def __init__(self, rare_words_handling=None, words_min_occur=None):
"""
If an arg is None, it will get its value from config.active_config.
Args:
rare_words_handling: {'nothing'|'discard'|'change'}
words_min_occur: words whose occurrences are less than this are
considered rare words
"""
self._tokenizer = Tokenizer(num_words=100000)
self._rare_words_handling = (rare_words_handling or
active_config().rare_words_handling)
self._words_min_occur = (words_min_occur or
active_config().words_min_occur)
self._word_of = {}
def get_dataset_instance(dataset_name=None,
lemmatize_caption=None,
single_caption=False):
"""
If an arg is None, it will get its value from config.active_config.
"""
dataset_name = dataset_name or active_config().dataset_name
lemmatize_caption = lemmatize_caption or active_config().lemmatize_caption
for dataset_class in [Flickr8kDataset]:
if dataset_class.DATASET_NAME == dataset_name:
return dataset_class(lemmatize_caption=lemmatize_caption,
single_caption=single_caption)
raise ValueError('Cannot find {} dataset!'.format(dataset_name))
def main(training_dir,
dataset_type='validation',
method='beam_search',
beam_size=3,
max_caption_length=20):
if method != 'beam_search':
raise NotImplementedError('inference method = {} is not implemented '
'yet!'.format(method))
if dataset_type not in ['validation', 'test']:
raise ValueError('dataset_type={} is not recognized!'.format(
dataset_type))
hyperparam_path = os.path.join(training_dir, 'hyperparams-config.yaml')
model_weights_path = os.path.join(training_dir, 'model-weights.hdf5')
logging('Loading hyperparams config..')
config = FileConfigBuilder(hyperparam_path).build_config()
config = config._replace(word_vector_init=None) # As we do an inference
active_config(config)
model = ImageCaptioningModel()
logging('Building model..')
model.build()
keras_model = model.keras_model
logging('Loading model weights..')
keras_model.load_weights(model_weights_path)
inference = BeamSearchInference(keras_model,
beam_size=beam_size,
max_caption_length=max_caption_length)
logging('Evaluating {} set..'.format(dataset_type))
if dataset_type == 'test':
metrics, predictions = inference.evaluate_test_set(
include_prediction=True)
else:
metrics, predictions = inference.evaluate_validation_set(
include_prediction=True)
logging('Writting result to files..')
metrics_path = os.path.join(training_dir,
'{}-metrics-{}-{}.yaml'.format(dataset_type, beam_size,
max_caption_length))
predictions_path = os.path.join(training_dir,
'{}-predictions-{}-{}.yaml'.format(dataset_type, beam_size,
max_caption_length))
write_yaml_file(metrics, metrics_path)
write_yaml_file(predictions, predictions_path)
logging('Done!')
def __init__(self,
training_label_prefix,
dataset_name=None,
epochs=None,
time_limit=None,
num_gpus=None):
if not ((epochs is None) ^ (time_limit is None)):
raise ValueError('epochs or time_limit must present, '
'but not both!')
self._training_label_prefix = training_label_prefix
self._dataset_name = dataset_name or active_config().dataset_name
self._validate_training_label_prefix()
self._epochs = epochs
self._time_limit = time_limit
fixed_config_keys = dict(dataset_name=self._dataset_name,
epochs=self._epochs,
time_limit=self._time_limit)
self._config_builder = Embed300FineRandomConfigBuilder(
fixed_config_keys)
try:
self._num_gpus = len(sh.nvidia_smi('-L').split('\n')) - 1
except sh.CommandNotFound:
self._num_gpus = 1
self._num_gpus = num_gpus or self._num_gpus
# TODO ! Replace set with a thread-safe set
self._available_gpus = set(range(self.num_gpus))
self._semaphore = Semaphore(self.num_gpus)
self._running_commands = [] # a list of (index, sh.RunningCommand)
self._stop_search = False
self._lock = Lock()
def __init__(self, image_augmentation=None):
self._image_data_generator = ImageDataGenerator(rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest')
if image_augmentation is None:
self._image_augmentation_switch = active_config().image_augmentation
else:
self._image_augmentation_switch = image_augmentation
def __init__(self,
training_label,
conf=None,
model_weights_path=None,
min_delta=1e-4,
min_lr=1e-7,
log_metrics_period=4,
explode_ratio=0.25,
explode_patience=2,
max_q_size=10,
workers=1,
verbose=1):
"""
Args
conf: an instance of config.Config; its properties:
epochs
time_limit
reduce_lr_factor
reduce_lr_patience
early_stopping_patience
"""
self._training_label = training_label
self._config = conf or config.active_config()
self._epochs = self._config.epochs
self._time_limit = self._config.time_limit
self._reduce_lr_factor = self._config.reduce_lr_factor
self._reduce_lr_patience = self._config.reduce_lr_patience
self._early_stopping_patience = self._config.early_stopping_patience
self._model_weights_path = model_weights_path
self._min_delta = min_delta
self._min_lr = min_lr
self._log_metrics_period = log_metrics_period
self._explode_ratio = explode_ratio
self._explode_patience = explode_patience
self._max_q_size = max_q_size
self._workers = workers
self._verbose = verbose
if not ((self._epochs is None) ^ (self._time_limit is None)):
raise ValueError('Either conf.epochs or conf.time_limit must be '
'set, but not both!')
