def split_data():
print("Data splitting . . ", end="")
data = pd.read_table(helper.root_dir() + '/data/data.txt', sep="*")
train_data, test_data = train_test_split(data, test_size=0.1, shuffle=True)
train_data.to_csv(helper.root_dir() + '/data/train.txt',
sep="*",
index=None)
test_data.to_csv(helper.root_dir() + '/data/test.txt', sep="*", index=None)
print("Done. Data saved to data folder. Test: {} rows. Train: {} rows.".
format(train_data.shape[0], test_data.shape[0]))
def run(epochs=50, batch_size=256):
data_path = helper.root_dir() + '/data/'
generator = Generator(batch_size=batch_size)
num_training_samples = generator.training_dataset.shape[0]
num_validation_samples = generator.validation_dataset.shape[0]
print('Number of training samples:', num_training_samples)
print('Number of validation samples:', num_validation_samples)
nn = model.generate(max_token_length=generator.MAX_TOKEN_LENGTH,
vocabulary_size=generator.VOCABULARY_SIZE)
nn.compile(loss='categorical_crossentropy', optimizer='adam')
models_path = data_path + '/weights/' + datetime.now().strftime(
'%Y-%m-%d-%H-%M')
os.makedirs(models_path)
model_names = (models_path + '/nn_weights.{epoch:02d}-{val_loss:.2f}.hdf5')
model_checkpoint = ModelCheckpoint(model_names,
monitor='loss',
verbose=1,
save_best_only=True,
mode='min')
tboard = TensorBoard(log_dir=data_path + '/tboard/' +
datetime.now().strftime('%Y-%m-%d-%H-%M'))
callbacks = [model_checkpoint, tboard]
nn.fit_generator(generator=generator.flow(train=True),
epochs=epochs,
steps_per_epoch=int(num_training_samples / batch_size),
validation_data=generator.flow(train=False),
validation_steps=int(num_validation_samples / batch_size),
callbacks=callbacks,
verbose=1)
def __init__(self):
self.image_model = InceptionV3(weights='imagenet')
self.generator = Generator(batch_size=32)
self.nn = model.generate(
max_token_length=self.generator.MAX_TOKEN_LENGTH,
vocabulary_size=self.generator.VOCABULARY_SIZE)
data_path = helper.root_dir() + '/data/'
models_path = data_path + '/weights/2018-03-25-14-18'
model_names = (models_path + '/nn_weights.25-2.94.hdf5')
self.nn.load_weights(model_names)
def __init__(self, run_inception=False, word_threshold=-1):
self.root_path = helper.root_dir()
self.word_threshold = word_threshold
self.max_caption_length = 0
self.run_inception = run_inception
self.IMG_FEATURES = 1000 # inception model
self.BOS = '<S>' # Beginning Of Sentence
self.EOS = '<E>' # End Of Sentence
self.PAD = '<P>'
self.word_frequencies = None
self.captions = None
self.image_files = None
self.image_features = None
self.word_to_id = None
self.id_to_word = None
self.extracted_features = None
self.features_file_names = None
self.image_feature_files = None
self.vocabulary_size = 0
def __init__(self, batch_size=1):
self.data_path = helper.root_dir() + '/data/'
self.dictionary = None
self.training_dataset = None
self.validation_dataset = None
self.image_names_to_features = None
data_logs = np.genfromtxt(self.data_path + 'data_parameters.log', delimiter=' ', dtype='str')
data_logs = dict(zip(data_logs[:, 0], data_logs[:, 1]))
self.MAX_TOKEN_LENGTH = int(data_logs['max_caption_length:']) + 2
self.IMG_FEATS = int(data_logs['IMG_FEATS:'])
self.BOS = str(data_logs['BOS:'])
self.EOS = str(data_logs['EOS:'])
self.PAD = str(data_logs['PAD:'])
self.VOCABULARY_SIZE = None
self.word_to_id = None
self.id_to_word = None
self.BATCH_SIZE = batch_size
self._load_dataset()
self._load_vocabulary()
self._load_image_features()