def _obtain_model(model_backend_name, violation_class,
nb_of_conv_layers_to_fine_tune):
if model_backend_name == 'VGG16':
model = HRA_VGG16(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
elif model_backend_name == 'VGG19':
model = HRA_VGG19(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
elif model_backend_name == 'ResNet50':
model = HRA_ResNet50(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
elif model_backend_name == 'VGG16_Places365':
model = HRA_VGG16_Places365(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
return model
def single_img_HRA_inference_return_only(img_path, violation_class,
model_backend_name,
nb_of_conv_layers_to_fine_tune):
"""Performs single image inference.
# Arguments
img_path: Path to image file
violation_class: one of `cl` (HRA dataset with 2 classes - [i]'child_labour' and [ii]'no violation')
or `dp` (HRA dataset with 2 classes - [i]'displaced_populations' and [ii]'no violation')
model_backend_name: One of `VGG16`, `VGG19`, `ResNet50` or `VGG16_Places365`.
nb_of_conv_layers_to_fine_tune: integer to indicate the number of convolutional layers to fine-tune.
# Returns
Three integer values corresponding to `valence`, `arousal` and `dominance`.
"""
if model_backend_name == 'VGG16':
model = HRA_VGG16(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
elif model_backend_name == 'VGG19':
model = HRA_VGG19(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
elif model_backend_name == 'ResNet50':
model = HRA_ResNet50(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
elif model_backend_name == 'VGG16_Places365':
model = HRA_VGG16_Places365(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
## Uncomment for extra verbosity
# print('[INFO] HRA-2class model has been loaded')
img = image.load_img(img_path, target_size=(224, 224))
from applications.hra_utils import predict as pd
raw_preds, decoded_preds = pd(violation_class=violation_class,
model=model,
img=img,
target_size=(224, 224))
return raw_preds
# ---------------------------------------------------- #
if violation_class == 'cl':
main_test_dir = '/home/sandbox/Desktop/HRA-2clas-full-test/ChildLabour'
elif violation_class =='dp':
main_test_dir = '/home/sandbox/Desktop/HRA-2clas-full-test/DisplacedPopulations'
if model_backend_name == 'ResNet50':
model = HRA_ResNet50(include_top=True, weights='HRA',
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune,
violation_class=violation_class)
elif model_backend_name == 'VGG16':
model = HRA_VGG16(include_top=True, weights='HRA',
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune,
violation_class=violation_class)
elif model_backend_name == 'VGG19':
model = HRA_VGG19(include_top=True, weights='HRA',
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune,
violation_class=violation_class)
elif model_backend_name == 'VGG16_Places365':
model = HRA_VGG16_Places365(include_top=True, weights='HRA',
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune,
violation_class=violation_class)
base_evaluator = BaseEvaluator(model=model,
model_backend_name=model_backend_name,
help='Number of predictions returned by the model')
parser.add_argument("--to_file",
type=str,
help='name to save the super imposed image to disk')
args = parser.parse_args()
return args
# model = baseline_model(classes=9, epochs=40, weights='HRA')
args = get_args()
if args.pre_trained_model == 'VGG16':
model = HRA_VGG16(weights='HRA',
mode='fine_tuning',
pooling_mode=args.pooling_mode)
layer2visualise = 'block5_conv3'
elif args.pre_trained_model == 'VGG19':
model = HRA_VGG19(weights='HRA',
mode='fine_tuning',
pooling_mode=args.pooling_mode)
layer2visualise = 'block5_conv3'
elif args.pre_trained_model == 'ResNet50':
model = HRA_ResNet50(weights='HRA',
mode='fine_tuning',
pooling_mode=args.pooling_mode)
elif args.pre_trained_model == 'VGG16_Places365':
def single_img_HRA_inference(img_path, violation_class, model_backend_name,
nb_of_conv_layers_to_fine_tune):
"""Performs single image inference.
# Arguments
img_path: Path to image file
violation_class: one of `cl` (HRA dataset with 2 classes - [i]'child_labour' and [ii]'no violation')
or `dp` (HRA dataset with 2 classes - [i]'displaced_populations' and [ii]'no violation')
model_backend_name: One of `VGG16`, `VGG19`, `ResNet50` or `VGG16_Places365`.
nb_of_conv_layers_to_fine_tune: integer to indicate the number of convolutional layers to fine-tune.
# Returns
Three integer values corresponding to `valence`, `arousal` and `dominance`.
