def load_and_get_model_for_inference(trained_model_arch,
trained_checkpoint_dir, filetype,
input_shape, num_classes):
model_factory = ModelFactory()
model = model_factory.get_model(
trained_model_arch,
input_shape,
is_training=False,
num_classes=num_classes,
learning_rate=0.001) # A dummy learning rate since it is test mode.
# The ModelCheckpoint in train pipeline saves the weights inside the checkpoint directory as follows.
if filetype == '.h5':
weights_path = trained_checkpoint_dir + "best_model_dir-auc.h5"
model = tf.keras.models.load_model(weights_path)
elif filetype == 'tf':
weights_path = os.path.join(trained_checkpoint_dir, "variables",
"variables")
model.load_weights(weights_path)
else:
raise ValueError(
"The provided saved model filetype not recognized: %s" % filetype)
print(
"The model has been created and the weights have been loaded from: %s"
% weights_path)
model.summary()
return model
def __init__(self, config):
self.manager = ModelManager(
config.flavor,
config.server,
config.database,
config.driver,
config.port,
config.schema
)
self.factory = ModelFactory(
config.environment,
config.flavor,
FieldFactory()
)
def createModel():
json_data = request.get_json(force=True)
# check if all fields are there
if json_data.get('model_name') is None:
abort(make_response("model_name field is missing.\n", 422))
if json_data.get('model_type') is None:
abort(make_response("model_type field is missing.\n", 422))
if json_data.get('retrain_counter') is None:
abort(make_response("no retrain information set.\n", 422))
# add model to list of models
app.r.sadd('models', json_data.get('model_name'))
# save model definition
mdl = ModelFactory.createModel(json_data.get('model_type'),
json_data.get('model_name'),
json_data.get('retrain_counter'))
if mdl is None:
return abort(make_response("No model available of type " +
json_data.get('model_type') + "\n",
422))
app.r.set(json_data.get('model_name') + '_object', pickle.dumps(mdl))
return "created model: " + str(mdl) + "\n", 201
def load_model():
model_file_path = 'src/best_weights_1555982768.7076797.h5'
#model_file_path = 'best_weights_1555982768.7076797.h5'
model_factory = ModelFactory()
model = model_factory.get_model(class_names,
model_name=model_type,
use_base_weights=False,
weights_path=model_file_path,
input_shape=(img_height, img_width, 3))
optimizer = keras.optimizers.Adam(lr=1e-3, beta_1=0.9, beta_2=0.999)
model.compile(optimizer=optimizer,
loss="binary_crossentropy",
metrics=["accuracy", "binary_accuracy"])
model.load_weights(model_file_path)
return model
def createModel():
json_data = request.get_json(force=True)
# check if all fields are there
if json_data.get('model_name') is None:
abort(make_response("model_name field is missing.\n", 422))
if json_data.get('model_type') is None:
abort(make_response("model_type field is missing.\n", 422))
if json_data.get('retrain_counter') is None:
abort(make_response("no retrain information set.\n", 422))
# add model to list of models
app.r.sadd('models', json_data.get('model_name'))
# save model definition
mdl = ModelFactory.createModel(json_data.get('model_type'),
json_data.get('model_name'),
json_data.get('retrain_counter'))
if mdl is None:
return abort(make_response("No model available of type " +
json_data.get('model_type') + "\n",
422))
app.r.set(json_data.get('model_name') + '_object', pickle.dumps(mdl))
return "created model: " + str(mdl) + "\n", 201
class Main:
def __init__(self, config):
self.manager = ModelManager(
config.flavor,
config.server,
config.database,
config.driver,
config.port,
config.schema
)
self.factory = ModelFactory(
config.environment,
config.flavor,
FieldFactory()
)
def __call__(self):
# Connect to the server and query model metadata
with self.manager as meta:
tables = meta.tables()
columns = [meta.columns(table=table[3]) for table in tables]
# Zip the serialized values together
zipped = zip(tables, columns)
# Map the Model.__repr__ onto the zipped records.
results = map(
lambda x: str(self.factory.make(x[0][1], x[0][2], x[0][3], fields=x[1])),
zipped
)
# Send the table creation back to the server.
for result in results:
print(result)
def main(dataset_name, model_name, epochs=50, batch_size=128):
dataset = DatasetFactory.get_by_name(dataset_name, train_percentage=0.95)
checkpoint = 'checkpoint/{}_{}'.format(dataset_name, model_name)
solver = Solver(dataset, checkpoint, train_batch_size=batch_size)
model = ModelFactory.get_by_name(model_name, dataset)
solver.train_model(model, epochs=epochs, warmup_epochs=10, num_epoch_to_log=5, learning_rate=1e-3, weight_decay=1e-4)
solver.test(model)
def __init__(self, parent=None):
super(MaeBird, self).__init__(parent)
self.models = ModelFactory()
self.table = None
self.languages = {'perimary': 'Fin', 'secondary': 'Eng',
'tertiary': 'Swe'}
self.dbtype = __DB__
self.dbfile = None
self.db = None
self.matches = []
self.currentsearchitem = 0
self.fullscreen = False
self.setupUi(self)
self.setWindowTitle(__APPNAME__ + ' ' + __VERSION__)
# TODO: loading settings should be moved to a separate method
settings = QSettings()
# Set up logging
loggingdir = settings.value("Logging/loggingDir")
if loggingdir is None:
loggingdir = __USER_DATA_DIR__
self.logger = Logger('root', loggingdir=loggingdir)
if settings.value("Settings/debugging"):
self.logger.debugging = int(settings.value("Settings/debugging"))
self.logger.debug('Logging initialized')
# Try to load previous session
if settings.value("Settings/saveSettings"):
self.saveSettings = int(settings.value("Settings/saveSettings"))
else:
self.saveSettings = 1
if self.saveSettings:
QTimer.singleShot(0, self.load_initial_data)
#QTimer.singleShot(0, self.load_initial_model)
self.header = self.tableView.horizontalHeader()
self.header.sectionDoubleClicked.connect(self.sort_table)
self.search.textEdited.connect(self.update_ui)
self.search.setFocus()
self.searchNextButton.clicked.connect(self.update_ui)
self.searchPrevButton.clicked.connect(self.update_ui)
self.tableView.pressed.connect(self.update_ui)
self.tableView.doubleClicked.connect(
lambda: self.handle_observation(ObservationDialog.SHOW))
self.addButton.clicked.connect(
lambda: self.handle_observation(ObservationDialog.ADD))
self.deleteButton.clicked.connect(
lambda: self.handle_observation(ObservationDialog.DELETE))
def get_parser():
"""
Constructs and returns a parser object to handle the command-line arguments.
:return: argparse.ArgumentParser
"""
# Get parsers for various model architectures.
model_parser = ModelFactory.get_all_parsers()
# Get parsers for various optimizers.
optimizer_parser = OptimizerFactory.get_all_parsers()
# Add parent parsers.
parent_parsers = model_parser + optimizer_parser
parser = argparse.ArgumentParser(parents=parent_parsers)
# Generic options
parser.add_argument('--checkpoint-step', type=int, default=1,
help='Number of epochs between successive checkpoint creations')
parser.add_argument('--config-file', type=str, default=[], nargs='*',
help='File(s) to read the command-line arguments from')
parser.add_argument('--continue', action='store_true',
help='Continue the execution of the last experiment saved into the export directory')
parser.add_argument('--debug', action='store_true', help='Show debug messages')
parser.add_argument('--export-dir', type=str, required=True, help='Export directory')
parser.add_argument('--no-gpu', action='store_true', help='Use CPU')
parser.add_argument("--wandb-directory", type=str, default="../wandb")
parser.add_argument("--disable-wandb", action="store_true", help="No Wandb logging")
# Data options
parser.add_argument('--batch-size', type=int, default=[16], nargs='*', help='Batch size(s)')
parser.add_argument('--dataset', type=str, default=[consts.SIGMORPHON2020], nargs='*',
choices=[consts.SIGMORPHON2020], help='Dataset(s) to train on')
parser.add_argument('--sigmorphon2020-root', type=str, help='Root directory for the SIGMORPHON 2020 dataset')
# Language options
parser.add_argument('--language-families', type=str, nargs='*', default=None,
help='The families of languages to load the data for.'
' If not provided, all available families will be used.')
parser.add_argument('--language-info-file', type=str, default='lang_config.tsv',
help='The language information file.')
parser.add_argument('--languages', type=str, nargs='*', default=None,
help='The languages to load the data for.'
' If not provided, all available languages will be used.')
# Optimizer options
parser.add_argument('--optimizer', type=str, default=[OptimizerFactory.optimizers[0]],
choices=OptimizerFactory.optimizers, nargs='*', help='Optimizer algorithm(s)')
parser.add_argument('--num-epochs', type=int, default=30, help='Number(s) of epochs')
# Model options
parser.add_argument('--model-architecture', type=str, default=[ModelFactory.architectures[0]], nargs='*',
choices=ModelFactory.architectures, help='Model architecture(s)')
# Parallelism Optoions, affect various
parser.add_argument('--loader-threads', type=int, default=0, help='Data loading threads. Default to 0 (load in main)')
parser.add_argument('--use-dataparallel', action='store_true', help='Use torch.nn.DataParallel to wrap the model?')
