Skip to main content

Angkor Wat, Cambodia

Tf Image Classifier

EfficientNetV2B0

import cv2 as cv
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
from sklearn.metrics import (
classification_report,
confusion_matrix,
ConfusionMatrixDisplay)
import seaborn as sns

import tensorflow as tf
from tensorflow.io import TFRecordWriter
from tensorflow.keras import Sequential
from tensorflow.keras.callbacks import (
Callback,
CSVLogger,
EarlyStopping,
LearningRateScheduler,
ModelCheckpoint
)
from tensorflow.keras.layers import (
Layer,
GlobalAveragePooling2D,
Conv2D,
MaxPool2D,
Dense,
Flatten,
InputLayer,
BatchNormalization,
Input,
Dropout,
RandomFlip,
RandomRotation,
RandomContrast,
RandomBrightness,
Resizing,
Rescaling
)
from tensorflow.keras.losses import BinaryCrossentropy, CategoricalCrossentropy, SparseCategoricalCrossentropy
from tensorflow.keras.metrics import CategoricalAccuracy, TopKCategoricalAccuracy, SparseCategoricalAccuracy
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.regularizers import L2, L1
from tensorflow.keras.utils import image_dataset_from_directory
from tensorflow.train import Feature, Features, Example, BytesList, Int64List
BATCH = 32
SIZE = 224
SEED = 42

EPOCHS = 20
LR = 0.001
FILTERS = 6
KERNEL = 3
STRIDES = 1
REGRATE = 0.0
POOL = 2
DORATE = 0.05
LABELS = ['Gladiolus', 'Adenium', 'Alpinia_Purpurata', 'Alstroemeria', 'Amaryllis', 'Anthurium_Andraeanum', 'Antirrhinum', 'Aquilegia', 'Billbergia_Pyramidalis', 'Cattleya', 'Cirsium', 'Coccinia_Grandis', 'Crocus', 'Cyclamen', 'Dahlia', 'Datura_Metel', 'Dianthus_Barbatus', 'Digitalis', 'Echinacea_Purpurea', 'Echinops_Bannaticus', 'Fritillaria_Meleagris', 'Gaura', 'Gazania', 'Gerbera', 'Guzmania', 'Helianthus_Annuus', 'Iris_Pseudacorus', 'Leucanthemum', 'Malvaceae', 'Narcissus_Pseudonarcissus', 'Nerine', 'Nymphaea_Tetragona', 'Paphiopedilum', 'Passiflora', 'Pelargonium', 'Petunia', 'Platycodon_Grandiflorus', 'Plumeria', 'Poinsettia', 'Primula', 'Protea_Cynaroides', 'Rose', 'Rudbeckia', 'Strelitzia_Reginae', 'Tropaeolum_Majus', 'Tussilago', 'Viola', 'Zantedeschia_Aethiopica']
NLABELS = len(LABELS)
DENSE1 = 1024
DENSE2 = 128

Dataset

train_directory = '../dataset/Flower_Dataset/split/train'
test_directory = '../dataset/Flower_Dataset/split/val'
train_dataset = image_dataset_from_directory(
train_directory,
labels='inferred',
label_mode='categorical',
class_names=LABELS,
color_mode='rgb',
batch_size=BATCH,
image_size=(SIZE, SIZE),
shuffle=True,
seed=SEED,
interpolation='bilinear',
follow_links=False,
crop_to_aspect_ratio=False
)

# Found 9206 files belonging to 48 classes.
test_dataset = image_dataset_from_directory(
test_directory,
labels='inferred',
label_mode='categorical',
class_names=LABELS,
color_mode='rgb',
batch_size=BATCH,
image_size=(SIZE, SIZE),
shuffle=True,
seed=SEED
)

# Found 3090 files belonging to 48 classes.
data_augmentation = Sequential([
# Resizing(224, 224),
RandomRotation(factor=0.25),
RandomFlip(mode='horizontal'),
RandomContrast(factor=0.1),
RandomBrightness(0.1)
],
name="img_augmentation",
)
training_dataset = (
train_dataset
.map(lambda image, label: (data_augmentation(image), label))
.prefetch(tf.data.AUTOTUNE)
)


testing_dataset = (
test_dataset.prefetch(
tf.data.AUTOTUNE
)
)

Building the Efficient TF Model

# transfer learning

backbone = tf.keras.applications.EfficientNetV2B0(
include_top=False,
weights="imagenet",
input_shape=(SIZE, SIZE, 3),
include_preprocessing=True
)
backbone.trainable = False
efficient_model = tf.keras.Sequential([
Input(shape=(SIZE, SIZE, 3)),
data_augmentation,
backbone,
GlobalAveragePooling2D(),
Dense(DENSE1, activation='relu'),
BatchNormalization(),
Dense(DENSE2, activation='relu'),
Dense(NLABELS, activation='softmax')
])

efficient_model.summary()
checkpoint_callback = ModelCheckpoint(
'../best_weights',
monitor='val_accuracy',
mode='max',
verbose=1,
save_best_only=True
)
loss_function = CategoricalCrossentropy()
metrics = [CategoricalAccuracy(name='accuracy')]
efficient_model.compile(
optimizer = Adam(learning_rate=LR),
loss = loss_function,
metrics = metrics
)

Model Training

efficient_history = efficient_model.fit(
training_dataset,
validation_data = testing_dataset,
epochs = EPOCHS,
verbose = 1
)

