In [0]:
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
Estimated completion time: 30 minutes
In this notebook we will build on the model we created in Exercise 1 to classify cats vs. dogs, and improve accuracy by employing a couple strategies to reduce overfitting: data augmentation and dropout.
We will follow these steps:
Let's get started!
Let's get familiar with the concept of data augmentation, an essential way to fight overfitting for computer vision models.
In order to make the most of our few training examples, we will "augment" them via a number of random transformations, so that at training time, our model will never see the exact same picture twice. This helps prevent overfitting and helps the model generalize better.
This can be done by configuring a number of random transformations to be performed on the images read by our ImageDataGenerator instance. Let's get started with an example:
In [0]:
from tensorflow.keras.preprocessing.image import ImageDataGenerator
datagen = ImageDataGenerator(
rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest')
These are just a few of the options available (for more, see the Keras documentation. Let's quickly go over what we just wrote:
rotation_range is a value in degrees (0–180), a range within which to randomly rotate pictures.width_shift and height_shift are ranges (as a fraction of total width or height) within which to randomly translate pictures vertically or horizontally.shear_range is for randomly applying shearing transformations.zoom_range is for randomly zooming inside pictures.horizontal_flip is for randomly flipping half of the images horizontally. This is relevant when there are no assumptions of horizontal assymmetry (e.g. real-world pictures).fill_mode is the strategy used for filling in newly created pixels, which can appear after a rotation or a width/height shift.Let's take a look at our augmented images. First let's set up our example files, as in Exercise 1.
NOTE: The 2,000 images used in this exercise are excerpted from the "Dogs vs. Cats" dataset available on Kaggle, which contains 25,000 images. Here, we use a subset of the full dataset to decrease training time for educational purposes.
In [0]:
!wget --no-check-certificate \
https://storage.googleapis.com/mledu-datasets/cats_and_dogs_filtered.zip -O \
/tmp/cats_and_dogs_filtered.zip
In [0]:
import os
import zipfile
local_zip = '/tmp/cats_and_dogs_filtered.zip'
zip_ref = zipfile.ZipFile(local_zip, 'r')
zip_ref.extractall('/tmp')
zip_ref.close()
base_dir = '/tmp/cats_and_dogs_filtered'
train_dir = os.path.join(base_dir, 'train')
validation_dir = os.path.join(base_dir, 'validation')
# Directory with our training cat pictures
train_cats_dir = os.path.join(train_dir, 'cats')
# Directory with our training dog pictures
train_dogs_dir = os.path.join(train_dir, 'dogs')
# Directory with our validation cat pictures
validation_cats_dir = os.path.join(validation_dir, 'cats')
# Directory with our validation dog pictures
validation_dogs_dir = os.path.join(validation_dir, 'dogs')
train_cat_fnames = os.listdir(train_cats_dir)
train_dog_fnames = os.listdir(train_dogs_dir)
Next, let's apply the datagen transformations to a cat image from the training set to produce five random variants. Rerun the cell a few times to see fresh batches of random variants.
In [0]:
%matplotlib inline
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
from tensorflow.keras.preprocessing.image import array_to_img, img_to_array, load_img
img_path = os.path.join(train_cats_dir, train_cat_fnames[2])
img = load_img(img_path, target_size=(150, 150)) # this is a PIL image
x = img_to_array(img) # Numpy array with shape (150, 150, 3)
x = x.reshape((1,) + x.shape) # Numpy array with shape (1, 150, 150, 3)
# The .flow() command below generates batches of randomly transformed images
# It will loop indefinitely, so we need to `break` the loop at some point!
i = 0
for batch in datagen.flow(x, batch_size=1):
plt.figure(i)
imgplot = plt.imshow(array_to_img(batch[0]))
i += 1
if i % 5 == 0:
break
Now let's add our data-augmentation transformations from Exploring Data Augmentation to our data preprocessing configuration:
In [0]:
# Adding rescale, rotation_range, width_shift_range, height_shift_range,
# shear_range, zoom_range, and horizontal flip to our ImageDataGenerator
train_datagen = ImageDataGenerator(
rescale=1./255,
rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,)