return parser
def test_factory_creation(self):
model_name = 'lin_reg'
retrain_counter = 10
model_obj = StandardModels.LinearRegression(model_name, retrain_counter)
model_obj2 = ModelFactory.createModel('LinearRegression',
model_name,
retrain_counter)
self.assertEqual(model_obj, model_obj2)
def test_online_linear_regression(self):
model_name = 'onl_lin_reg'
retrain_counter = 1
model_obj = ModelFactory.createModel('OnlineLinearRegression',
model_name, retrain_counter)
self.assertEqual(model_obj.model_name, model_name)
self.assertEqual(model_obj.retrain_counter, retrain_counter)
self.assertEqual(model_obj.model_type, "OnlineLinearRegression")
def test_factory_creation(self):
model_name = 'lin_reg'
retrain_counter = 10
model_obj = StandardModels.LinearRegression(model_name,
retrain_counter)
model_obj2 = ModelFactory.createModel('LinearRegression', model_name,
retrain_counter)
self.assertEqual(model_obj, model_obj2)
def test_online_linear_regression(self):
model_name = 'onl_lin_reg'
retrain_counter = 1
model_obj = ModelFactory.createModel('OnlineLinearRegression',
model_name,
retrain_counter)
self.assertEqual(model_obj.model_name, model_name)
self.assertEqual(model_obj.retrain_counter, retrain_counter)
self.assertEqual(model_obj.model_type, "OnlineLinearRegression")
def __init__(self, dataset_config: Config.Dataset, buffer_size: int = 64):
super(SyntheticDataset).__init__()
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.label = dataset_config.classes[-1]
self.model = ModelFactory.create(
dataset_config.gan_model,
n_classes=len(dataset_config.classes)).to(self.device)
self.buffer_size = buffer_size
self.pointer = 0
self.buffer = self.generate_buffer()
def train(self):
"""Enqueue an event to train person group and report train progress"""
group = ModelFactory.registered_group()
group.cognitive.train()
# Keep querying training status every 1 second
while True:
train_result = group.cognitive.trainingStatus()
# Completed training event will have status succeeded or failed
if (train_result['status'] == 'succeeded' or
train_result['status'] == 'failed'):
break
time.sleep(1)
def __init__(self, api, model_factory=None):
self.api = api
self.model_factory = model_factory or ModelFactory()
self.mockups = {}
for resource_name, resource in self.api._registry.items():
model_class = resource._meta.object_class
self.register(resource)
try:
model_factory[model_class]
except:
self.model_factory.register(model_class)
def identify(self, register_new):
"""Identify a person or a group of people captured by camera"""
registered_group = ModelFactory.registered_group()
faces = self.record(False)
recognized_people = registered_group.identify(faces)
if len(recognized_people) > 0:
for person in recognized_people:
print 'Xin chao %s' % person
Speaker.speak(person)
# If register on premise mode is turn on then register person
# in case we cannot identify who he is
if register_new is True and len(recognized_people) == 0:
self.record(True)
def __init__(self, env: MultiEnv, model_factory: ModelFactory, curiosity_factory: CuriosityFactory,
normalize_state: bool, normalize_reward: bool, reporter: Reporter = NoReporter()) -> None:
self.env = env
self.reporter = reporter
self.state_converter = Converter.for_space(self.env.observation_space)
self.action_converter = Converter.for_space(self.env.action_space)
self.model = model_factory.create(self.state_converter, self.action_converter)
self.curiosity = curiosity_factory.create(self.state_converter, self.action_converter)
self.reward_normalizer = StandardNormalizer() if normalize_reward else NoNormalizer()
self.state_normalizer = self.state_converter.state_normalizer() if normalize_state else NoNormalizer()
self.normalize_state = normalize_state
self.device: torch.device = None
self.dtype: torch.dtype = None
self.numpy_dtype: object = None
def register(self, data_path):
"""Register a group of people specified in data folder and train
Args:
data_path (string): train data folder
"""
if data_path == 'default':
data_path = Utilities.train_data_path()
if Utilities.file_exists(data_path):
group = ModelFactory.registered_group()
group.save()
for alias_name in os.listdir(data_path):
# Ignore gitkeep file and collect data from all folders
if alias_name != '.gitkeep':
logger.log('Registering %s...' % alias_name)
Person.register(group, alias_name)
# After everything is done, call api to train newly created group
self.train()
class MaeBird(QMainWindow, Ui_MainWindow):
def __init__(self, parent=None):
super(MaeBird, self).__init__(parent)
self.models = ModelFactory()
self.table = None
self.languages = {'perimary': 'Fin', 'secondary': 'Eng',
'tertiary': 'Swe'}
self.dbtype = __DB__
self.dbfile = None
self.db = None
self.matches = []
self.currentsearchitem = 0
self.fullscreen = False
self.setupUi(self)
self.setWindowTitle(__APPNAME__ + ' ' + __VERSION__)
# TODO: loading settings should be moved to a separate method
settings = QSettings()
# Set up logging
loggingdir = settings.value("Logging/loggingDir")
if loggingdir is None:
loggingdir = __USER_DATA_DIR__
self.logger = Logger('root', loggingdir=loggingdir)
if settings.value("Settings/debugging"):
self.logger.debugging = int(settings.value("Settings/debugging"))
self.logger.debug('Logging initialized')
# Try to load previous session
if settings.value("Settings/saveSettings"):
self.saveSettings = int(settings.value("Settings/saveSettings"))
else:
self.saveSettings = 1
if self.saveSettings:
QTimer.singleShot(0, self.load_initial_data)
#QTimer.singleShot(0, self.load_initial_model)
self.header = self.tableView.horizontalHeader()
self.header.sectionDoubleClicked.connect(self.sort_table)
self.search.textEdited.connect(self.update_ui)
self.search.setFocus()
self.searchNextButton.clicked.connect(self.update_ui)
self.searchPrevButton.clicked.connect(self.update_ui)
self.tableView.pressed.connect(self.update_ui)
self.tableView.doubleClicked.connect(
lambda: self.handle_observation(ObservationDialog.SHOW))
self.addButton.clicked.connect(
lambda: self.handle_observation(ObservationDialog.ADD))
self.deleteButton.clicked.connect(
lambda: self.handle_observation(ObservationDialog.DELETE))
def closeEvent(self, event):
settings = QSettings()
if self.saveSettings:
db = self.dbfile if self.db is not None else ''
settings.setValue("Database/LastDb", db)
if self.tableView.model() is not None:
settings.setValue("Database/DefaultModel", self.tableView.model().name)
visible_fields = [not bool(self.tableView.isColumnHidden(i)) for i in range(0, self.tableView.model().columnCount())]
settings.setValue("Database/visibleFields", visible_fields)
settings.setValue("Settings/debugging", int(self.logger.debugging))
settings.setValue("Settings/saveSettings", int(self.saveSettings))
def load_initial_data(self):
settings = QSettings()
dbfile = unicode(settings.value("Database/LastDb"))
modelname = unicode(settings.value("Database/DefaultModel"))
if dbfile and QFile.exists(dbfile):
self.load_db(dbfile, modelname=modelname)
self.logger.debug("Loaded database %s with model %s" % (dbfile,
modelname))
if settings.value("Database/visibleFields"):
visible_fields = [item for item in settings.value("Database/visibleFields")]
# FIXME: in absence of QVariant, deal with values
visible_fields = [False if item == 'false' else True for item in visible_fields]
if not all(visible_fields):
self.logger.debug("Hiding fields %s" % visible_fields)
self.show_fields(visible_fields)
def load_db(self, dbname, modelname=None):
self.db = QSqlDatabase.addDatabase(self.dbtype)
self.db.setDatabaseName(dbname)
if not self.db.open():
QMessageBox.warning(self, "Batabase connection",
"Database Error: %s" % (self.db.lastError().text()))
return
self.dbfile = dbname
if modelname not in self.models.model_names:
modeldlg = ModelDialog(self.models.model_names)
if modeldlg.exec_():
modelname = modeldlg.selected_model()
if modelname:
self.load_model(modelname)
def load_model(self, modelname):
''' Loads a specific database model and sets it to view.
