C2W3: Transfer Learning
C2W3: Transfer Learning#
import tensorflow as tf
import matplotlib.pyplot as plt
import os
from tensorflow.keras import layers, optimizers, losses
from tensorflow.keras.applications.inception_v3 import InceptionV3, preprocess_input
training_dataset_url = 'https://storage.googleapis.com/tensorflow-1-public/course2/week3/horse-or-human.zip'
validation_dataset_url = 'https://storage.googleapis.com/tensorflow-1-public/course2/week3/validation-horse-or-human.zip'
training_dataset_file = tf.keras.utils.get_file(cache_subdir='datasets/horse-or-human/training',
origin=training_dataset_url, extract=False)
validation_dataset_file = tf.keras.utils.get_file(cache_subdir='datasets/horse-or-human/validation',
origin=validation_dataset_url, extract=False)
train_dir = os.path.dirname(training_dataset_file)
validation_dir = os.path.dirname(validation_dataset_file)
train_dataset = tf.keras.utils.image_dataset_from_directory(
train_dir,
label_mode='binary',
batch_size=32,
image_size=(150, 150)
).cache().prefetch(tf.data.AUTOTUNE)
validation_dataset = tf.keras.utils.image_dataset_from_directory(
validation_dir,
label_mode='binary',
batch_size=32,
image_size=(150, 150)
).cache().prefetch(tf.data.AUTOTUNE)
Found 1027 files belonging to 2 classes.
Found 256 files belonging to 2 classes.
model_weights_url = 'https://storage.googleapis.com/mledu-datasets/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5'
local_weights_file = tf.keras.utils.get_file(cache_subdir='models', origin=model_weights_url)
def create_pre_trained_model(local_weights_file):
pre_trained_model = InceptionV3(input_shape=(150, 150, 3),
include_top=False,
weights=None)
pre_trained_model.load_weights(local_weights_file)
pre_trained_model.trainable = False
return pre_trained_model
pre_trained_model = create_pre_trained_model(local_weights_file)
Last layer
avg_pool (GlobalAveragePooling2D) (None, 2048)
predictions (Dense) (None, 1000)
name_layer = 'mixed7'
idx_layer = [layer.name for layer in pre_trained_model.layers].index(name_layer)
print(f'{name_layer} layer is the number {idx_layer+1} with output_shape {pre_trained_model.layers[idx_layer].output_shape}')
# for layer in pre_trained_model.layers[:idx_layer+1]:
# layer.trainable = False
mixed7 layer is the number 229 with output_shape (None, 7, 7, 768)
print(f'Num layers: {len(pre_trained_model.layers)}')
print(f'Num none trainable layers: {[layer.trainable for layer in pre_trained_model.layers].count(False)}')
Num layers: 311
Num none trainable layers: 311
class myCallback(tf.keras.callbacks.Callback):
def on_epoch_end(self, epoch, logs={}):
if(logs.get('accuracy')>0.999):
print("\nReached 99.9% accuracy so cancelling training!")
self.model.stop_training = True
data_augmentation = tf.keras.Sequential([
layers.RandomRotation(0.2),
# layers.RandomTranslation(0, 0.2),
layers.RandomZoom(0.2),
layers.RandomFlip('horizontal')
])
def create_final_model(pre_trained_model, last_output):
inputs = layers.Input(shape=(150, 150, 3))
# x = data_augmentation(inputs)
x = preprocess_input(inputs)
x = pre_trained_model(x)
x = layers.GlobalAveragePooling2D()(x)
# x = layers.Dropout(0.2)(x)
x = layers.Dense(1)(x)
model = tf.keras.Model(inputs, x)
model.compile(optimizer=optimizers.RMSprop(learning_rate=0.0001),
loss=losses.BinaryCrossentropy(from_logits=True),
metrics=['accuracy'])
return model
last_output = pre_trained_model.output
model = create_final_model(pre_trained_model, last_output)
model.summary()
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_2 (InputLayer) [(None, 150, 150, 3)] 0