if self._time_limit:
self._epochs = sys.maxsize
self._activate_config_and_init_dataset_provider()
self._init_result_dir()
self._init_callbacks()
self._model = ImageCaptioningModel()
self._write_active_config()
self._stop_training = False
def __init__(self,
batch_size=None,
dataset=None,
image_preprocessor=None,
caption_preprocessor=None,
single_caption=False):
"""
If an arg is None, it will get its value from config.active_config.
"""
self._batch_size = batch_size or active_config().batch_size
self._dataset = (dataset or
get_dataset_instance(single_caption=single_caption))
self._image_preprocessor = image_preprocessor or ImagePreprocessor()
self._caption_preprocessor = (caption_preprocessor
or CaptionPreprocessor())
self._single_caption = single_caption
self._build()
def __init__(self,
learning_rate=None,
vocab_size=None,
embedding_size=None,
rnn_output_size=None,
dropout_rate=None,
bidirectional_rnn=None,
rnn_type=None,
rnn_layers=None,
l1_reg=None,
l2_reg=None,
initializer=None,
word_vector_init=None):
"""
If an arg is None, it will get its value from config.active_config.
"""
self._learning_rate = learning_rate or active_config().learning_rate
self._vocab_size = vocab_size or active_config().vocab_size
self._embedding_size = embedding_size or active_config().embedding_size
self._rnn_output_size = (rnn_output_size
or active_config().rnn_output_size)
self._dropout_rate = dropout_rate or active_config().dropout_rate
self._rnn_type = rnn_type or active_config().rnn_type
self._rnn_layers = rnn_layers or active_config().rnn_layers
self._word_vector_init = (word_vector_init
or active_config().word_vector_init)
self._initializer = initializer or active_config().initializer
if self._initializer == 'vinyals_uniform':
self._initializer = RandomUniform(-0.08, 0.08)
if bidirectional_rnn is None:
self._bidirectional_rnn = active_config().bidirectional_rnn
else:
self._bidirectional_rnn = bidirectional_rnn
l1_reg = l1_reg or active_config().l1_reg
l2_reg = l2_reg or active_config().l2_reg
self._regularizer = l1_l2(l1_reg, l2_reg)
self._keras_model = None
if self._vocab_size is None:
raise ValueError('config.active_config().vocab_size cannot be '
'None! You should check your config or you can '
'explicitly pass the vocab_size argument.')
if self._rnn_type not in ('lstm', 'gru'):
raise ValueError('rnn_type must be either "lstm" or "gru"!')
if self._rnn_layers < 1:
raise ValueError('rnn_layers must be >= 1!')
if self._word_vector_init is not None and self._embedding_size != 300:
raise ValueError('If word_vector_init is not None, embedding_size '
'must be 300')
def _write_active_config(self):
CONFIG_FILENAME = 'hyperparams-config.yaml'
self._config_filepath = self._path_from_result_dir(CONFIG_FILENAME)
config.write_to_file(config.active_config(), self._config_filepath)
def _activate_config_and_init_dataset_provider(self):
config.active_config(self._config)
self._dataset_provider = DatasetProvider()
config.init_vocab_size(self._dataset_provider.vocab_size)
def config_used(self):
config = active_config()
config = config._replace(epochs=2, time_limit=None, batch_size=2)
return config