# Note that the validation data should not be augmented!
val_datagen = ImageDataGenerator(rescale=1./255)
# Flow training images in batches of 32 using train_datagen generator
train_generator = train_datagen.flow_from_directory(
train_dir, # This is the source directory for training images
target_size=(150, 150), # All images will be resized to 150x150
batch_size=20,
# Since we use binary_crossentropy loss, we need binary labels
class_mode='binary')
# Flow validation images in batches of 32 using val_datagen generator
validation_generator = val_datagen.flow_from_directory(
validation_dir,
target_size=(150, 150),
batch_size=20,
class_mode='binary')
If we train a new network using this data augmentation configuration, our network will never see the same input twice. However the inputs that it sees are still heavily intercorrelated, so this might not be quite enough to completely get rid of overfitting.
TIP: To learn more about dropout, see Training Neural Networks in Machine Learning Crash Course.
Let's reconfigure our convnet architecture from Exercise 1 to add some dropout, right before the final classification layer:
In [0]:
from tensorflow.keras import layers
from tensorflow.keras import Model
from tensorflow.keras.optimizers import RMSprop
# Our input feature map is 150x150x3: 150x150 for the image pixels, and 3 for
# the three color channels: R, G, and B
img_input = layers.Input(shape=(150, 150, 3))
# First convolution extracts 16 filters that are 3x3
# Convolution is followed by max-pooling layer with a 2x2 window
x = layers.Conv2D(16, 3, activation='relu')(img_input)
x = layers.MaxPooling2D(2)(x)
# Second convolution extracts 32 filters that are 3x3
# Convolution is followed by max-pooling layer with a 2x2 window
x = layers.Conv2D(32, 3, activation='relu')(x)
x = layers.MaxPooling2D(2)(x)
# Third convolution extracts 64 filters that are 3x3
# Convolution is followed by max-pooling layer with a 2x2 window
x = layers.Convolution2D(64, 3, activation='relu')(x)
x = layers.MaxPooling2D(2)(x)
# Flatten feature map to a 1-dim tensor
x = layers.Flatten()(x)
# Create a fully connected layer with ReLU activation and 512 hidden units
x = layers.Dense(512, activation='relu')(x)
# Add a dropout rate of 0.5
x = layers.Dropout(0.5)(x)
# Create output layer with a single node and sigmoid activation
output = layers.Dense(1, activation='sigmoid')(x)
# Configure and compile the model
model = Model(img_input, output)
model.compile(loss='binary_crossentropy',
optimizer=RMSprop(lr=0.001),
metrics=['acc'])
In [0]:
# WRITE CODE TO TRAIN THE MODEL ON ALL 2000 IMAGES FOR 30 EPOCHS, AND VALIDATE
# ON ALL 1,000 VALIDATION IMAGES
In [0]:
history = model.fit_generator(
train_generator,
steps_per_epoch=100,
epochs=30,
validation_data=validation_generator,
validation_steps=50,
verbose=2)
Note that with data augmentation in place, the 2,000 training images are randomly transformed each time a new training epoch runs, which means that the model will never see the same image twice during training.
In [0]:
# Retrieve a list of accuracy results on training and validation data
# sets for each training epoch
acc = history.history['acc']
val_acc = history.history['val_acc']
# Retrieve a list of list results on training and validation data
# sets for each training epoch
loss = history.history['loss']
val_loss = history.history['val_loss']
# Get number of epochs
epochs = range(len(acc))
# Plot training and validation accuracy per epoch
plt.plot(epochs, acc)
plt.plot(epochs, val_acc)
plt.title('Training and validation accuracy')
plt.figure()
# Plot training and validation loss per epoch
plt.plot(epochs, loss)
plt.plot(epochs, val_loss)
plt.title('Training and validation loss')
Much better! We are no longer overfitting, and we have gained ~3 validation accuracy percentage points (see the green line in the top chart). In fact, judging by our training profile, we could keep fitting our model for 30+ more epochs and we could probably get to ~80%!
In [0]:
import os, signal
os.kill(os.getpid(), signal.SIGKILL)