'''
try:
model = self.models.get_model(modelname)
except NotImplementedError, e:
QMessageBox.warning(self, "Database model",
"Database Model Error: %s" % str(e))
return
self.tableView.setModel(model(self))
self.tableView.setItemDelegate(QSqlRelationalDelegate(self))
self.tableView.setSelectionMode(QTableView.SingleSelection)
self.tableView.setSelectionBehavior(QTableView.SelectRows)
self.tableView.setColumnHidden(0, True)
self.tableView.resizeColumnsToContents()
self.update_ui()
class Trainer():
def __init__(self, image_shape, io):
self.mf = ModelFactory(image_shape)
self.io = io
def train(self, path1, path2):
print('path to imgs:', path1, " path to arts:", path2)
g1, g2, d1, d2, c1, c2 = self.mf.training_models()
photo_imgs = io.imgs_to_np(path1)
impr_imgs = io.imgs_to_np(path2)
#photo_imgs, impr_imgs = IOHandler().load_npz('real-images.npz')
self._model_train(d1, d2, g1, g2, c1, c2, (photo_imgs, impr_imgs))
def _generate_real_samples(self, dataset, n_samples, patch_shape):
ix = np.random.randint(0, dataset.shape[0], n_samples)
X = dataset[ix]
y = np.ones((n_samples, patch_shape, patch_shape, 1))
return X, y
def _summarize_performance(self, step, g_model, trainX, name, n_samples=5):
# select a sample of input images
X_in, _ = self._generate_real_samples(trainX, n_samples, 0)
# generate translated images
X_out, _ = g_model.generate_fake_samples(X_in, 0)
# scale all pixels from [-1,1] to [0,1]
X_in = (X_in + 1) / 2.0
X_out = (X_out + 1) / 2.0
self.io.save_fig('figs', X_in, X_out)
#image pool abstraction function
def _update_image_pool(self, pool, images, max_size=50):
selected = list()
for image in images:
if len(pool) < max_size:
# stock the pool
pool.append(image)
selected.append(image)
elif np.random.random() < 0.5:
# use image, but don't add it to the pool
selected.append(image)
else:
# replace an existing image and use replaced image
ix = np.random.randint(0, len(pool))
selected.append(pool[ix])
pool[ix] = image
return np.asarray(selected)
# train cyclegan models
def _model_train(self, d_model_A, d_model_B, g_model_AtoB, g_model_BtoA,
c_model_AtoB, c_model_BtoA, dataset):
# define properties of the training run
n_epochs, n_batch, = 25, 1 #* tpu_strategy.num_replicas_in_sync
# determine the output square shape of the discriminator
n_patch = d_model_A.model.output_shape[1]
# unpack dataset
trainA, trainB = dataset
# prepare image pool for fakes
poolA, poolB = list(), list()
# calculate the number of batches per training epoch
bat_per_epo = int(len(trainA) / n_batch)
# calculate the number of training iterations
n_steps = bat_per_epo * n_epochs
print("doing {} steps".format(n_steps))
print(bat_per_epo)
# manually enumerate epochs
for i in range(n_steps):
# select a batch of real samples
X_realA, y_realA = self._generate_real_samples(
trainA, n_batch, n_patch)
X_realB, y_realB = self._generate_real_samples(
trainB, n_batch, n_patch)
# generate a batch of fake samples
X_fakeA, y_fakeA = g_model_BtoA.generate_fake_samples(
X_realB, n_patch)
X_fakeB, y_fakeB = g_model_AtoB.generate_fake_samples(
X_realA, n_patch)
# update fakes from pool
X_fakeA = self._update_image_pool(poolA, X_fakeA)
X_fakeB = self._update_image_pool(poolB, X_fakeB)
# update generator B->A via adversarial and cycle loss
c_model_BtoA.set_trainable(True)
g_loss2, _, _, _, _ = c_model_BtoA.model.train_on_batch(
[X_realB, X_realA], [y_realA, X_realA, X_realB, X_realA])
c_model_BtoA.set_trainable(False)
# update discriminator for A -> [real/fake]
dA_loss1 = d_model_A.model.train_on_batch(X_realA, y_realA)
dA_loss2 = d_model_A.model.train_on_batch(X_fakeA, y_fakeA)
# update generator A->B via adversarial and cycle loss
c_model_AtoB.set_trainable(True)
g_loss1, _, _, _, _ = c_model_AtoB.model.train_on_batch(
[X_realA, X_realB], [y_realB, X_realB, X_realA, X_realB])
c_model_AtoB.set_trainable(False)
# update discriminator for B -> [real/fake]
dB_loss1 = d_model_B.model.train_on_batch(X_realB, y_realB)
dB_loss2 = d_model_B.model.train_on_batch(X_fakeB, y_fakeB)
# summarize performance
s = ('>%d, dA[%.3f,%.3f] dB[%.3f,%.3f] g[%.3f,%.3f]' %
(i + 1, dA_loss1, dA_loss2, dB_loss1, dB_loss2, g_loss1,
g_loss2))
sys.stdout.write(s)
sys.stdout.flush()
# evaluate the model performance every so often
if (i + 1) % int(300) == 0:
# plot A->B translation
self._summarize_performance(i, g_model_AtoB, trainA, 'AtoB')
# plot B->A translation
self._summarize_performance(i, g_model_BtoA, trainB, 'BtoA')
if (i + 1) % (1000) == 0:
# save the models
self.io.save_models('models',
step=i,
models=[
d_model_A, d_model_B, g_model_AtoB,
g_model_BtoA, c_model_AtoB,
c_model_BtoA
])
from keras.models import Model, Sequential, load_model
from keras.utils.vis_utils import plot_model
import pandas as pd
from utils import RgbGenerator
from models import ModelFactory, ModelType
train = pd.read_csv('20bnjester_csv_files/train.csv')
valid = pd.read_csv('20bnjester_csv_files/valid.csv')
model = ModelFactory(rgbpath='trained_models/rgblstm.h5',
trained=True).getModel(ModelType.RGB)
model.summary()
plot_model(model,
to_file='model_plot.png',
show_shapes=True,
show_layer_names=True)
predict_gen = RgbGenerator(valid)
y_pred = model.predict_generator(predict_gen, 1750)
y_pred = np.argmax(y_pred, axis=1)
print(y_pred)
y_true = np.argmax(valid.iloc[:14000, 1:].values, axis=1)
from sklearn.metrics import classification_report
from sklearn.metrics import confusion_matrix
import seaborn as sns
import matplotlib.pyplot as plt
class Engine:
def __init__(self,
args: Args,
cfg: ConfigTree,
local_rank: int,
final_validate=False):
self.args = args
self.cfg = cfg
self.local_rank = local_rank
self.model_factory = ModelFactory(cfg)
self.data_loader_factory = DataLoaderFactoryV3(cfg, final_validate)
self.final_validate = final_validate
self.device = torch.device(
f'cuda:{local_rank}' if torch.cuda.is_available() else 'cpu')
model_type = cfg.get_string('model_type')
if model_type == '1stream':
self.model = self.model_factory.build(local_rank) # basic model
elif model_type == 'multitask':
self.model = self.model_factory.build_multitask_wrapper(local_rank)
else:
raise ValueError(f'Unrecognized model_type "{model_type}"')
if not final_validate:
self.train_loader = self.data_loader_factory.build(
vid=False, # need label to gpu
split='train',
device=self.device)
self.validate_loader = self.data_loader_factory.build(
vid=False, split='val', device=self.device)
if final_validate:
self.n_crop = cfg.get_int(
'temporal_transforms.validate.final_n_crop')
else:
self.n_crop = cfg.get_int('temporal_transforms.validate.n_crop')
self.criterion = nn.CrossEntropyLoss()
self.learning_rate = self.cfg.get_float('optimizer.lr')
optimizer_type = self.cfg.get_string('optimizer.type', default='sgd')
if optimizer_type == 'sgd':
self.optimizer = torch.optim.SGD(
self.model.parameters(),
lr=self.learning_rate,
momentum=self.cfg.get_float('optimizer.momentum'),
dampening=self.cfg.get_float('optimizer.dampening'),
weight_decay=self.cfg.get_float('optimizer.weight_decay'),
nesterov=self.cfg.get_bool('optimizer.nesterov'),
)
elif optimizer_type == 'adam':
self.optimizer = torch.optim.Adam(
self.model.parameters(),
lr=self.learning_rate,
eps=self.cfg.get_float('optimizer.eps'),
)
else:
raise ValueError(f'Unknown optimizer {optimizer_type})')
self.num_epochs = cfg.get_int('num_epochs')
self.schedule_type = self.cfg.get_string('optimizer.schedule')
if self.schedule_type == "plateau":
self.scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer=self.optimizer,
mode='min',
patience=self.cfg.get_int('optimizer.patience'),
verbose=True)
elif self.schedule_type == "multi_step":
self.scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer=self.optimizer,
milestones=self.cfg.get("optimizer.milestones"),
)
elif self.schedule_type == "cosine":
self.scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer=self.optimizer,
T_max=self.num_epochs,
eta_min=self.learning_rate / 1000)
elif self.schedule_type == 'none':
self.scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer=self.optimizer,
lr_lambda=lambda epoch: 1,
)
else:
raise ValueError("Unknow schedule type")
self.arch = cfg.get_string('model.arch')
if local_rank == 0:
self.summary_writer = SummaryWriter(
log_dir=str(args.experiment_dir))
else:
self.summary_writer = None
self.best_acc1 = 0.
self.current_epoch = 0
self.next_epoch = None
logger.info('Engine: n_crop=%d', self.n_crop)
self.checkpoint_manager = CheckpointManager(self.args.experiment_dir,
keep_interval=None)
self.loss_meter = None
def has_next_epoch(self):
return not self.final_validate and self.current_epoch < self.num_epochs - 1
def load_checkpoint(self, checkpoint_path):
states = torch.load(checkpoint_path, map_location=self.device)
if states['arch'] != self.arch:
raise ValueError(
f'Loading checkpoint arch {states["arch"]} does not match current arch {self.arch}'
)
logger.info('Loading checkpoint from %s', checkpoint_path)
self.model.module.load_state_dict(states['model'])
logger.info('Checkpoint loaded')
self.optimizer.load_state_dict(states['optimizer'])
self.scheduler.load_state_dict(states['scheduler'])
self.current_epoch = states['epoch']
self.best_acc1 = states['best_acc1']
def load_moco_checkpoint(self, checkpoint_path: str):
cp = torch.load(checkpoint_path, map_location=self.device)
if 'model' in cp and 'arch' in cp:
logger.info('Loading MoCo checkpoint from %s (epoch %d)',
checkpoint_path, cp['epoch'])
moco_state = cp['model']
prefix = 'encoder_q.'
else:
# This checkpoint is from third-party
logger.info('Loading third-party model from %s', checkpoint_path)
if 'state_dict' in cp:
moco_state = cp['state_dict']
else:
# For c3d
moco_state = cp
logger.warning(
'if you are not using c3d sport1m, maybe you use wrong checkpoint'
)
if next(iter(moco_state.keys())).startswith('module'):
prefix = 'module.'
else:
prefix = ''
"""
fc -> fc. for c3d sport1m. Beacuse fc6 and fc7 is in use.
