lr: 0.0010
Epoch 67/100
1562/1563 [============================>.] - ETA: 0s - loss: 0.9799 - accuracy: 0.6839
Epoch 67: val_accuracy did not improve from 0.77110
Restoring model weights from the end of the best epoch: 57.
1563/1563 [==============================] - 37s 24ms/step - loss: 0.9799 - accuracy: 0.6838 - val_loss: 0.8824 - val_accuracy: 0.7192 - lr: 0.0010
Epoch 67: early stopping

avaliando o treinamento¶

Vamos dar uma olhada n curva de loss e acuracia do treinamento

In [20]:

from matplotlib import pyplot as plt

## exibe history com plot de loss e
#acuracia
def plot_history(history):
  # summarize history for loss
  plt.plot(history.history['loss'])
  plt.plot(history.history['val_loss'])
  plt.title('Model loss')
  plt.ylabel('loss')
  plt.xlabel('epoch')
  plt.legend(['train', 'test'], loc='upper left')
  plt.show()
  # summarize history for accuracy
  plt.plot(history.history['accuracy'])
  plt.plot(history.history['val_accuracy'])
  plt.title('model accuracy')
  plt.ylabel('accuracy')
  plt.xlabel('epoch')
  plt.legend(['train', 'test'], loc='upper left')
  plt.show()

plot_history(history)

from matplotlib import pyplot as plt ## exibe history com plot de loss e #acuracia def plot_history(history): # summarize history for loss plt.plot(history.history['loss']) plt.plot(history.history['val_loss']) plt.title('Model loss') plt.ylabel('loss') plt.xlabel('epoch') plt.legend(['train', 'test'], loc='upper left') plt.show() # summarize history for accuracy plt.plot(history.history['accuracy']) plt.plot(history.history['val_accuracy']) plt.title('model accuracy') plt.ylabel('accuracy') plt.xlabel('epoch') plt.legend(['train', 'test'], loc='upper left') plt.show() plot_history(history)

Avaliando o Modelo e Carregando o Melhor Modelo¶

Vamos carregar conjunto de dados de teste e normalize os valores dos pixels:

In [11]:

# Carregando o conjunto de dados de teste CIFAR-10
(x_test, y_test) = cifar10.load_data()[1]

# Normalizando os valores dos pixels para o intervalo [0, 1]
x_test = x_test.astype('float32') / 255.0

# Convertendo os rótulos para vetores one-hot
y_test = to_categorical(y_test, 10)

# Carregando o conjunto de dados de teste CIFAR-10 (x_test, y_test) = cifar10.load_data()[1] # Normalizando os valores dos pixels para o intervalo [0, 1] x_test = x_test.astype('float32') / 255.0 # Convertendo os rótulos para vetores one-hot y_test = to_categorical(y_test, 10)

Avaliação do Modelo¶

Vamos avaliar o modelo nos dados de teste.

Durante o treinamento salvamos no google drive o modelo para garantir que estamos utilizando o modelo com melhor desempenho:

In [13]:

# Carregando o melhor modelo salvo durante o treinamento
best_model = tf.keras.models.load_model(model_save_path)

# Avaliando o melhor modelo salvo nos dados de teste
best_test_loss, best_test_accuracy = best_model.evaluate(x_test, y_test, verbose=2)
print(f'Best test loss: {best_test_loss}')
print(f'Best test accuracy: {best_test_accuracy}')

# Carregando o melhor modelo salvo durante o treinamento best_model = tf.keras.models.load_model(model_save_path) # Avaliando o melhor modelo salvo nos dados de teste best_test_loss, best_test_accuracy = best_model.evaluate(x_test, y_test, verbose=2) print(f'Best test loss: {best_test_loss}') print(f'Best test accuracy: {best_test_accuracy}')

313/313 - 1s - loss: 0.6726 - accuracy: 0.7711 - 1s/epoch - 4ms/step
Best test loss: 0.6726096868515015
Best test accuracy: 0.7710999846458435

fazendo predições nas imagens de teste¶

vamos fazer previsões no conjunto de dados de teste:

In [21]:

import numpy as np

# Fazendo previsões no conjunto de dados de teste
predictions = best_model.predict(x_test)
predicted_classes = np.argmax(predictions, axis=1)
true_classes = np.argmax(y_test, axis=1)

# Nomes das classes CIFAR-10
class_names = ['Airplane', 'Automobile', 'Bird', 'Cat', 'Deer', 'Dog', 'Frog', 'Horse', 'Ship', 'Truck']

import numpy as np # Fazendo previsões no conjunto de dados de teste predictions = best_model.predict(x_test) predicted_classes = np.argmax(predictions, axis=1) true_classes = np.argmax(y_test, axis=1) # Nomes das classes CIFAR-10 class_names = ['Airplane', 'Automobile', 'Bird', 'Cat', 'Deer', 'Dog', 'Frog', 'Horse', 'Ship', 'Truck']

313/313 [==============================] - 1s 2ms/step

Explorando a matriz de confusão¶

A matriz de confusão é uma ferramenta para avaliar a performance de um modelo de classificação. Ela apresenta uma tabela que resume os resultados das previsões do modelo, comparando os rótulos previstos com os rótulos reais.

