def train_vgg16(nameOverride=None, layers_to_train=None):
"""
Build, train and save VGG16 model with block4 and block5 fine-tuning.
:return: null
"""
if nameOverride is None:
nameOverride = "block4and5"
# load data
training_sets = load_augmented_dataset()
# build models
model_vgg = build_vgg_trainable(fine_tune=layers_to_train)
baseWeights_t = model_vgg.get_weights()
# NOTE: You can still leave this alone if you've only downloaded the fully augmented set.
for training_set in training_sets:
print(" Starting training for set {}".format(str(training_set)))
model_vgg.set_weights(baseWeights_t) # Resets model
train_x = np.load(
os.path.join("./model_cache/train_data",
training_sets[training_set][0]))
train_y = np.load(
os.path.join("./model_cache/train_data",
training_sets[training_set][1]))
early_stopping_monitor = EarlyStopping(patience=2)
history = model_vgg.fit(train_x,
train_y,
batch_size=32,
epochs=20,
verbose=1,
validation_split=0.2,
shuffle=True,
callbacks=[early_stopping_monitor])
mpu.plot_accuracy_loss(
history, "./model_cache/train_data/{}_{}_vgg16_plots.png".format(
str(training_set), nameOverride))
upload_blob(
BUCKET_NAME,
"./model_cache/train_data/{}_{}_vgg16_plots.png".format(
str(training_set),
nameOverride), "model_charts/{}_{}_vgg16_plots.png".format(
str(training_set), nameOverride))
model_vgg.save("./model_cache/train_data/{}_{}_vgg16.h5".format(
str(training_set), nameOverride))
upload_blob(
BUCKET_NAME, "./model_cache/train_data/{}_{}_vgg16.h5".format(
str(training_set), nameOverride),
"saved_models/{}_{}_vgg16.h5".format(str(training_set),
nameOverride))
def train_mobilenetv2():
"""
Build, train and save MobileNetV2 model with 80% fine-tuning.
:return: null
"""
# load data
training_sets = load_augmented_dataset()
# build models
model_mobile = build_mobilenetv2()
# store base weights
baseWeights_t = model_mobile.get_weights()
# NOTE: You can still leave this alone if you've only downloaded the fully augmented set.
for training_set in training_sets:
print(" Starting training for set {}".format(str(training_set)))
model_mobile.set_weights(baseWeights_t) # Resets model
train_x = np.load(
os.path.join("./model_cache/train_data",
training_sets[training_set][0]))
train_y = np.load(
os.path.join("./model_cache/train_data",
training_sets[training_set][1]))
early_stopping_monitor = EarlyStopping(patience=2)
history = model_mobile.fit(train_x,
train_y,
batch_size=32,
epochs=20,
verbose=1,
validation_split=0.2,
shuffle=True,
callbacks=[early_stopping_monitor])
mpu.plot_accuracy_loss(
history,
"./model_cache/train_data/{}_mobilenetv2_plots.png".format(
str(training_set)))
upload_blob(
BUCKET_NAME,
"./model_cache/train_data/{}_mobilenetv2_plots.png".format(
str(training_set)),
"model_charts/{}_mobilenetv2_plots.png".format(str(training_set)))
model_mobile.save("./model_cache/train_data/{}_mobilenetv2.h5".format(
str(training_set)))
upload_blob(
BUCKET_NAME, "./model_cache/train_data/{}_mobilenetv2.h5".format(
str(training_set)),
"saved_models/{}_mobilenetv2.h5".format(str(training_set)))
def train_efficientnet():
"""
Build, train and save Efficientnet model.
:return: null
"""
# load data
training_sets = load_augmented_dataset()
# build models
model_efn = build_efficientnet()
baseWeights_t = model_efn.get_weights()
for training_set in training_sets:
print(" Starting training for set {}".format(str(training_set)))
model_efn.set_weights(baseWeights_t) # Resets model
train_x = np.load(
os.path.join("./model_cache/train_data",
training_sets[training_set][0]))
train_y = np.load(
os.path.join("./model_cache/train_data",
training_sets[training_set][1]))
early_stopping_monitor = EarlyStopping(patience=5)
history = model_efn.fit(train_x,
train_y,
batch_size=32,
epochs=40,
verbose=1,
validation_split=0.2,
shuffle=True,
callbacks=[early_stopping_monitor])
mpu.plot_accuracy_loss(
history,
"./model_cache/train_data/{}_efficientnet_plots.png".format(
str(training_set)))
upload_blob(
BUCKET_NAME,
"./model_cache/train_data/{}_efficientnet_plots.png".format(
str(training_set)),
"model_charts/{}_efficientnet_plots.png".format(str(training_set)))
model_efn.save("./model_cache/train_data/{}_efficientnet.h5".format(
str(training_set)))
upload_blob(
BUCKET_NAME, "./model_cache/train_data/{}_efficientnet.h5".format(
str(training_set)),
"saved_models/{}_efficientnet.h5".format(str(training_set)))