tf.math.truediv (TFOpLambda (None, 150, 150, 3) 0
)
tf.math.subtract (TFOpLambd (None, 150, 150, 3) 0
a)
inception_v3 (Functional) (None, 3, 3, 2048) 21802784
global_average_pooling2d (G (None, 2048) 0
lobalAveragePooling2D)
dense (Dense) (None, 1) 2049
=================================================================
Total params: 21,804,833
Trainable params: 2,049
Non-trainable params: 21,802,784
_________________________________________________________________
callbacks = myCallback()
history = model.fit(train_dataset,
validation_data = validation_dataset,
epochs = 20,
verbose = 2)
Epoch 1/20
33/33 - 13s - loss: 0.6430 - accuracy: 0.5998 - val_loss: 0.4667 - val_accuracy: 0.7656 - 13s/epoch - 400ms/step
Epoch 2/20
33/33 - 2s - loss: 0.2786 - accuracy: 0.8851 - val_loss: 0.2543 - val_accuracy: 0.8906 - 2s/epoch - 69ms/step
Epoch 3/20
33/33 - 2s - loss: 0.1438 - accuracy: 0.9572 - val_loss: 0.1432 - val_accuracy: 0.9766 - 2s/epoch - 69ms/step
Epoch 4/20
33/33 - 2s - loss: 0.0839 - accuracy: 0.9854 - val_loss: 0.0862 - val_accuracy: 0.9883 - 2s/epoch - 68ms/step
Epoch 5/20
33/33 - 2s - loss: 0.0553 - accuracy: 0.9912 - val_loss: 0.0564 - val_accuracy: 0.9961 - 2s/epoch - 69ms/step
Epoch 6/20
33/33 - 2s - loss: 0.0400 - accuracy: 0.9932 - val_loss: 0.0400 - val_accuracy: 1.0000 - 2s/epoch - 68ms/step
Epoch 7/20
33/33 - 2s - loss: 0.0309 - accuracy: 0.9942 - val_loss: 0.0302 - val_accuracy: 1.0000 - 2s/epoch - 69ms/step
Epoch 8/20
33/33 - 2s - loss: 0.0247 - accuracy: 0.9942 - val_loss: 0.0240 - val_accuracy: 1.0000 - 2s/epoch - 69ms/step
Epoch 9/20
33/33 - 2s - loss: 0.0201 - accuracy: 0.9951 - val_loss: 0.0198 - val_accuracy: 1.0000 - 2s/epoch - 69ms/step
Epoch 10/20
33/33 - 2s - loss: 0.0165 - accuracy: 0.9971 - val_loss: 0.0167 - val_accuracy: 1.0000 - 2s/epoch - 69ms/step
Epoch 11/20
33/33 - 2s - loss: 0.0136 - accuracy: 0.9971 - val_loss: 0.0143 - val_accuracy: 1.0000 - 2s/epoch - 69ms/step
Epoch 12/20
33/33 - 2s - loss: 0.0112 - accuracy: 0.9971 - val_loss: 0.0124 - val_accuracy: 1.0000 - 2s/epoch - 70ms/step
Epoch 13/20
33/33 - 2s - loss: 0.0093 - accuracy: 0.9971 - val_loss: 0.0109 - val_accuracy: 1.0000 - 2s/epoch - 69ms/step
Epoch 14/20
33/33 - 2s - loss: 0.0077 - accuracy: 0.9971 - val_loss: 0.0097 - val_accuracy: 1.0000 - 2s/epoch - 69ms/step
Epoch 15/20
33/33 - 2s - loss: 0.0064 - accuracy: 0.9981 - val_loss: 0.0086 - val_accuracy: 1.0000 - 2s/epoch - 69ms/step
Epoch 16/20
33/33 - 2s - loss: 0.0053 - accuracy: 1.0000 - val_loss: 0.0078 - val_accuracy: 1.0000 - 2s/epoch - 70ms/step
Epoch 17/20
33/33 - 2s - loss: 0.0044 - accuracy: 1.0000 - val_loss: 0.0070 - val_accuracy: 1.0000 - 2s/epoch - 70ms/step
Epoch 18/20
33/33 - 2s - loss: 0.0037 - accuracy: 1.0000 - val_loss: 0.0064 - val_accuracy: 1.0000 - 2s/epoch - 69ms/step
Epoch 19/20
33/33 - 2s - loss: 0.0030 - accuracy: 1.0000 - val_loss: 0.0058 - val_accuracy: 1.0000 - 2s/epoch - 68ms/step
Epoch 20/20
33/33 - 2s - loss: 0.0025 - accuracy: 1.0000 - val_loss: 0.0053 - val_accuracy: 1.0000 - 2s/epoch - 70ms/step
acc = history.history['accuracy']
val_acc = history.history['val_accuracy']
loss = history.history['loss']
val_loss = history.history['val_loss']
plt.figure(figsize=(8, 8))
plt.subplot(1, 2, 1)
plt.plot(acc, label='Training Accuracy')
plt.plot(val_acc, label='Validation Accuracy')
plt.legend(loc='lower right')
plt.title('Training y Validation Accuracy')
plt.subplot(1, 2, 2)
plt.plot(loss, label='Training loss')
plt.plot(val_loss, label='Validation loss')
plt.legend(loc='upper right')
plt.title('Trainind and Validation Loss')
plt.show()