"""
blacklist = ['fc.', 'linear', 'head', 'new_fc', 'fc8']
blacklist += ['encoder_fuse']
def filter(k):
return k.startswith(prefix) and not any(
k.startswith(f'{prefix}{fc}') for fc in blacklist)
model_state = {
k[len(prefix):]: v
for k, v in moco_state.items() if filter(k)
}
msg = self.model.module.load_state_dict(model_state, strict=False)
# assert set(msg.missing_keys) == {"fc.weight", "fc.bias"} or \
# set(msg.missing_keys) == {"linear.weight", "linear.bias"} or \
# set(msg.missing_keys) == {'head.projection.weight', 'head.projection.bias'} or \
# set(msg.missing_keys) == {'new_fc.weight', 'new_fc.bias'},\
# msg
logger.warning(
f'Missing keys: {msg.missing_keys}, Unexpected keys: {msg.unexpected_keys}'
)
def train_context(self):
return EpochContext(self,
name='Train',
n_crop=1,
dataloader=self.train_loader,
tensorboard_prefix='train')
def validate_context(self):
return EpochContext(self,
name='Validate',
n_crop=self.n_crop,
dataloader=self.validate_loader,
tensorboard_prefix='val')
def train_epoch(self):
epoch = self.next_epoch
if epoch is None:
epoch = self.train_context()
self.next_epoch = self.validate_context()
self.model.train()
with epoch:
for loss, *_ in epoch.forward():
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
self.loss_meter = epoch.loss_meter
def validate_epoch(self):
epoch = self.next_epoch
if epoch is None:
epoch = self.validate_context()
if self.has_next_epoch():
self.next_epoch = self.train_context()
else:
self.next_epoch = None
self.model.eval()
all_logits = torch.empty(0,
device=next(self.model.parameters()).device)
indices = []
with epoch:
with torch.no_grad():
for _, logits, others in epoch.forward():
all_logits = torch.cat((all_logits, logits), dim=0)
if others:
assert len(others[0]) == logits.size(0), \
f'Length of indices and logits not match. {others[0]} vs {logits.size(0)}'
indices.extend(others[0])
epoch.sync_meters()
logger.info(
'Validation finished.\n\tLoss = %f\n\tAcc@1 = %.2f%% (%d/%d)\n\tAcc@5 = %.2f%% (%d/%d)',
epoch.loss_meter.avg.item(),
epoch.top1_meter.avg.item(),
epoch.top1_meter.sum.item() / 100,
epoch.top1_meter.count.item(),
epoch.top5_meter.avg.item(),
epoch.top5_meter.sum.item() / 100,
epoch.top5_meter.count.item(),
)
if self.final_validate:
ds = self.validate_loader.dataset
if hasattr(ds, 'save_results'):
assert indices, 'Dataset should return indices to sort logits'
assert len(indices) == all_logits.size(0), \
f'Length of indices and logits not match. {len(indices)} vs {all_logits.size(0)}'
with (self.args.experiment_dir /
f'results_{self.local_rank}.json').open('w') as f:
ds.save_results(f, indices, all_logits)
return epoch.top1_meter.avg.item()
def run(self):
num_epochs = 1 if self.args.debug else self.num_epochs
self.model.train()
while self.current_epoch < num_epochs:
logger.info("Current LR:{}".format(self.scheduler._last_lr))
if self.summary_writer is not None:
self.summary_writer.add_scalar('train/lr',
utils.get_lr(self.optimizer),
self.current_epoch)
self.train_epoch()
acc1 = self.validate_epoch()
if self.schedule_type == "plateau":
self.scheduler.step(self.loss_meter.val.item())
else:
self.scheduler.step()
self.current_epoch += 1
if self.local_rank == 0:
is_best = acc1 > self.best_acc1
self.best_acc1 = max(acc1, self.best_acc1)
# save_checkpoint({
# 'epoch': self.current_epoch,
# 'arch': self.arch,
# 'model': self.model.module.state_dict(),
# 'best_acc1': self.best_acc1,
# 'optimizer': self.optimizer.state_dict(),
# 'scheduler': self.scheduler.state_dict(),
# }, is_best, self.args.experiment_dir)
self.checkpoint_manager.save(
{
'epoch': self.current_epoch,
'arch': self.arch,
'model': self.model.module.state_dict(),
'best_acc1': self.best_acc1,
'optimizer': self.optimizer.state_dict(),
'scheduler': self.scheduler.state_dict(),
}, is_best, self.current_epoch)
if self.summary_writer is not None:
self.summary_writer.flush()
def __init__(self, image_shape, io):
self.mf = ModelFactory(image_shape)
self.io = io
def __init__(self,
args: Args,
cfg: ConfigTree,
local_rank: int,
final_validate=False):
self.args = args
self.cfg = cfg
self.local_rank = local_rank
self.model_factory = ModelFactory(cfg)
self.data_loader_factory = DataLoaderFactoryV3(cfg, final_validate)
self.final_validate = final_validate
self.device = torch.device(
f'cuda:{local_rank}' if torch.cuda.is_available() else 'cpu')
model_type = cfg.get_string('model_type')
if model_type == '1stream':
self.model = self.model_factory.build(local_rank) # basic model
elif model_type == 'multitask':
self.model = self.model_factory.build_multitask_wrapper(local_rank)
else:
raise ValueError(f'Unrecognized model_type "{model_type}"')
if not final_validate:
self.train_loader = self.data_loader_factory.build(
vid=False, # need label to gpu
split='train',
device=self.device)
self.validate_loader = self.data_loader_factory.build(
vid=False, split='val', device=self.device)
if final_validate:
self.n_crop = cfg.get_int(
'temporal_transforms.validate.final_n_crop')
else:
self.n_crop = cfg.get_int('temporal_transforms.validate.n_crop')
self.criterion = nn.CrossEntropyLoss()
self.learning_rate = self.cfg.get_float('optimizer.lr')
optimizer_type = self.cfg.get_string('optimizer.type', default='sgd')
if optimizer_type == 'sgd':
self.optimizer = torch.optim.SGD(
self.model.parameters(),
lr=self.learning_rate,
momentum=self.cfg.get_float('optimizer.momentum'),
dampening=self.cfg.get_float('optimizer.dampening'),
weight_decay=self.cfg.get_float('optimizer.weight_decay'),
nesterov=self.cfg.get_bool('optimizer.nesterov'),
)
elif optimizer_type == 'adam':
self.optimizer = torch.optim.Adam(
self.model.parameters(),
lr=self.learning_rate,
eps=self.cfg.get_float('optimizer.eps'),
)
else:
raise ValueError(f'Unknown optimizer {optimizer_type})')
self.num_epochs = cfg.get_int('num_epochs')
self.schedule_type = self.cfg.get_string('optimizer.schedule')
if self.schedule_type == "plateau":
self.scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer=self.optimizer,
mode='min',
patience=self.cfg.get_int('optimizer.patience'),
verbose=True)
elif self.schedule_type == "multi_step":
self.scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer=self.optimizer,
milestones=self.cfg.get("optimizer.milestones"),
)
elif self.schedule_type == "cosine":
self.scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer=self.optimizer,
T_max=self.num_epochs,
eta_min=self.learning_rate / 1000)
elif self.schedule_type == 'none':
self.scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer=self.optimizer,
lr_lambda=lambda epoch: 1,
)
else:
raise ValueError("Unknow schedule type")
self.arch = cfg.get_string('model.arch')
if local_rank == 0:
self.summary_writer = SummaryWriter(
log_dir=str(args.experiment_dir))
else:
self.summary_writer = None
self.best_acc1 = 0.
self.current_epoch = 0
self.next_epoch = None
logger.info('Engine: n_crop=%d', self.n_crop)
self.checkpoint_manager = CheckpointManager(self.args.experiment_dir,
keep_interval=None)
self.loss_meter = None
def __init__(self, config, experiment_id, load_from_directory=None):
"""
Initializes an experiment by initializing its variables, loading the corresponding dataset, loading the
language information, and instantiating the model and the optimizer.
:param config: The configurations of this experiment as a dictionary object.
:param experiment_id: A unique identifier for this experiment.
:param load_from_directory: A string indicating the path to a directory. If omitted, a new experiment is
created. Else, an already serialized experiment will be read from the directory that is pointed to by this
argument.
"""
self.config = config
self.id = experiment_id
self.current_epoch = 0
self.best_test_score = float('inf')
self.best_epoch_number = -1
self.loss_function = self.AutoPaddedLoss(
torch.nn.CrossEntropyLoss(ignore_index=Alphabet.unknown_integer))
self.model = None
self.optimizer = None
# Find the corresponding dataset.
assert config[consts.DATASET] in [consts.SIGMORPHON2020]
if config[consts.DATASET] == consts.SIGMORPHON2020:
# Create a data loader factory.
data_loader_factory = SigmorphonData_Factory(
config[consts.SIGMORPHON2020_ROOT],
config[consts.LANGUAGE_INFO_FILE])
# Create a data loader for the training data.
logging.getLogger(
consts.MAIN).info('Creating the training dataset.')
dataloader_kwargs = {
'batch_size': self.config[consts.BATCH_SIZE],
'collate_type': 'unpacked'
}
self.train_loader = data_loader_factory.get_dataset(
type=[consts.TRAIN],
families=self.config[consts.LANGUAGE_FAMILIES],
languages=self.config[consts.LANGUAGES],
dataloader_kwargs=dataloader_kwargs)
# Create a data loader for the testing data.
logging.getLogger(
consts.MAIN).info('Creating the testing dataset.')
self.test_loader = data_loader_factory.get_dataset(
type=[consts.DEV],
families=self.config[consts.LANGUAGE_FAMILIES],
languages=self.config[consts.LANGUAGES],
dataloader_kwargs=dataloader_kwargs)
else:
raise Exception('Unsupported dataset.')
# Instantiate the model indicated by the configurations.
self.model = ModelFactory.create_model(
config[consts.MODEL_ARCHITECTURE], self.config,
self.train_loader.dataset.get_dimensionality())
# Move to the preferred device
self.loss_function = self.loss_function.to(self.config[consts.DEVICE])
self.model = self.model.to(self.config[consts.DEVICE])
if self.config[consts.DATA_PARALLEL]:
self.model = torch.nn.DataParallel(self.model)
# Instantiate the optimizer indicated by the configurations.
self.optimizer = OptimizerFactory.create_optimizer(
config[consts.OPTIMIZER], self.model, self.config)