Como Interpretar:

• True Positives (TP): Número de previsões corretas da classe positiva.
• True Negatives (TN): Número de previsões corretas da classe negativa.
• False Positives (FP): Número de previsões incorretas onde a classe negativa foi prevista como positiva (erro tipo I).
• False Negatives (FN): Número de previsões incorretas onde a classe positiva foi prevista como negativa (erro tipo II).

A matriz de confusão ajuda a entender não apenas a acurácia do modelo, mas também os tipos de erros que ele comete. Isso pode ser útil para:

• Identificar classes que são frequentemente confundidas.
• Melhorar o modelo ao ajustar os hiperparâmetros ou coletar mais dados para classes específicas.

In [23]:

import seaborn as sns
from sklearn.metrics import confusion_matrix

# Função para plotar a matriz de confusão
def plot_confusion_matrix(true_labels, predicted_labels, class_names):
    cm = confusion_matrix(true_labels, predicted_labels)
    plt.figure(figsize=(10, 8))
    sns.heatmap(cm, annot=True, fmt='d', cmap='Blues', xticklabels=class_names, yticklabels=class_names)
    plt.xlabel('Predicted')
    plt.ylabel('True')
    plt.show()


# Visualizando a matriz de confusão
print("Matriz de Confusão:")
plot_confusion_matrix(true_classes, predicted_classes, class_names)

import seaborn as sns from sklearn.metrics import confusion_matrix # Função para plotar a matriz de confusão def plot_confusion_matrix(true_labels, predicted_labels, class_names): cm = confusion_matrix(true_labels, predicted_labels) plt.figure(figsize=(10, 8)) sns.heatmap(cm, annot=True, fmt='d', cmap='Blues', xticklabels=class_names, yticklabels=class_names) plt.xlabel('Predicted') plt.ylabel('True') plt.show() # Visualizando a matriz de confusão print("Matriz de Confusão:") plot_confusion_matrix(true_classes, predicted_classes, class_names)

Matriz de Confusão:

Exibindo os resultados¶

Vamos plotar algumas imagens para visualizar os resultados

In [25]:

import matplotlib.pyplot as plt

# Função para plotar imagens com rótulos verdadeiros e previstos
def plot_images(images, true_labels, predicted_labels, class_names, num_images=10):
    plt.figure(figsize=(15, 15))
    for i in range(num_images):
        plt.subplot(5, 5, i + 1)
        plt.imshow(images[i])
        plt.title(f'True: {class_names[true_labels[i]]}\nPred: {class_names[predicted_labels[i]]}')
        plt.axis('off')
    plt.show()

# Função para plotar imagens de resultados corretos
def plot_correct_predictions(images, true_labels, predicted_labels, class_names):
    correct_indices = np.where(predicted_labels == true_labels)[0]
    plot_images(images[correct_indices], true_labels[correct_indices], predicted_labels[correct_indices], class_names)

# Função para plotar imagens de resultados incorretos
def plot_incorrect_predictions(images, true_labels, predicted_labels, class_names):
    incorrect_indices = np.where(predicted_labels != true_labels)[0]
    plot_images(images[incorrect_indices], true_labels[incorrect_indices], predicted_labels[incorrect_indices], class_names)

import matplotlib.pyplot as plt # Função para plotar imagens com rótulos verdadeiros e previstos def plot_images(images, true_labels, predicted_labels, class_names, num_images=10): plt.figure(figsize=(15, 15)) for i in range(num_images): plt.subplot(5, 5, i + 1) plt.imshow(images[i]) plt.title(f'True: {class_names[true_labels[i]]}\nPred: {class_names[predicted_labels[i]]}') plt.axis('off') plt.show() # Função para plotar imagens de resultados corretos def plot_correct_predictions(images, true_labels, predicted_labels, class_names): correct_indices = np.where(predicted_labels == true_labels)[0] plot_images(images[correct_indices], true_labels[correct_indices], predicted_labels[correct_indices], class_names) # Função para plotar imagens de resultados incorretos def plot_incorrect_predictions(images, true_labels, predicted_labels, class_names): incorrect_indices = np.where(predicted_labels != true_labels)[0] plot_images(images[incorrect_indices], true_labels[incorrect_indices], predicted_labels[incorrect_indices], class_names)

chamando as funções

In [26]:

# Visualizando algumas imagens de resultados corretos
print("Imagens de Resultados Corretos:")
plot_correct_predictions(x_test, true_classes, predicted_classes, class_names)

# Visualizando algumas imagens de resultados incorretos
print("Imagens de Resultados Incorretos:")
plot_incorrect_predictions(x_test, true_classes, predicted_classes, class_names)

# Visualizando algumas imagens de resultados corretos print("Imagens de Resultados Corretos:") plot_correct_predictions(x_test, true_classes, predicted_classes, class_names) # Visualizando algumas imagens de resultados incorretos print("Imagens de Resultados Incorretos:") plot_incorrect_predictions(x_test, true_classes, predicted_classes, class_names)

Imagens de Resultados Corretos:

Imagens de Resultados Incorretos: