def generate_text(self, start_string: str, num_words: int,
temperature: float) -> str:
"""Generate string starting with start_string of length num_characters using rebuilt model.
Resetting of the graph must be done with every POST request otherwise the model won't run.
:param start_string: str user wishes to start generation with. Can be a single letter.
:param num_words: number of words you wish to be generated. Note time to generate increases.
:param temperature: parameter that determines how 'surprising' the predictions are. value of 1 is neutral,
lower is more predictable, higher is more surprising.
:return: string of generated text
"""
load_start_time = datetime.now()
tf.compat.v1.reset_default_graph()
self.sess = gpt2.start_tf_sess()
gpt2.load_gpt2(self.sess,
run_name=self.model_folder,
checkpoint_dir=self.checkpoint_directory)
time_to_load = datetime.now() - load_start_time
logging.info(f'Time taken to load model {time_to_load}')
gen_start_time = datetime.now()
txt = \
gpt2.generate(self.sess, run_name=self.model_folder, checkpoint_dir=self.checkpoint_directory,
length=num_words, temperature=temperature, prefix=start_string, return_as_list=True)[0]
time_to_generate = datetime.now() - gen_start_time
logging.info(f'Time taken to generate lyrics {time_to_generate}')
txt = CleanOutput.remove_lastline(text_in=txt)
txt = CleanOutput.capitalise_first_character(text_in=txt)
txt = CleanOutput.clean_line(text_in=txt)
txt = CleanOutput.sanitise_string(text_in=txt,
custom_badwords=custom_badwords)
self.sess.close()
return txt
def test_remove_lastline(text_in, expected):
"""Check that custom profanities are being redacted."""
assert CleanOutput.remove_lastline(text_in=text_in) == expected