def trainTopModel(self, trainLabels, validLabels, fold):
from keras import optimizers
print('[INFO] Training top model type ' + (str(self.model.topType)))
trainData, validationData = self.logger.loadBottlenecks(fold)
trainLabels = utils.transformLabelsToCategorical(
trainLabels, self.numClasses)
validationLabels = utils.transformLabelsToCategorical(
validLabels, self.numClasses)
self.model.topModel.compile(optimizer=optimizers.RMSprop(lr=1e-6),
loss='categorical_crossentropy',
metrics=['accuracy'])
import time
t = time.process_time()
historyGenerated = self.model.topModel.fit(
trainData,
trainLabels,
epochs=self.epochs,
batch_size=self.batchSize,
validation_data=(validationData, validationLabels))
trainingTime = time.process_time() - t
self.logger.saveWeightsTopModel(self.model, fold)
return trainingTime, historyGenerated
def trainTopModel(self, trainData, trainLabels, validData, validLabels,
fold):
from keras import optimizers
import math
print('[INFO] Training top model type ' + (str(self.model.topType)))
self.model.topModel.compile(
optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'],
)
trainSamples = len(trainLabels)
validationSamples = len(validLabels)
predictSizeTrain = int(math.ceil(trainSamples / self.batchSize))
predictSizeValidation = int(
math.ceil(validationSamples / self.batchSize))
trainLabels = utils.transformLabelsToCategorical(
trainLabels, self.numClasses)
validLabels = utils.transformLabelsToCategorical(
validLabels, self.numClasses)
# Data augmentation
trainDatagen = ImageDataGenerator(rescale=1. / 255,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True)
generatorTrain = trainDatagen.flow(trainData,
trainLabels,
shuffle=False,
batch_size=self.batchSize)
valDatagen = ImageDataGenerator(rescale=1. / 255)
generatorVal = valDatagen.flow(validData,
validLabels,
shuffle=False,
batch_size=self.batchSize)
import time
t = time.process_time()
historyGenerated = self.model.topModel.fit_generator(
generatorTrain,
steps_per_epoch=predictSizeTrain,
epochs=self.epochs,
validation_data=generatorVal,
validation_steps=predictSizeValidation)
trainingTime = time.process_time() - t
self.logger.saveWeightsTopModel(self.model, fold)
return trainingTime, historyGenerated
def testFineTuneModel(self, validationData, validationLabels, trainingTime,
historyGenerated, fold, numLayersFreeze):
print('[INFO] testing fine tune model')
fullModel = self.addTopModel()
validationSamples = len(validationLabels)
validationLabels = utils.transformLabelsToCategorical(
validationLabels, self.numClasses)
testDatagen = ImageDataGenerator(rescale=1. / 255)
generator = testDatagen.flow(validationData,
validationLabels,
shuffle=False,
batch_size=self.batchSize)
from keras import optimizers
fullModel.compile(optimizer=optimizers.RMSprop(lr=1e-6),
loss='categorical_crossentropy',
metrics=['accuracy'])
testLoss, testAccuracy = fullModel.evaluate_generator(
generator, steps=validationSamples / self.batchSize)
self.logger.saveDataFineTuneModel(historyGenerated,
(testLoss, testAccuracy),
trainingTime, fold, numLayersFreeze,
self.model.topType)
import gc
del fullModel
gc.collect()
def testFineTuneModel(self, validationData, validationLabels, trainingTime,
historyGenerated, fold, numLayersFreeze):
print('[INFO] testing fine tune model')
validationSamples = len(validationLabels)
validationLabels = utils.transformLabelsToCategorical(
validationLabels, self.numClasses)
testDatagen = ImageDataGenerator(rescale=1. / 255)
generator = testDatagen.flow(validationData,
validationLabels,
shuffle=False,
batch_size=self.batchSize)
testLoss, testAccuracy = self.model.topModel.evaluate_generator(
generator, steps=validationSamples / self.batchSize)
self.logger.saveDataFineTuneModel(historyGenerated,
(testLoss, testAccuracy),
trainingTime, fold, numLayersFreeze,
self.model.topType)
import gc
del self.model.topModel
gc.collect()
def testTopModel(self, validationData, validationLabels, trainingTime,
historyGenerated, fold):
print('[INFO] testing top model type ' + (str(self.model.topType)))
validationSamples = len(validationLabels)
validationLabels = utils.transformLabelsToCategorical(
validationLabels, self.numClasses)
testDatagen = ImageDataGenerator(rescale=1. / 255)
generator = testDatagen.flow(validationData,
validationLabels,
shuffle=False,
batch_size=self.batchSize)
testLoss, testAccuracy = self.model.topModel.evaluate_generator(
generator, steps=validationSamples / self.batchSize)
self.logger.saveDataTopModel(historyGenerated,
(testLoss, testAccuracy), trainingTime,
fold, self.model.topType)
def fineTune(self, trainData, trainLabels, validData, validLabels,
numLayersFreeze, bestTopType, pathBestTop, fold):
import math
self.model.topModel = self.logger.loadWeights(self.model.topModel,
pathBestTop)
self.model.topType = modelType.modelType[bestTopType]
fullModel = self.addTopModel()
# Freeze layers
for layer in fullModel.layers[:numLayersFreeze]:
layer.trainable = False
import tensorflow as tf
run_opts = tf.RunOptions(report_tensor_allocations_upon_oom=True)
from keras import optimizers
fullModel.compile(loss='categorical_crossentropy',
optimizer=optimizers.SGD(lr=1e-4, momentum=0.9),
metrics=['accuracy'],
options=run_opts)
trainSamples = len(trainLabels)
validationSamples = len(validLabels)
predictSizeTrain = int(math.ceil(trainSamples / self.batchSize))
predictSizeValidation = int(
math.ceil(validationSamples / self.batchSize))
trainLabels = utils.transformLabelsToCategorical(
trainLabels, self.numClasses)
validLabels = utils.transformLabelsToCategorical(
validLabels, self.numClasses)
print('[INFO] Fine tuning with ' + str(numLayersFreeze) +
' layers freeze.')
# Data augmentation
trainDatagen = ImageDataGenerator(rescale=1. / 255,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True)
generatorTrain = trainDatagen.flow(trainData,
trainLabels,
shuffle=False,
batch_size=self.batchSize)
valDatagen = ImageDataGenerator(rescale=1. / 255)
generatorVal = valDatagen.flow(validData,
validLabels,
shuffle=False,
batch_size=self.batchSize)
import time
t = time.process_time()
historyGenerated = fullModel.fit_generator(
generatorTrain,
steps_per_epoch=predictSizeTrain,
epochs=10,
validation_data=generatorVal,
validation_steps=predictSizeValidation)
trainingTime = time.process_time() - t
self.logger.saveWeightsFineTune(fullModel, fold, numLayersFreeze)
import gc
del fullModel
gc.collect()
return trainingTime, historyGenerated