# If the `load_from_directory` argument is given, load the state of the experiment from a file.
if load_from_directory is not None:
self.deserialize(load_from_directory)
#!/usr/bin/env python
# coding: utf-8
# In[1]:
import cv2
# In[2]:
from models import ModelType, ModelFactory
# In[3]:
model = ModelFactory(rgbpath='trained_models/rgblstm.h5',
trained=True).getModel(ModelType.RGB)
# In[4]:
model.summary()
# In[5]:
from keras.layers import Input
from keras.models import Model
# In[6]:
rgbinput = Input((150, 100, 3))
x = model.layers[1].layer(rgbinput)
for layer in model.layers[2:-3]:
def main():
print("\n###############################################################")
print("##########################DATA PREPARATION#####################")
print("###############################################################\n")
ROOT_DIR = os.getcwd()
print(ROOT_DIR)
INPUT_DIR = os.path.join(ROOT_DIR, config.INPUT_FOLDER)
print(INPUT_DIR)
PATIENTS_INFO = os.path.join(INPUT_DIR, config.INFO_PATIENTS)
print(PATIENTS_INFO)
IMAGES_REGEX = os.path.join(INPUT_DIR, config.IMAGES_ACESS)
images_paths = config_func.getImages(IMAGES_REGEX)
print(images_paths[:5])
data = pd.read_csv(PATIENTS_INFO)
print(data.iloc[0])
data = data.sort_values(config.IMAGE_ID, ascending=True)
print(data.head(5))
#ADD NEW COLUMN (PATH IMAGE) AND POPULATE WITH COHERENT PATH FOR EACH IMAGE
data = config_func.addNewColumn_Populate_DataFrame(data, config.PATH, images_paths)
data = data.sort_index()
print(data.head(5))
print(data.iloc[0][config.PATH])
#IMPUTATE NULL VALUES
data = config_func.impute_null_values(data, config.AGE, mean=True)
print(data.isnull().sum())
print(data.head(5))
data.dx = data.dx.astype('category')
print(data.info())
#GET IMAGE DATASET WITH SPECIFIC SIZE
X, Y = config_func.getDataFromImages(dataframe=data, size=config.WANTED_IMAGES)
print(X.shape)
print(Y.shape)
#number_by_perc = [sum(Y == i) for i in range(len(data.dx.unique()))]
# STRATIFY X_TEST, X_VAL AND X_TEST
indexes = np.arange(X.shape[0])
X_train, X_val, y_train, y_val, indeces_train, indices_val = train_test_split(X, Y, indexes, test_size=config.VALIDATION_SPLIT, shuffle=True,
random_state=config.RANDOM_STATE, stratify=Y)
indexes = indeces_train
X_train, X_test, y_train, y_test, indices_train, indices_test = train_test_split(X_train, y_train, indexes, test_size=config.TEST_SPLIT,
shuffle=True, random_state=config.RANDOM_STATE, stratify=y_train)
print(X_train.shape)
print(y_train.shape)
print(X_val.shape)
print(y_val.shape)
print(X_test.shape)
print(y_test.shape)
if config.FLAG_SEGMENT_IMAGES == 1:
## ---------------------------U-NET APPLICATION ------------------------------------
dataset = Data.Data(X_train=X_train, X_val=X_val, X_test=X_test,
y_train=y_train, y_val=y_val, y_test=y_test)
unet_args = (0, 0) # args doesn't matter --> any tuple is valid here, only in U-Net model
fact = ModelFactory.ModelFactory()
unet = fact.getModel(config.U_NET, dataset, *unet_args) # args doesn't matter
## check save and load predictions array to file
PREDICTIONS_TEMP_FILE_PATH = os.path.join(INPUT_DIR, config.TEMP_ARRAYS)
if os.path.exists(PREDICTIONS_TEMP_FILE_PATH):
with open(PREDICTIONS_TEMP_FILE_PATH, 'rb') as f:
predictions = np.load(f)
else: ## if not exists
with open(PREDICTIONS_TEMP_FILE_PATH, 'wb') as f:
model, predictions, history = unet.template_method()
predictions = np.array(predictions) ## transform list to numpy array
np.save(f, predictions)
## create folder if not exists
masks_path_folder = os.path.join(INPUT_DIR, config.MASKS_FOLDER)
if not os.path.exists(masks_path_folder):
os.makedirs(masks_path_folder)
if not os.listdir(masks_path_folder): ## if folder is empty (no images inside)
## insert mask images in mask folder
for i in range(predictions.shape[0]):
cv2.imwrite(os.path.join(masks_path_folder, data.at[indices_train[i], config.IMAGE_ID]+'.jpg'), predictions[i])
# plt.figure(figsize=(16, 16))
# plt.imshow(cv2.cvtColor(self.data.X_train[2], cv2.COLOR_BGR2RGB))
# plt.title('Original Image')
# plt.show()
# plt.imshow(mask, plt.cm.binary_r)
# plt.title('Binary Mask')
# plt.show()
# plt.imshow(cv2.cvtColor(concatenated_mask, cv2.COLOR_BGR2RGB))
# plt.title('Segmented Image')
# plt.show()
# NORMALIZE DATA
X_train, X_val, X_test = config_func.normalize(X_train, X_val, X_test)
# ONE HOT ENCODING TARGETS
y_train, y_val, y_test = config_func.one_hot_encoding(y_train, y_val, y_test)
print("\n###############################################################")
print("##########################CLASSIFICATION#######################")
print("###############################################################\n")
# CREATION OF DATA OBJECT
data_obj = Data.Data(X_train=X_train, X_val=X_val, X_test=X_test,
y_train=y_train, y_val=y_val, y_test=y_test)
## INSTANCE OF MODEL FACTORY
model_fact = ModelFactory.ModelFactory()
## STRATEGIES OF TRAIN INSTANCES
undersampling = UnderSampling.UnderSampling()
oversampling = OverSampling.OverSampling()
data_augment = DataAugmentation.DataAugmentation()
## ---------------------------ALEXNET APPLICATION ------------------------------------
## DEFINITION OF NUMBER OF CNN AND DENSE LAYERS
args = (6,1)
# CREATE MODEL
alexNet = model_fact.getModel(config.ALEX_NET, data_obj, *args)
# APPLY STRATEGIES OF TRAIN
#alexNet.addStrategy(undersampling)
alexNet.addStrategy(oversampling)
alexNet.addStrategy(data_augment)
# VALUES TO POPULATE ON CONV AND DENSE LAYERS
# definition of args to pass to template_method (conv's number of filters, dense neurons and batch size)
alex_args = (
3, # number of normal convolutional layer (+init conv)
1, # number of stack cnn layers
73, # number of feature maps of initial conv layer
23, # growth rate
1, # number of FCL Layers
65, # number neurons of Full Connected Layer
12# batch size
)
# APPLY BUILD, TRAIN AND PREDICT
#model, predictions, history = alexNet.template_method(*alex_args)
#alexNet.save(model, config.ALEX_NET_WEIGHTS_FILE)
## PLOT FINAL RESULTS
#config_func.print_final_results(data_obj.y_test, predictions, history, dict=False)
## ---------------------------VGGNET APPLICATION ------------------------------------
## DEFINITION OF NUMBER OF CNN AND DENSE LAYERS
vggLayers = (5, 1)
## GET VGGNET MODEL
vggnet = model_fact.getModel(config.VGG_NET, data_obj, *vggLayers)
## ATTRIBUTION OS TRAIN STRATEGIES
vggnet.addStrategy(oversampling)
vggnet.addStrategy(data_augment)
# VALUES TO POPULATE ON CONV AND DENSE LAYERS
vgg_args = (
4, # number of stack cnn layers (+ init stack)
71, # number of feature maps of initial conv layer
18, # growth rate
1, # number of FCL Layers
61, # number neurons of Full Connected Layer
12 # batch size
)
# APPLY BUILD, TRAIN AND PREDICT
#model, predictions, history = vggnet.template_method(*vgg_args)
#vggnet.save(model, config.VGG_NET_WEIGHTS_FILE)
## PLOT FINAL RESULTS
#config_func.print_final_results(data_obj.y_test, predictions, history, dict=False)
## ---------------------------RESNET APPLICATION ------------------------------------
# number of conv and dense layers respectively
number_cnn_dense = (5, 1)
# creation of ResNet instance
resnet = model_fact.getModel(config.RES_NET, data_obj, *number_cnn_dense)
# apply strategies to resnet
resnet.addStrategy(oversampling)
resnet.addStrategy(data_augment)
# definition of args to pass to template_method (conv's number of filters, dense neurons and batch size)
resnet_args = (
56, # number of filters of initial CNN layer
4, # number of consecutive conv+identity blocks
2, # number of identity block in each (conv+identity) block
42, # growth rate
12, # batch size
)
# APPLY BUILD, TRAIN AND PREDICT
#model, predictions, history = resnet.template_method(*resnet_args)
#resnet.save(model, config.RES_NET_WEIGHTS_FILE)
## PLOT FINAL RESULTS
#config_func.print_final_results(data_obj.y_test, predictions, history, dict=False)
## ---------------------------DENSENET APPLICATION ------------------------------------
# # DICTIONARIES DEFINITION
numberLayers = (
4, # BLOCKS
1 # DENSE LAYERS
)
valuesLayers = (
59, # initial number of Feature Maps
4, # number of dense blocks
5, # number of layers in each block
11, # growth rate
1.0, # compression rate
21 # batch size
)
densenet = model_fact.getModel(config.DENSE_NET, data_obj, *numberLayers)
densenet.addStrategy(oversampling)
densenet.addStrategy(data_augment)
#model, predictions, history = densenet.template_method(*valuesLayers)
#densenet.save(model, config.DENSE_NET_WEIGHTS_FILE)
#config_func.print_final_results(data_obj.y_test, predictions, history)
## --------------------------- ENSEMBLE OF MODELS ------------------------------------
# get weights of all methods from files
alexNet2 = load_model(config.ALEX_NET_WEIGHTS_FILE)
vggnet2 = load_model(config.VGG_NET_WEIGHTS_FILE)
#vggnet2.name = 'model_2'
#vggnet.save(vggnet2, config.VGG_NET_WEIGHTS_FILE)
resnet2 = load_model(config.RES_NET_WEIGHTS_FILE)
#resnet2.name = 'model_3'
#resnet.save(resnet2, config.RES_NET_WEIGHTS_FILE)
densenet2 = load_model(config.DENSE_NET_WEIGHTS_FILE)
#densenet2.name = 'model_4'
#densenet.save(densenet2, config.DENSE_NET_WEIGHTS_FILE)
models = [alexNet2, vggnet2, resnet2, densenet2]
##call ensemble method
ensemble_model = config_func.ensemble(models=models)
predictions = ensemble_model.predict(data_obj.X_test)
argmax_preds = np.argmax(predictions, axis=1) # BY ROW, BY EACH SAMPLE
argmax_preds = keras.utils.to_categorical(argmax_preds)
## print final results
config_func.print_final_results(data_obj.y_test, argmax_preds, history=None, dict=True)
# save ensemble model
ensemble_model.save(config.ENSEMBLE_ALL)
del ensemble_model
def _choose_slave(args):
