def main():
print(
"****** Run short training with InceptionV3 and save results. ******")
num_classes = 10
config_dict = {
'base_model': 'InceptionV3',
'num_dense_layers': 3,
'num_dense_units_0': 500,
'num_dense_units_1': 250,
'num_dense_units_2': 50,
'activation': 'relu',
'dropout': True,
'dropout_0': 0.5,
'dropout_1': 0.5,
'dropout_2': 0.5,
'optimizer': "SGD",
'learning_rate': 0.001,
'cnn_unlock_epoch': 8,
'unfreeze_percentage': 0.1,
'batch_size': 16
}
_, _, histories = train(config_dict, epochs=16, num_classes=num_classes)
print("HISTORIES:")
print(histories)
run_name = tools.get_run_name()
tools.save_learning_curves(histories, run_name)
csv_path = os.path.join("plots/", run_name, "data.csv")
ut.write_csv_dict(histories,
keys=['loss', 'acc', 'val_loss', 'val_acc'],
filename=csv_path)
def draw_num_classes_graphs():
"""Train network and save learning curves for different values for num_classes."""
values = [10, 50, 100, 250, 1000, 4000]
for num_classes in values:
print("Training model on {} most common classes.".format(num_classes))
model = create_pretrained_model(num_classes=num_classes)
histories = train(model, num_classes, epochs=50)
run_name = get_run_name("{}classes".format(num_classes))
save_learning_curves(histories, run_name)
csv_path = os.path.join("plots/", run_name, "data.csv")
ut.write_csv_dict(histories,
keys=['loss', 'acc', 'val_loss', 'val_acc'],
filename=csv_path)
def main():
print("Run short training with InceptionV3 and save results.")
num_classes = 10
model = create_pretrained_model(num_classes=num_classes)
histories = train(model, num_classes, epochs=2, cnn_epochs=2)
print("HISTORIES:")
print(histories)
run_name = tools.get_run_name()
tools.save_learning_curves(histories, run_name)
csv_path = os.path.join("plots/", run_name, "data.csv")
ut.write_csv_dict(histories,
keys=['loss', 'acc', 'val_loss', 'val_acc'],
filename=csv_path)
def draw_num_classes_graphs():
print("Will likely not work because")
print("keras_tools.draw_num_classes_graphs() was not yet adapted")
print("to the usage of config_dict in keras_model.py")
"""Train network and save learning curves for different values for num_classes."""
values = [10, 50, 100, 250, 1000, 4000]
for num_classes in values:
print("Training model on {} most common classes.".format(num_classes))
model = create_pretrained_model(num_classes=num_classes)
histories = train(model, num_classes, epochs=50)
run_name = get_run_name("{}classes".format(num_classes))
save_learning_curves(histories, run_name)
csv_path = os.path.join("plots/", run_name, "data.csv")
ut.write_csv_dict(histories,
keys=['loss', 'acc', 'val_loss', 'val_acc'],
filename=csv_path)
def train(config_dict,
epochs,
model=None,
num_classes=10,
save_model_path=None,
save_data_path="plots",
train_dir="data/model_train",
valid_dir="data/model_valid",
train_valid_split=0.7):
start_time = time.time()
#
# extract relevant parts of configuration
#
cnn_unlock_epoch = config_dict["cnn_unlock_epoch"]
unfreeze_percentage = config_dict["unfreeze_percentage"]
batch_size = config_dict['batch_size']
#
# get model to train, determine training times
#
if model is None:
model = _create_pretrained_model(config_dict, num_classes)
if epochs <= cnn_unlock_epoch:
training_epochs_dense = epochs
training_epochs_wholemodel = 0
else:
training_epochs_dense = cnn_unlock_epoch
training_epochs_wholemodel = epochs - cnn_unlock_epoch
if model.name == 'InceptionV3' or model.name == 'Xception' or model.name == 'InceptionResNetV2':
target_size = (299, 299)
elif model.name == 'ResNet50' or model.name == 'MobileNet':
target_size = (224, 224)
else:
print("invalid model: ", model.name)
print("training model", model.name)
#
# prepare training data
#
# create environment on filesystem with new random train/valid split
num_train_imgs, num_valid_imgs = ut.create_small_case(
sel_whales=np.arange(1, num_classes+1),
train_dir=train_dir,
valid_dir=valid_dir,
train_valid=train_valid_split,
sub_dirs=True)
train_gen = image.ImageDataGenerator(
# featurewise_center=True,
# featurewise_std_normalization=True,
rescale=1./255, # redundant with featurewise_center ?
# preprocessing_function=preprocess_input, not used in most examples
# horizontal_flip = True, # no, as individual shapes are looked for
fill_mode="nearest",
zoom_range=0.3,
width_shift_range=0.3,
height_shift_range=0.3,
rotation_range=30)
train_flow = train_gen.flow_from_directory(
train_dir,
# save_to_dir="data/model_train/augmented",
# color_mode="grayscale",
target_size=target_size,
batch_size=batch_size,
class_mode="categorical")
valid_gen = image.ImageDataGenerator(
rescale=1./255,
fill_mode="nearest")
valid_flow = valid_gen.flow_from_directory(
valid_dir,
target_size=target_size,
class_mode="categorical")
#
# train fully connected part
#
hist_dense = model.fit_generator(
train_flow,
steps_per_epoch=num_train_imgs//batch_size,
verbose=2,
validation_data=valid_flow,
validation_steps=num_valid_imgs//batch_size,
epochs=training_epochs_dense)
histories = hist_dense.history
#
# train the whole model with parts of the cnn unlocked (fixed optimizer!)
#
if training_epochs_wholemodel > 0:
model = _unfreeze_cnn_layers(model, config_dict)
hist_wholemodel = model.fit_generator(
train_flow,
steps_per_epoch = num_train_imgs//batch_size,
verbose = 2,
validation_data = valid_flow,
validation_steps = num_valid_imgs//batch_size,
epochs=training_epochs_wholemodel)
# concatenate training history
for key in histories.keys():
if type(histories[key]) == list:
histories[key].extend(hist_wholemodel.history[key])
#
# do final cleanup
#
if save_model_path is not None:
model.save(save_model_path)
if save_data_path is not None:
run_name = tools.get_run_name()
tools.save_learning_curves(histories, run_name, base_path=save_data_path)
csv_path = os.path.join(save_data_path, run_name, run_name + ".csv")
ut.write_csv_dict(histories,
keys=['loss', 'acc', 'val_loss', 'val_acc'],
filename=csv_path)
config_file_path = os.path.join(save_data_path, run_name, "config.txt")
ut.append_to_file("configuration of this run:", config_file_path)
ut.append_to_file(config_dict, config_file_path)
ut.append_to_file("epochs=" + str(epochs), config_file_path)
ut.append_to_file("num_classes=" + str(num_classes), config_file_path)
ut.append_to_file("train_valid_split=" + str(train_valid_split), config_file_path)
hpbandster_loss = 1.0 - histories['val_acc'][-1]
runtime = time.time() - start_time
return (hpbandster_loss, runtime, histories)