"""
(head, tail) = os.path.split(img_path)
filename_only = os.path.splitext(tail)[0]
if model_backend_name == 'VGG16':
model = HRA_VGG16(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
print("[INFO] Loading and preprocessing image...")
x = prepare_input_data(img_path=img_path,
objects_or_places_flag='objects')
elif model_backend_name == 'VGG19':
model = HRA_VGG19(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
print("[INFO] Loading and preprocessing image...")
x = prepare_input_data(img_path=img_path,
objects_or_places_flag='objects')
elif model_backend_name == 'ResNet50':
model = HRA_ResNet50(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
print("[INFO] Loading and preprocessing image...")
x = prepare_input_data(img_path=img_path,
objects_or_places_flag='objects')
elif model_backend_name == 'VGG16_Places365':
model = HRA_VGG16_Places365(
weights='HRA',
violation_class=violation_class,
nb_of_conv_layers_to_fine_tune=nb_of_conv_layers_to_fine_tune)
print("[INFO] Loading and preprocessing image...")
x = prepare_input_data(img_path=img_path,
objects_or_places_flag='places')
# There seems to be relatively high confidence scores with the following method
# preds = model.predict(x)[0]
#
# print ('Raw predictions: ', preds)
#
# top_preds = np.argsort(preds)[::-1][0:2]
#
# print('Sorted predictions: ', top_preds)
#
# if violation_class == 'cl':
# file_name = 'categories_HRA_2classCL.txt'
# elif violation_class == 'dp':
# file_name = 'categories_HRA_2classDP.txt'
#
# classes = list()
# with open(file_name) as class_file:
# for line in class_file:
# classes.append(line.strip().split(' ')[0][3:])
# classes = tuple(classes)
#
#
# print ('\n')
# print('--PREDICTED HRA 2 CLASSES:')
# # output the prediction
# for i in range(0, 2):
# print(classes[top_preds[i]], '->', preds[top_preds[i]])
img = image.load_img(img_path, target_size=(224, 224))
from applications.hra_utils import predict as pd
raw_preds, decoded_preds = pd(violation_class=violation_class,
model=model,
img=img,
target_size=(224, 224))
# print('Raw preds: ', raw_preds)
# print ('Decoded preds: ', decoded_preds)
# print (type(raw_preds))
# print('Raw preds: ', raw_preds[0])
# print(type(raw_preds[0]))
top_1_predicted_probability = decoded_preds[0][2]
# top_1_predicted = np.argmax(preds)
top_1_predicted_label = decoded_preds[0][1]
# print(top_1_predicted_label, '->' , top_1_predicted_probability)
overlayed_text = str(top_1_predicted_label) + ' (' + str(
round(top_1_predicted_probability, 2)) + ')'
return raw_preds, overlayed_text, top_1_predicted_label
# ==== CompoundNet model / Early-fusion========================================================================================================== # model= CompoundNet_VGG16(weights='HRA', mode= 'fine_tuning', fusion_strategy='average', pooling_mode='avg', data_augm_enabled=False) # model.summary() # =============================================================================================================================================== # ==== Object-centric CompoundNet model ========================================================================================================= # model= CompoundNet_VGG16_VGG19(weights='HRA', mode= 'fine_tuning', fusion_strategy='maximum', pooling_mode='max') # model.summary() # =============================================================================================================================================== # ==== Late-fusion ========================================================================================================= # pooling_mode = 'max' model_a = HRA_VGG16(weights='HRA', mode='fine_tuning', pooling_mode='max', include_top=True, data_augm_enabled=False) model_a.summary() model_b = HRA_VGG16_Places365(weights='HRA', mode='fine_tuning', pooling_mode='flatten', include_top=True, data_augm_enabled=False) model_b.summary() # =============================================================================================================================================== model_name='BEST_COVERAGE_late_fusion' metrics = HRA_metrics(main_test_dir ='/home/sandbox/Desktop/Human_Rights_Archive_DB/test_uniform')
def train(self):
"""Loads the selected model & starts the training process.