# prob not threadsafe
X, y, X_val, y_val, model_type, m, params = args
model = ModelFactory.create_model(model_type, m, **params)
model.fit(X, y)
return (model.score(X_val, y_val), model, params)
def main():
print("\n###############################################################")
print("##########################DATA PREPARATION#####################")
print("###############################################################\n")
# acess image data
PROJECT_DIR = os.getcwd()
INPUT_DIR = os.path.join(PROJECT_DIR,
config.INPUT_DIR) # path of input directory
IMAGES_DIR = os.path.join(INPUT_DIR, config.IMAGES_ACESS)
# define paths for all classes (stroma, tumor, mucosa, empty, lympho, adipose, complex, debris)
STROMA_FOLDER = os.path.join(IMAGES_DIR, config.STROMA_DIR,
config.IMAGES_REGEX)
TUMOR_FOLDER = os.path.join(IMAGES_DIR, config.TUMOR_DIR,
config.IMAGES_REGEX)
MUCOSA_FOLDER = os.path.join(IMAGES_DIR, config.MUCOSA_DIR,
config.IMAGES_REGEX)
EMPTY_FOLDER = os.path.join(IMAGES_DIR, config.EMPTY_DIR,
config.IMAGES_REGEX)
LYMPHO_FOLDER = os.path.join(IMAGES_DIR, config.LYMPHO_DIR,
config.IMAGES_REGEX)
ADIPOSE_FOLDER = os.path.join(IMAGES_DIR, config.ADIPOSE_DIR,
config.IMAGES_REGEX)
COMPLEX_FOLDER = os.path.join(IMAGES_DIR, config.COMPLEX_DIR,
config.IMAGES_REGEX)
DEBRIS_FOLDER = os.path.join(IMAGES_DIR, config.DEBRIS_DIR,
config.IMAGES_REGEX)
LIST_CLASSES_FOLDER = [
STROMA_FOLDER, TUMOR_FOLDER, MUCOSA_FOLDER, EMPTY_FOLDER,
LYMPHO_FOLDER, ADIPOSE_FOLDER, COMPLEX_FOLDER, DEBRIS_FOLDER
]
# get images from all folders
# classes targets --> 0: Stroma, 1: Tumor, 2: Mucosa, 3: Empty, 4: Lympho, 5: Adipose, 6: Complex, 7: Debris
images = []
labels = []
for i, j in zip(LIST_CLASSES_FOLDER, range(config.NUMBER_CLASSES)):
images.append(config_func.getImages(i))
labels.extend([j for i in range(len(images[j]))])
# flatten images list
images = [path for sublist in images for path in sublist]
# construct DataFrame with two columns: (image_path, target)
data = pd.DataFrame(list(zip(images, labels)),
columns=[config.IMAGE_PATH, config.TARGET])
# subsample data, if not wanted, rate 1 should be passed
if config.SUBSAMPLE_PERCENTAGE != 1:
data = config_func.get_subsample_of_data(1, data)
print(data.head(5))
print(data.shape)
print(data[config.TARGET].value_counts())
# get pixel data from images and respectives targets
X, Y = config_func.resize_images(config.WIDTH, config.HEIGHT, data)
print(X.shape)
print(Y.shape)
# STRATIFY X_TEST, X_VAL AND X_TEST
X_train, X_val, y_train, y_val = train_test_split(
X,
Y,
test_size=config.VALIDATION_SPLIT,
shuffle=True,
random_state=config.RANDOM_STATE,
stratify=Y)
X_train, X_test, y_train, y_test = train_test_split(
X_train,
y_train,
test_size=config.TEST_SPLIT,
shuffle=True,
random_state=config.RANDOM_STATE,
stratify=y_train)
# normalization of data
X_train, X_val, X_test = config_func.normalize(X_train, X_val, X_test)
# one-hot encoding targets
y_train, y_val, y_test = config_func.one_hot_encoding(y_train=y_train,
y_val=y_val,
y_test=y_test)
print("\n###############################################################")
print("##########################CLASSIFICATION#######################")
print("###############################################################\n")
# creation of Data instance
data_obj = Data.Data(X_train=X_train,
X_val=X_val,
X_test=X_test,
y_train=y_train,
y_val=y_val,
y_test=y_test)
# creation of Factory model's instance
model_factory = ModelFactory.ModelFactory()
# creation of Factory optimization algorithms instance
optimization_factory = OptimizerFactory.OptimizerFactory()
# definition of train strategies instances
data_augment = DataAugmentation.DataAugmentation()
## ---------------------------ALEXNET APPLICATION ------------------------------------
# number of conv layers and dense respectively
alex_number_layers = (5, 1)
# creation of AlexNet instance
alexNet = model_factory.getModel(config.ALEX_NET, data_obj,
*alex_number_layers)
# apply strategies to alexNet
alexNet.addStrategy(data_augment)
# definition of args to pass to template_method (conv's number of filters, dense neurons and batch size)
alex_args = (
2, # number of normal convolutional layer (+init conv)
2, # number of stack cnn layers
70, # number of feature maps of initial conv layer
19, # growth rate
1, # number of FCL Layers
43, # number neurons of Full Connected Layer
9 # batch size
)
# apply build, train and predict
#model, predictions, history = alexNet.template_method(*alex_args)
##alexNet.save(model, config.ALEX_NET_WEIGHTS_FILE)
# print final results
#config_func.print_final_results(y_test=data_obj.y_test, predictions=predictions, history=history, dict=False)
## ---------------------------VGGNET APPLICATION ------------------------------------
# number of conv layers and dense respectively
vgg_number_layers = (4, 1)
# creation of VGGNet instance
vggnet = model_factory.getModel(config.VGG_NET, data_obj,
*vgg_number_layers)
# apply strategies to vggnet
vggnet.addStrategy(data_augment)
# definition of args to pass to template_method (conv's number of filters, dense neurons and batch size)
vgg_args = (
4, # number of stack cnn layers (+ init stack)
64, # number of feature maps of initial conv layer
12, # growth rate
1, # number of FCL Layers
16, # number neurons of Full Connected Layer
config.BATCH_SIZE_ALEX_AUG # batch size
)
# apply build, train and predict
#model, predictions, history = vggnet.template_method(*vgg_args)
##vggnet.save(model, config.VGG_NET_WEIGHTS_FILE)
# print final results
#config_func.print_final_results(y_test=data_obj.y_test, predictions=predictions, history=history, dict=False)
## ---------------------------RESNET APPLICATION ------------------------------------
# number of conv and dense layers respectively
number_cnn_dense = (5, 1)
# creation of ResNet instance
resnet = model_factory.getModel(config.RES_NET, data_obj,
*number_cnn_dense)
# apply strategies to resnet
resnet.addStrategy(data_augment)
# definition of args to pass to template_method (conv's number of filters, dense neurons and batch size)
resnet_args = (
48, # number of filters of initial CNN layer
4, # number of consecutive conv+identity blocks
0, # repetition of identity block's, by default resnet-18 is 1 (1conv block + 1 identity block) for all layers
8, # growth rate
config.BATCH_SIZE_ALEX_AUG, # batch size
)
# apply build, train and predict
#model, predictions, history = resnet.template_method(*resnet_args)
##resnet.save(model, config.RES_NET_WEIGHTS_FILE)
# print final results
#config_func.print_final_results(y_test=data_obj.y_test, predictions=predictions, history=history, dict=False)
## ---------------------------DENSENET APPLICATION ------------------------------------
# # DICTIONARIES DEFINITION
numberLayers = (
4, #BLOCKS
1 #DENSE LAYERS
)
valuesLayers = (
24, # initial number of Feature Maps
4, # number of dense blocks
5, # number of layers in each block
12, # growth rate
0.5, # compression rate
config.BATCH_SIZE_ALEX_AUG # batch size
)
densenet = model_factory.getModel(config.DENSE_NET, data_obj,
*numberLayers)
densenet.addStrategy(data_augment)
model, predictions, history = densenet.template_method(*valuesLayers)
config_func.print_final_results(data_obj.y_test, predictions, history)
def main(
config_path: str = 'configs/classification.yaml',
dataset_name: str = 'svhn',
imbalance_ratio: int = 1,
oversampling: str = 'none', # none, oversampling, gan
ada: bool = False, # only for gan training
seed: int = 1, # No seed if 0
wandb_logs: bool = False,
test: bool = False,
load_model: bool = False):
# Ensure output directory exists
if not os.path.exists(OUTPUT_PATH):
os.mkdir(OUTPUT_PATH)
# Set a seed
if seed:
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
# Load configuration
logger.info(f'Loading config at "{config_path}"...')
config = load_config(config_path, dataset_name, imbalance_ratio,
oversampling, ada, load_model)
if config.trainer.task == 'generation' and test:
raise ValueError('Cannot test the generation models')
# Init logging with WandB
mode = 'offline' if wandb_logs else 'disabled'
wandb.init(mode=mode,
dir=OUTPUT_PATH,
entity=WANDB_TEAM,
project=PROJECT_NAME,
group=config.trainer.task,
config=dataclasses.asdict(config))
# Load model
logger.info('Loading model...')
model = ModelFactory.create(model_config=config.model,
n_classes=len(config.dataset.classes))
# Load dataset
logger.info('Loading dataset...')
train_dataset, valid_dataset, test_dataset = DatasetFactory.create(
dataset_config=config.dataset)
# Instatiate trainer
logger.info('Loading trainer...')
trainer = TrainerFactory.create(trainer_config=config.trainer,
train_dataset=train_dataset,
valid_dataset=valid_dataset,
model=model,
classes=config.dataset.classes)
if test:
logger.info('Testing...')
trainer.test(test_dataset)
else:
logger.info('Training...')
trainer.train()
# Cleanup
wandb.finish()
logger.info('done :)')
def train(train_metadata_file_path,
val_metadata_file_path,
images_dir_path,
out_dir,
model_arch,
num_classes,
label_name=None,
sequence_image_count=1,
data_pipeline_mode="mode_flat_all",
class_weight=None,
whole_epochs=100,
batch_size=32,
learning_rate=0.001,
patience=2,
min_delta_auc=0.01,
input_size=(224, 224, 3)):
"""
Train a VGG16 model based on single image.
:param train_metadata_file_path: The path to the metadata '.csv' file containing training image names.
:param val_metadata_file_path: The path to the metadata '.csv' file containing validation image names.
:param images_dir_path: The path containing the images.
:param out_dir: The path to which the saved models need to be written.
