def classifier(self):
"""
Returns a classifier object, which is created on demand.
This means if the ClipClassifier is copied to a new process a new Classifier instance will be created.
"""
if globs._classifier is None:
t0 = datetime.now()
logging.info("classifier loading")
globs._classifier = Model(
tools.get_session(disable_gpu=not self.config.use_gpu))
globs._classifier.load(self.config.classify.model)
logging.info("classifier loaded ({})".format(datetime.now() - t0))
return globs._classifier
def get_full_classifier(config):
from ml_tools.model import Model
"""
Returns a classifier object, which is created on demand.
This means if the ClipClassifier is copied to a new process a new Classifier instance will be created.
"""
t0 = datetime.now()
logging.info("classifier loading")
classifier = Model(
train_config=config.train,
session=tools.get_session(disable_gpu=not config.use_gpu),
)
classifier.load(config.classify.model)
logging.info("classifier loaded ({})".format(datetime.now() - t0))
return classifier
def classifier(self):
"""
Returns a classifier object, which is created on demand.
This means if the ClipClassifier is copied to a new process a new Classifier instance will be created.
"""
if globs._classifier is None:
t0 = datetime.now()
logging.info("classifier loading")
if self.kerasmodel:
model = KerasModel(self.config.train)
model.load_weights(self.model_file)
globs._classifier = model
else:
globs._classifier = Model(
train_config=self.config.train,
session=tools.get_session(
disable_gpu=not self.config.use_gpu),
)
globs._classifier.load(self.model_file)
logging.info("classifier loaded ({})".format(datetime.now() - t0))
return globs._classifier
def __init__(self,
train_config=None,
session=None,
training=False,
tflite=False):
self.tflite = tflite
self.training = training
self.use_gru = train_config.use_gru
self.name = self.model_name()
self.session = session or tools.get_session()
self.saver = None
tf.compat.v1.disable_eager_execution()
# datasets
self.datasets = namedtuple("Datasets", "train, validation, test")
# ------------------------------------------------------
# placeholders, used to feed data to the model
# ------------------------------------------------------
self.X = None
self.y = None
self.keep_prob = None
self.is_training = None
self.global_step = None
# ------------------------------------------------------
# tensorflow nodes used to evaluate
# ------------------------------------------------------
# prediction for each class(probability distribution)
self.prediction = None
# accuracy of batch
self.accuracy = None
# total loss of batch
self.loss = None
# training operation
self.train_op = None
self.novelty = None
self.novelty_distance = None
self.state_in = None
self.state_out = None
self.logits_out = None
self.hidden_out = None
self.lstm_out = None
# we store 1000 samples and use these to plot projections during training
self.train_samples = None
self.val_samples = None
# number of samples to use when evaluating the model, 1000 works well but is a bit slow,
# 100 should give results to within a few percent.
self.eval_samples = 500
# number of samples to use when generating the model report,
# atleast 1000 is recommended for a good representation
self.report_samples = 2000
# how often to do an evaluation + print
self.print_every = 6000
# restore best weights found during training rather than the most recently one.
self.use_best_weights = True
# the score this model got on it's final evaluation
self.eval_score = None
# our current global step
self.step = 0
# enabled parallel loading and training on data (much faster)
self.enable_async_loading = True
# folder to write tensorboard logs to
if train_config:
self.log_dir = os.path.join(train_config.train_dir, "logs")
self.checkpoint_folder = os.path.join(train_config.train_dir,
"checkpoints")
else:
self.log_dir = "./logs"
self.checkpoint_folder = "./checkpoints"
self.log_id = ""
# number of frames per segment during training
self.training_segment_frames = 27
# number of frames per segment during testing
self.testing_segment_frames = 27
# dictionary containing current hyper parameters
self.params = {
# augmentation
"augmentation": True,
"thermal_threshold": 10,
"scale_frequency": 0.5,
# dropout
"keep_prob": 0.5,
# training
"batch_size": 16,
}
""" List of labels this model can classifiy. """
self.labels = []
# used for tensorboard
self.writer_train = None
self.writer_val = None
self.merged_summary = None
# this defines our input shape
self.frame_count = 1
if self.training:
self.frame_count = self.training_segment_frames
def __init__(self, session=None):
self.name = "model"
self.session = session or tools.get_session()
self.saver = None
# datasets
self.datasets = namedtuple('Datasets', 'train, validation, test')
# ------------------------------------------------------
# placeholders, used to feed data to the model
# ------------------------------------------------------
self.X = None
self.y = None
self.keep_prob = None
self.is_training = None
self.global_step = None
# ------------------------------------------------------
# tensorflow nodes used to evaluate
# ------------------------------------------------------
# prediction for each class(probability distribution)
self.prediction = None
# accuracy of batch
self.accuracy = None
# total loss of batch
self.loss = None
# training operation
self.train_op = None
self.novelty = None
self.novelty_distance = None
self.state_in = None
self.state_out = None
self.logits_out = None
self.hidden_out = None
self.lstm_out = None
# we store 1000 samples and use these to plot projections during training
self.train_samples = None
self.val_samples = None
# number of samples to use when evaluating the model, 1000 works well but is a bit slow,
# 100 should give results to within a few percent.
self.eval_samples = 500
# number of samples to use when generating the model report,
# atleast 1000 is recommended for a good representation
self.report_samples = 2000
# how often to do an evaluation + print
self.print_every = 6000
# restore best weights found during training rather than the most recently one.
self.use_best_weights = True
# the score this model got on it's final evaluation
self.eval_score = None
# our current global step
self.step = 0
# enabled parallel loading and training on data (much faster)
self.enable_async_loading = True
# folder to write tensorboard logs to
self.log_dir = './logs'
self.log_id = ''
# number of frames per segment during training
self.training_segment_frames = 27
# number of frames per segment during testing
self.testing_segment_frames = 27
# dictionary containing current hyper parameters
self.params = {
# augmentation
'augmentation': True,
'thermal_threshold': 10,
'scale_frequency': 0.5,
# dropout
'keep_prob': 0.5,
# training
'batch_size': 16
}
""" List of labels this model can classifiy. """
self.labels = []
# used for tensorboard
self.writer_train = None
self.writer_val = None
self.merged_summary = None