# loss: 0.2039
# accuracy: 0.9343
# val_loss: 0.3764
# val_accuracy: 0.9026

Model Evaluation

efficient_model.evaluate(testing_dataset)
# loss: 0.3764 - accuracy: 0.9026

Model Finetuning

backbone.trainable = True
efficient_model.compile(
optimizer = Adam(learning_rate=LR/100),
loss = loss_function,
metrics = metrics,
# callbacks = ['checkpoint_callback']
)
efficient_history = efficient_model.fit(
training_dataset,
validation_data = testing_dataset,
epochs = EPOCHS,
verbose = 1
)

# loss: 0.2119
# accuracy: 0.9344
# val_loss: 0.3779
# val_accuracy: 0.9084

Model Evaluation

efficient_model.evaluate(testing_dataset)
# loss: 0.3490 - accuracy: 0.9013
plt.plot(efficient_history.history['loss'])
plt.plot(efficient_history.history['val_loss'])
plt.title('Model Loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train_loss', 'val_loss'])
plt.savefig('assets/EfficientNetV2B0_01.webp', bbox_inches='tight')

tf Emotion Detection

plt.plot(efficient_history.history['accuracy'])
plt.plot(efficient_history.history['val_accuracy'])
plt.title('Model Accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train_accuracy', 'val_accuracy'])
plt.savefig('assets/EfficientNetV2B0_02.webp', bbox_inches='tight')

tf Emotion Detection

test_image_bgr = cv.imread('../dataset/snapshots/Viola_Tricolor.jpg')
test_image_resized = cv.resize(test_image_bgr, (SIZE, SIZE))
test_image_rgb = cv.cvtColor(test_image_resized, cv.COLOR_BGR2RGB)
img = tf.constant(test_image_rgb, dtype=tf.float32)
img = tf.expand_dims(img, axis=0)

probs = efficient_model(img).numpy()
label = LABELS[tf.argmax(probs, axis=1).numpy()[0]]

print(label, str(probs[0]))

plt.imshow(test_image_rgb)
plt.title(label)
plt.axis('off')

plt.savefig('assets/EfficientNetV2B0_Prediction_01.webp', bbox_inches='tight')

tf Emotion Detection

test_image_bgr = cv.imread('../dataset/snapshots/Strelitzia.jpg')
test_image_resized = cv.resize(test_image_bgr, (SIZE, SIZE))
test_image_rgb = cv.cvtColor(test_image_resized, cv.COLOR_BGR2RGB)
img = tf.constant(test_image_rgb, dtype=tf.float32)
img = tf.expand_dims(img, axis=0)

probs = efficient_model(img).numpy()
label = LABELS[tf.argmax(probs, axis=1).numpy()[0]]

print(label, str(probs[0]))

plt.imshow(test_image_rgb)
plt.title(label)
plt.axis('off')

plt.savefig('assets/EfficientNetV2B0_Prediction_02.webp', bbox_inches='tight')

tf Emotion Detection

test_image_bgr = cv.imread('../dataset/snapshots/Water_Lilly.jpg')
test_image_resized = cv.resize(test_image_bgr, (SIZE, SIZE))
test_image_rgb = cv.cvtColor(test_image_resized, cv.COLOR_BGR2RGB)
img = tf.constant(test_image_rgb, dtype=tf.float32)
img = tf.expand_dims(img, axis=0)

probs = efficient_model(img).numpy()
label = LABELS[tf.argmax(probs, axis=1).numpy()[0]]

print(label, str(probs[0]))

plt.imshow(test_image_rgb)
plt.title(label)
plt.axis('off')

plt.savefig('assets/EfficientNetV2B0_Prediction_03.webp', bbox_inches='tight')

tf Emotion Detection

plt.figure(figsize=(16,16))

for images, labels in testing_dataset.take(1):
for i in range(16):
ax = plt.subplot(4,4,i+1)
true = "True: " + LABELS[tf.argmax(labels[i], axis=0).numpy()]
pred = "Predicted: " + LABELS[
tf.argmax(efficient_model(tf.expand_dims(images[i], axis=0)).numpy(), axis=1).numpy()[0]
]
plt.title(
true + "\n" + pred
)
plt.imshow(images[i]/255.)
plt.axis('off')

plt.savefig('assets/EfficientNetV2B0_03.webp', bbox_inches='tight')

tf Emotion Detection

y_pred = []
y_test = []

for img, label in testing_dataset:
y_pred.append(efficient_model(img))
y_test.append(label.numpy())
conf_mtx = ConfusionMatrixDisplay(
confusion_matrix=confusion_matrix(
np.argmax(y_test[:-1], axis=-1).flatten(),
np.argmax(y_pred[:-1], axis=-1).flatten()
),
display_labels=LABELS
)

fig, ax = plt.subplots(figsize=(16,12))
conf_mtx.plot(ax=ax, cmap='plasma', include_values=True, xticks_rotation='vertical',)

plt.savefig('assets/EfficientNetV2B0_04.webp', bbox_inches='tight')

tf Emotion Detection

Saving the Model

tf.keras.saving.save_model(
efficient_model, '../saved_model/efficientv2b0_model', overwrite=True, save_format='tf'
)
# restore the model
restored_model = tf.keras.saving.load_model('../saved_model/efficientv2b0_model')
# Check its architecture
restored_model.summary()
restored_model.evaluate(testing_dataset)
# loss: 0.3779 - accuracy: 0.9084