:param model_arch: The model architecture provided as string, which are present in the 'models' module.
:param num_classes: The number of classes present in the data. If num_classes=1, it requires the 'label_name'.
:param label_name: Required if num_classes=1. The name of the label to pick from the data.
:param sequence_image_count: The number of images in the sequence dataset. Default: 1.
:param data_pipeline_mode: The mode of the data pipeline. Default: "mode_flat_all".
:param class_weight: The class_weights for imbalanced data. Example: {0: 1.0, 1: 0.5}, if class "0" is twice less
represented than class "1" in your data. Default: None.
:param whole_epochs: The maximum number of epochs to be trained. Note that the model maybe early-stopped. Default: 100.
:param batch_size: The batch size used for the data. Ensure that it fits within the GPU memory. Default: 32.
:param learning_rate: The constant learning rate to be used for the Adam optimizer. Default: 0.001.
:param patience: The number of epochs (full train dataset) to wait before early stopping. Default: 2.
:param min_delta_auc: The minimum delta of validation auc for early stopping after patience. Default: 0.01.
:param input_size: The shape of the tensors returned by the data pipeline mode. Default: (224, 224, 3).
"""
if num_classes == 1 and label_name is None:
raise ValueError(
"Since num_classes equals 1, the label_name must be provided.")
train_data_epoch_subdivisions = 4
early_stop_monitor = "val_auc"
early_stop_min_delta = min_delta_auc
early_stop_patience = patience * train_data_epoch_subdivisions # One run through the train dataset.
prefetch_buffer_size = 3 # Can be also be set to tf.data.experimental.AUTOTUNE
os.makedirs(out_dir)
# Build model architecture.
model_factory = ModelFactory()
model = model_factory.get_model(model_arch,
input_size,
is_training=True,
num_classes=num_classes,
learning_rate=learning_rate)
print("Created the model architecture: %s" % model.name)
model.summary()
# Prepare the training dataset.
print("Preparing training and validation datasets.")
train_data_pipeline = PipelineGenerator(
train_metadata_file_path,
images_dir_path, # XXX: This function calls requires this path to end with slash.
# This needs to be handled in the PipelineGenerator.
is_training=True,
sequence_image_count=sequence_image_count,
label_name=label_name,
mode=data_pipeline_mode)
train_dataset = train_data_pipeline.get_pipeline()
train_dataset = train_dataset.batch(batch_size).prefetch(
prefetch_buffer_size)
# Prepare the validation dataset
val_data_pipeline = PipelineGenerator(
val_metadata_file_path,
images_dir_path,
is_training=False,
sequence_image_count=sequence_image_count,
label_name=label_name,
mode=data_pipeline_mode)
val_dataset = val_data_pipeline.get_pipeline()
val_dataset = val_dataset.batch(batch_size).prefetch(prefetch_buffer_size)
# TODO: Find a way to log the activation maps, either during training, or after the training has completed.
# Prepare the callbacks.
print("Preparing Tensorflow Keras Callbacks.")
earlystop_callback = keras.callbacks.EarlyStopping(
monitor=early_stop_monitor,
min_delta=early_stop_min_delta,
patience=early_stop_patience)
# XXX: We use the HDF5 method to store the sequence models due to a bug in tensorflow TimeDistributed wrapper
if data_pipeline_mode in PipelineGenerator.TIMESTEP_MODES:
model_extension = ".h5"
else:
model_extension = ".ckpt"
best_model_checkpoint_auc_callback = keras.callbacks.ModelCheckpoint(
filepath=os.path.join(out_dir, "best_model_dir-auc" + model_extension),
mode='max',
monitor='val_auc',
save_best_only=True,
save_weights_only=False,
verbose=1)
best_model_checkpoint_loss_callback = keras.callbacks.ModelCheckpoint(
filepath=os.path.join(out_dir,
"best_model_dir-loss" + model_extension),
mode='min',
monitor='val_loss',
save_best_only=True,
save_weights_only=False,
verbose=1)
tensorboard_callback = keras.callbacks.TensorBoard(log_dir=os.path.join(
out_dir, "TBGraph"),
write_graph=True,
write_images=True)
callbacks = [
earlystop_callback, best_model_checkpoint_auc_callback,
best_model_checkpoint_loss_callback, tensorboard_callback
]
# Start model training.
# Defining an 'epoch' to be a quarter of the train dataset.
num_train_samples = train_data_pipeline.get_size()
num_val_samples = val_data_pipeline.get_size()
# Number of batches per one run through the train dataset.
num_training_steps_per_whole_dataset = int(num_train_samples / batch_size)
num_val_steps_per_whole_dataset = int(num_val_samples / batch_size)
steps_per_epoch = int(num_training_steps_per_whole_dataset /
train_data_epoch_subdivisions)
max_num_epochs = int(whole_epochs * train_data_epoch_subdivisions)
max_train_steps = int(max_num_epochs * steps_per_epoch)
print(
"Number of train samples: %s, which correspond to ~%s batches for one complete run through the "
"train dataset. Number of validation samples: %s, which correspond to ~%s batches for complete iteration. "
"Considering a 1/%s fraction of the train dataset as an epoch (steps_per_epoch: %s) "
"after which validation and model checkpoints are saved. Running training for a maximum of %s steps, "
"which correspond to max_num_epochs: %s (whole_epochs: %s). "
"Early stopping has been set based on '%s' of min_delta of %s with a patience of %s."
% (num_train_samples, num_training_steps_per_whole_dataset,
num_val_samples, num_val_steps_per_whole_dataset,
train_data_epoch_subdivisions, steps_per_epoch, max_train_steps,
max_num_epochs, whole_epochs, early_stop_monitor,
early_stop_min_delta, early_stop_patience))
print("\nStarting the model training.")
start_time = time.time()
model.fit(train_dataset,
epochs=max_num_epochs,
steps_per_epoch=steps_per_epoch,
validation_data=val_dataset,
validation_steps=num_val_steps_per_whole_dataset,
callbacks=callbacks,
class_weight=class_weight)
time_taken = time.time() - start_time
print(
"Training completed and the output has been saved in %s. Time taken: %s seconds."
% (out_dir, time_taken))
pin_memory=True)
dataloader_valid = DataLoader(dset_valid,
batch_size=BATCH_SIZE,
sampler=sampler_valid,
num_workers=N_WORKERS,
pin_memory=True)
dataloader_test = DataLoader(dset_test,
batch_size=BATCH_SIZE,
sampler=sampler_test,
num_workers=N_WORKERS,
pin_memory=True)
# Create model
model = ModelFactory.create(config)
model.to(device, dtype=dtype)
lr = float(LEARNING_RATE)
optimizer = model.get_optimizer(model, lr)
lr_scheduler = model.get_lr_scheduler(optimizer)
if DPATH_LOAD_CKPT:
if not fpath_load_ckpt:
fpath_load_ckpt = get_ckpt(DPATH_LOAD_CKPT, LOAD_POLICY) #get_best_ckpt_with_criterion(dpath_load_ckpt, LOAD_POLICY)
load_model(fpath_load_ckpt, model)
print("[%s]"%(LOAD_POLICY.upper()), fpath_load_ckpt, "has been loaded...")
# end of if
model = nn.DataParallel(model)
loss_ce = nn.CrossEntropyLoss()
def classification_loss(logits, target_labels):
def test_with_cover_stego_biased_proportions(prob_cover=0.99, prob_stego=0.01):
print("[Testing with prob_stego: %.3f" % (prob_stego))
# Create dataset and dataloader
dset_test = TimitTestSet(dpath_cover,
dpath_stego,
seed=SEED,
prob_cover=prob_cover,
prob_stego=prob_stego,
dtype=np.float32)
n_data = len(dset_test)
n_train = math.floor(0.8 * n_data)
ix_end_valid = n_train
indices = np.arange(n_data)
sampler_test = SubsetRandomSampler(indices[ix_end_valid:])
# Create dataloader_train
dataloader_test = DataLoader(dset_test,
batch_size=BATCH_SIZE,
sampler=sampler_test,
num_workers=N_WORKERS,
pin_memory=True)
# Create model
model = ModelFactory.create(config)
model.to(device, dtype=dtype)
model = nn.DataParallel(model)
global fpath_load_ckpt
if dpath_load_ckpt:
if not fpath_load_ckpt:
fpath_load_ckpt = get_ckpt(dpath_load_ckpt, policy=LOAD_POLICY)
load_model(fpath_load_ckpt, model)
print("[%s]" % LOAD_POLICY.upper(), fpath_load_ckpt,
"has been loaded...")