"""
if self.pre_trained_model == 'vgg16':
print('[INFO] Instantiating HRA-2CLASS-VGG16...')
if self.train_mode == 'feature_extraction':
model = HRA_VGG16(include_top=True,
weights=None,
input_tensor=None,
input_shape=None,
nb_of_conv_layers_to_fine_tune=self.
nb_of_conv_layers_to_fine_tune,
first_phase_trained_weights=None,
verbose=1)
else:
if os.path.isfile(self.first_phase_trained_weights) is False:
raise IOError("No such weights file: `" +
self.first_phase_trained_weights + "`. ")
model = HRA_VGG16(include_top=True,
weights=None,
input_tensor=None,
input_shape=None,
nb_of_conv_layers_to_fine_tune=self.
nb_of_conv_layers_to_fine_tune,
first_phase_trained_weights=self.
first_phase_trained_weights,
verbose=1)
print('[INFO] HRA-2CLASS-VGG16 model loaded')
elif self.pre_trained_model == 'vgg19':
print('[INFO] Instantiating HRA-2CLASS-VGG19...')
if self.train_mode == 'feature_extraction':
model = HRA_VGG19(include_top=True,
weights=None,
input_tensor=None,
input_shape=None,
nb_of_conv_layers_to_fine_tune=self.
nb_of_conv_layers_to_fine_tune,
first_phase_trained_weights=None,
verbose=1)
else:
if os.path.isfile(self.first_phase_trained_weights) is False:
raise IOError("No such weights file: `" +
self.first_phase_trained_weights + "`. ")
model = HRA_VGG19(include_top=True,
weights=None,
input_tensor=None,
input_shape=None,
nb_of_conv_layers_to_fine_tune=self.
nb_of_conv_layers_to_fine_tune,
first_phase_trained_weights=self.
first_phase_trained_weights,
verbose=1)
print('[INFO] HRA-2CLASS-VGG19 model loaded')
elif self.pre_trained_model == 'resnet50':
print('[INFO] Instantiating HRA-2CLASS-ResNet50...')
if self.train_mode == 'feature_extraction':
model = HRA_ResNet50(include_top=True,
weights=None,
input_tensor=None,
input_shape=None,
nb_of_conv_layers_to_fine_tune=self.
nb_of_conv_layers_to_fine_tune,
first_phase_trained_weights=None,
verbose=1)
else:
if os.path.isfile(self.first_phase_trained_weights) is False:
raise IOError("No such weights file: `" +
self.first_phase_trained_weights + "`. ")
model = HRA_ResNet50(include_top=True,
weights=None,
input_tensor=None,
input_shape=None,
nb_of_conv_layers_to_fine_tune=self.
nb_of_conv_layers_to_fine_tune,
first_phase_trained_weights=self.
first_phase_trained_weights,
verbose=1)
print('[INFO] HRA-2CLASS-ResNet50 model loaded')
elif self.pre_trained_model == 'vgg16_places365':
print('[INFO] Instantiating HRA-2CLASS-VGG16_Places365...')
if self.train_mode == 'feature_extraction':
model = HRA_VGG16_Places365(
include_top=True,
weights=None,
input_tensor=None,
input_shape=None,
nb_of_conv_layers_to_fine_tune=self.
nb_of_conv_layers_to_fine_tune,
first_phase_trained_weights=None,
verbose=1)
else:
if os.path.isfile(self.first_phase_trained_weights) is False:
raise IOError("No such weights file: `" +
self.first_phase_trained_weights + "`. ")
model = HRA_VGG16_Places365(
include_top=True,
weights=None,
input_tensor=None,
input_shape=None,
nb_of_conv_layers_to_fine_tune=self.
nb_of_conv_layers_to_fine_tune,
first_phase_trained_weights=self.
first_phase_trained_weights,
verbose=1)
print('[INFO] HRA-2CLASS-VGG16-Places365 model loaded')
# Finally start fitting the dataset
if self.train_mode == 'feature_extraction':
print(
'[INFO] Start training the randomly initialised classifier on top of the pre-trained conv. base...'
)
start_time = time.time()
history = model.fit_generator(self.train_generator,
epochs=self.nb_of_epochs,
steps_per_epoch=self.steps_per_epoch,
validation_data=self.val_generator,
validation_steps=self.val_steps,
callbacks=self.callbacks_list)
end_time = time.time()
print(
"[INFO] It took {} to train the randomly initialised classifier on top of the pre-trained conv. base"
.format(hms_string(end_time - start_time)))
print('[INFO] Saved trained model as: %s ' % self.weights_to_file)
else:
print('[INFO] Start fine-tuning the model...')
start_time = time.time()
history = model.fit_generator(self.train_generator,
epochs=self.nb_of_epochs,
steps_per_epoch=self.steps_per_epoch,
validation_data=self.val_generator,
validation_steps=self.val_steps,
callbacks=self.callbacks_list)
end_time = time.time()
print(
"[INFO] It took {} to fine-tune the top layers of the frozen conv. base"
.format(hms_string(end_time - start_time)))
print('[INFO] Saved trained model as: %s ' % self.weights_to_file)