Image Classification using tf.keras CNN questions about kernel and how to create test dataset?

Question

I have a colab with which I am working. It is about using keras to build CNN. My code so far which works is as follows (epochs set to 5 due to time reasons):

import os as os
import numpy as np
import glob as gl
import matplotlib.pyplot as plt
import shutil

import tensorflow as tf

from tensorflow.keras.preprocessing.image import ImageDataGenerator

URL = 'https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz'

zip_dir = tf.keras.utils.get_file('flower_photos.tgz', URL, extract=True)

base_dir = os.path.join(os.path.dirname(zip_dir), 'flower_photos')

labels = ['daisy', 'dandelion', 'roses', 'sunflowers', 'tulips']

for label in labels:
  imagepath = os.path.join(base_dir, label)
  images = gl.glob(imagepath + '/*.jpg')
  print("{}: {} Images".format(label, len(images)))
  train, val = images[:round(len(images)*0.8)], images[round(len(images)*0.8):]

  for t in train:
    if not os.path.exists(os.path.join(base_dir, 'train', label)):
      os.makedirs(os.path.join(base_dir, 'train', label))
    shutil.move(t, os.path.join(base_dir, 'train', label))

  for v in val:
    if not os.path.exists(os.path.join(base_dir, 'val', label)):
      os.makedirs(os.path.join(base_dir, 'val', label))
    shutil.move(v, os.path.join(base_dir, 'val', label))

traindir = os.path.join(base_dir, 'train')
val_dir = os.path.join(base_dir, 'val')

batch_size = 100
img_size = 150

image_gen_train = ImageDataGenerator(rescale = 1./255,
    rotation_range=45,
    width_shift_range=0.15,
    zoom_range=0.5,
    height_shift_range=0.15,
    horizontal_flip=True)


image_gen_val = ImageDataGenerator(rescale = 1./255)


train_data_gen = image_gen_train.flow_from_directory(batch_size=batch_size,
                                               directory=traindir,
                                               shuffle=True,
                                               target_size=(img_size,img_size),
                                               class_mode='sparse')

val_data_gen = image_gen_val.flow_from_directory(batch_size=batch_size,
                                               directory=val_dir,
                                               target_size=(img_size,img_size),
                                               class_mode='sparse')

def plotImages(images_arr):
    fig, axes = plt.subplots(1, 5, figsize=(20,20))
    axes = axes.flatten()
    for img, ax in zip( images_arr, axes):
        ax.imshow(img)
    plt.tight_layout()
    plt.show()


augmented_images = [train_data_gen[0][0][0] for i in range(5)]
plotImages(augmented_images)

model = tf.keras.models.Sequential([
    tf.keras.layers.Conv2D(16, (3,3), padding='same',  activation='relu', input_shape=(img_size, img_size, 3)),
    tf.keras.layers.MaxPooling2D(2, 2),

    tf.keras.layers.Conv2D(32, (3,3), padding='same', activation='relu'),
    tf.keras.layers.MaxPooling2D(2,2),

    tf.keras.layers.Conv2D(64, (3,3), padding='same', activation='relu'),
    tf.keras.layers.MaxPooling2D(2,2),

    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(512, activation='relu'),
    tf.keras.layers.Dense(5, activation='softmax')
])

model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

model.summary()

epochs = 5

history = model.fit(
    train_data_gen,
    epochs=epochs,
    validation_data=val_data_gen
)

Now I have two questions (the second one being more important): I use a kernel of 3,3 ( tf.keras.layers.Conv2D(16, (3,3)). In the solution however, they just use:

model.add(Conv2D(16, 3, padding='same', activation='relu', input_shape=(IMG_SHAPE,IMG_SHAPE, 3)))

So just the integer 3. Why this? The task clearly states "All convolutional filters should be 3 x 3".

My second question: I would like to have not only a train and validation set, but also a test set. At the end I want to check my performance on the test dataset. So I do use the training and validation dataset in the model training, but furthermore at the end I want to compare it with the following statment:

# Evaluate against the test set.
print("
 Evaluate the new model against the test set:")
my_model.evaluate(x=x_test_normalized, y=y_test, batch_size=batch_size)

How does the code needs to be modified in order to get a training, validation and test set?

I tried to use the validation split and hoped I could use the up to know called validation set as a test set:

history = model.fit(
    train_data_gen,
    epochs=epochs,
    validation_split= 0.2
)

my_model.evaluate(x=val_data_gen, y=train_data_gen, batch_size=batch_size)

However, this does not work as my input in the model fit does not support validation split, the error message tells me.

Image Classification using tf.keras CNN questions about kernel and how to create test dataset?

Answers (1)

Related Questions