elif fpath_load_ckpt:
load_model(fpath_load_ckpt, model)
loss_ce = nn.CrossEntropyLoss()
def classification_loss(logits, target_labels):
return loss_ce(logits, target_labels)
def classify(model, batch):
audios, labels = batch
audios = audios.to(device, non_blocking=True)
labels = labels.to(device, non_blocking=True)
logits = model(audios)
loss = classification_loss(logits, labels)
acc = accuracy_top1(logits, labels)
ps = labels.to(torch.float32).mean().item()
return (logits, loss.item(), acc, ps)
def compute_rates(labels_pred, labels_true):
# Calculate confusion matrix
cm = confusion_matrix(labels_pred, labels_true, labels=(1, 0))
tp = cm[0, 0] # True stegos (stegos)
tn = cm[1, 1] # True covers (covers)
fp = cm[0, 1] # False stegos (covers)
fn = cm[1, 0] # False covers (stegos)
p = tp + fn
n = tn + fp
tpr = tp / p # Sensitivity
fpr = fp / n # False alarm (1 - specificity)
tnr = tn / n # Specificity
fnr = fn / p # Miss rate
return tpr, fpr, tnr, fnr
# Lists for statistics
list_stats = []
list_acc = []
list_loss = []
list_prob = []
# Lists for true labels, scores, predictions
list_scores = []
list_labels = []
num_audios = 0
model.eval()
for epoch in tqdm.tqdm(range(N_REPEATS)):
# Testing model
sum_acc = 0
sum_loss = 0
sum_prob_stego = 0
list_single_test_preds = []
list_single_test_labels = []
for step, batch in enumerate(dataloader_test):
num_audios += 2 * len(batch)
with torch.no_grad():
# ps denotes prob. of fetching stegos
logits, loss, acc, ps = classify(model, batch)
sum_acc += acc
sum_loss += loss
sum_prob_stego += ps
# Compute score for roc_curve
sm = torch.softmax(logits, dim=0)
list_scores.append(sm[:, 1].cpu().numpy())
_, labels = batch
list_single_test_labels.append(labels.cpu().numpy())
preds = logits.topk(1).indices.view(-1) # Predictions
list_single_test_preds.append(preds.cpu().numpy())
# end of for
avg_acc = sum_acc / len(dataloader_test)
avg_loss = sum_loss / len(dataloader_test)
avg_prob_stego = sum_prob_stego / len(dataloader_test)
# Compute the rates
labels_pred = np.concatenate(list_single_test_preds)
labels_true = np.concatenate(list_single_test_labels)
tpr, fpr, tnr, fnr = compute_rates(labels_pred, labels_true)
fstr = "- Acc:%.4f, Loss:%.6f, Ps:%.4f, " \
"FA(fpr):%.4f, MD(fnr):%.4f, PE:%.4f"
print()
print(fstr % (avg_acc, avg_loss, avg_prob_stego, fpr, fnr, 0.5 *
(fpr + fnr)))
# end of for
list_acc.append(avg_acc)
list_loss.append(avg_loss)
list_prob.append(avg_prob_stego)
list_labels.append(labels_true)
list_stats.append({
"test_avg_acc": avg_acc,
"test_avg_loss": avg_loss,
"test_avg_prob_stego": avg_prob_stego,
"test_avg_prob_cover": 1 - avg_prob_stego,
"test_tpr": tpr,
"test_fpr": fpr,
"test_tnr": tnr,
"test_fnr": fnr,
})
# end of for
# Compute ROC
labels_true = np.concatenate(list_labels)
y_score = np.concatenate(list_scores)
roc_fpr, roc_tpr, roc_thr = roc_curve(labels_true, y_score)
roc_auc = roc_auc_score(labels_true, y_score)
print()
print("- Avg. acc:", "%.4f ± %.4f" % (np.mean(list_acc), np.std(list_acc)))
print("- Avg. loss:",
"%.6f ± %.4f" % (np.mean(list_loss), np.std(list_loss)))
print("- Avg. prob:",
"%.4f ± %.4f" % (np.mean(list_prob), np.std(list_prob)))
print("- Total num. tested audios:", num_audios)
print()
df_stats = pd.DataFrame(list_stats)
dict_stats = {
"model": MODEL,
"steganography": STEGANOGRAPHY.lower(),
"ps": ps,
"stats": df_stats,
"roc": {
"roc_auc": roc_auc,
"roc_tpr": roc_tpr,
"roc_fpr": roc_fpr,
"roc_thr": roc_thr
}
}
return dict_stats
def main():
'''''' '''''' '''''' '''''' '''''' '''''' '''''' '''''' '''''' '''''
' DATA PREPARATION (PRE-PROCESSING, CLEAN, TRANSFORM) '
''' '''''' '''''' '''''' '''''' '''''' '''''' '''''' '''''' '''''' ''
print("#################", "DATA PREPARATION", "####################\n")
# CREATION OF DATAFRAME WITH ALL IMAGES --> [ID_PATIENT, PATH_IMAGE, TARGET]
data = pd.DataFrame(index=np.arange(0, config.SIZE_DATAFRAME),
columns=[config.ID, config.IMAGE_PATH, config.TARGET])
# POPULATE DATAFRAME
data = config_func.populate_DataFrame(data)
#TRANSFORM DATA INTO NUMPY ARRAY'S
X, Y = config_func.resize_images(config.WIDTH, config.HEIGHT, data)
#DIVISION OF DATASET'S BETWEEN TRAIN, VALIDATION AND TEST --> I NEED ATTENTION, BECAUSE CLASSES ARE UNBALANCED
indexes = np.arange(X.shape[0])
X_train, X_val, y_train, y_val, indeces_train, indices_val = train_test_split(
X,
Y,
indexes,
test_size=config.VALIDATION_SIZE,
stratify=Y,
shuffle=True,
random_state=config.RANDOM_STATE
) #RANDOM STATE IS NEEDED TO GUARANTEES REPRODUCIBILITY
indexes = indeces_train
X_train, X_test, y_train, y_test, indices_train, indices_test = train_test_split(
X_train,
y_train,
indexes,
test_size=config.TEST_SIZE,
stratify=y_train,
shuffle=True,
random_state=config.RANDOM_STATE)
print(X_train.shape)
print(X_val.shape)
print(X_test.shape)
#NORMALIZE DATA
X_train, X_val, X_test = config_func.normalize(X_train, X_val, X_test)
#ONE HOT ENCODING TARGETS
y_train, y_val, y_test = config_func.one_hot_encoding(
y_train, y_val, y_test)
print("#################", "DATA PREPARATION CONCLUDED",
"####################\n")
#CREATE OBJECT DATA
d = Data.Data(X_train=X_train,
X_val=X_val,
X_test=X_test,
y_train=y_train,
y_val=y_val,
y_test=y_test)
factoryModel = ModelFactory.ModelFactory()
numberLayers = (
4, #CNN LAYERS
1 #DENSE LAYERS
)
## STRATEGIES OF TRAIN INSTANCES
underSampling = UnderSampling.UnderSampling()
data_aug = DataAugmentation.DataAugmentation()
## ---------------------------ALEXNET APPLICATION ------------------------------------
## DICTIONARIES DEFINITION
numberLayers = (
4, #CNN LAYERS
1 #DENSE LAYERS
)
valuesLayers = (
2, ## number of normal convolutional layers
2, ## number of stacked cnn layers
16, ## number of feature maps of first conv layer
16, ## growth rate
2, ## number of FCL's preceding output layer (sigmoid layer)
16, ## number of neurons of Full Connected Layer
config.BATCH_SIZE_ALEX_AUG #batch size
)
# CREATION OF MODEL
alexNetModel = factoryModel.getModel(config.ALEX_NET, d, *numberLayers)
## APPLY STRATEGIES OF TRAIN
alexNetModel.addStrategy(underSampling)
alexNetModel.addStrategy(data_aug)
#model, predictions, history = alexNetModel.template_method(*valuesLayers)
#config_func.print_final_results(d.y_test, predictions, history)
## ---------------------------VGGNET APPLICATION ------------------------------------
## DICTIONARIES DEFINITION
numberLayers = (
4, #CNN LAYERS
1 #DENSE LAYERS
)
valuesLayers = (
5, # conv stacks
24, # number of feature maps of initial convolution layer
16, # growth rate
1, ## number of FCL's preceding output layer (sigmoid layer)
16, # number neurons of Full Connected Layer
config.BATCH_SIZE_ALEX_AUG # batch size
)
vggNetModel = factoryModel.getModel(config.VGG_NET, d, *numberLayers)
vggNetModel.addStrategy(underSampling)
vggNetModel.addStrategy(data_aug)
#model, predictions, history = vggNetModel.template_method(*valuesLayers)
#config_func.print_final_results(d.y_test, predictions, history)
## ---------------------------RESNET APPLICATION ------------------------------------
# number of conv and dense layers respectively
number_cnn_dense = (5, 1)
# creation of ResNet instance
resnet = factoryModel.getModel(config.RES_NET, d, *number_cnn_dense)
# apply strategies to resnet
resnet.addStrategy(underSampling)
resnet.addStrategy(data_aug)
# definition of args to pass to template_method (conv's number of filters, dense neurons and batch size)
resnet_args = (
48, # number of filters of initial CNN layer
4, # number of consecutive conv+identity blocks
1, # repetition of identity block's, by default resnet-18 is 1 (1conv block + 1 identity block) for all layers
8, # growth rate
config.BATCH_SIZE_ALEX_AUG, # batch size
)
# apply build, train and predict
#model, predictions, history = resnet.template_method(*resnet_args)
##resnet.save(model, config.RES_NET_WEIGHTS_FILE)
# print final results
#config_func.print_final_results(y_test=d.y_test, predictions=predictions, history=history, dict=False)
## ---------------------------DENSENET APPLICATION ------------------------------------
# # DICTIONARIES DEFINITION
numberLayers = (
4, #BLOCKS
1 #DENSE LAYERS
)
valuesLayers = (
24, # initial number of Feature Maps
5, # number of dense blocks
2, # number of layers in each block
12, # growth rate
0.5, # compression rate
config.BATCH_SIZE_ALEX_AUG # batch size
)
densenet = factoryModel.getModel(config.DENSE_NET, d, *numberLayers)
densenet.addStrategy(underSampling)
densenet.addStrategy(data_aug)
#model, predictions, history = densenet.template_method(*valuesLayers)
#config_func.print_final_results(d.y_test, predictions, history)
## ------------------------PSO OPTIMIZATION ------------------------------------------
#PSO OPTIMIZATION
optFact = OptimizerFactory.OptimizerFactory()
# definition optimizers for models
pso_alex = optFact.createOptimizer(config.PSO_OPTIMIZER, alexNetModel,
*config.pso_init_args_alex)
pso_vgg = optFact.createOptimizer(config.PSO_OPTIMIZER, vggNetModel,
*config.pso_init_args_vgg)
pso_res = optFact.createOptimizer(config.PSO_OPTIMIZER, resnet,
*config.pso_init_args_resnet)
pso_dense = optFact.createOptimizer(config.PSO_OPTIMIZER, densenet,
*config.pso_init_args_densenet)
# call optimize function
cost, pos, optimizer = pso_alex.optimize()
#plot cost history and plot position history
print("Custo: {}".format(cost))
config_func.print_Best_Position_PSO(pos, config.ALEX_NET) # print position
pso_alex.plotCostHistory(optimizer=optimizer)
pso_alex.plotPositionHistory(optimizer, np.array(config.X_LIMITS),
np.array(config.Y_LIMITS), config.POS_VAR_EXP,
config.LABEL_X_AXIS, config.LABEL_Y_AXIS)