loading_timer = tm.time()
schema = Schema(model_args.schema_path, None)
data_generator = get_data_generator(args, model_args, schema)
mappings = data_generator.mappings
if args.vocab_only:
import sys; sys.exit()
if args.verbose:
print("Finished loading and pre-processing data, took {:.1f} seconds".format(tm.time() - loading_timer))
# TODO: load from checkpoint
ckpt = None
# Build the model
model = build_model(model_args, args, mappings, ckpt, model_path=args.agent_checkpoint)
tally_parameters(model)
create_path(args.model_path)
config_path = os.path.join(args.model_path, 'config.json')
write_json(vars(args), config_path)
builder = UtteranceBuilder(mappings['tgt_vocab'], 1, has_tgt=True)
# Build optimizer and trainer
optim = build_optim(args, model, ckpt)
# vocab is used to make_loss, so use target vocab
trainer = build_trainer(args, model, mappings['tgt_vocab'], optim)
trainer.builder = builder
# Perform actual training
trainer.learn(args, data_generator, report_func)
examples = read_examples(args.transcripts, args.max_examples, Scenario)
parsed_dialogues = []
templates = Templates()
lexicon = Lexicon(['ball', 'hat', 'book'])
for example in examples:
utterances = parse_example(example, lexicon, templates)
parsed_dialogues.append(utterances)
templates.finalize()
templates.save(args.templates_output)
templates.dump(n=10)
if args.transcripts_output:
write_json([e.to_dict() for e in examples], args.transcripts_output)
# Train n-gram model
sequences = []
for d in parsed_dialogues:
sequences.append([u.lf.intent for u in d])
manager = Manager.from_train(sequences)
manager.save(args.model_output)
generator = Generator(templates)
action = manager.choose_action(None, context=('<start>', '<start>'))
print action
print generator.retrieve('<start>', context_tag='<start>',
tag=action).template