def __init__(self, pc, model_hyperparams, best_model_path=None):
self.hyperparams = model_hyperparams
print 'Loading vocabulary from {}:'.format(
self.hyperparams['VOCAB_PATH'])
self.vocab = Vocab.from_file(self.hyperparams['VOCAB_PATH'])
print 'Loading feature vocabulary from {}:'.format(
self.hyperparams['FEAT_VOCAB_PATH'])
self.feat_vocab = Vocab.from_file(self.hyperparams['FEAT_VOCAB_PATH'])
self.BEGIN = self.vocab.w2i[BEGIN_CHAR]
self.STOP = self.vocab.w2i[STOP_CHAR]
self.UNK = self.vocab.w2i[UNK_CHAR]
self.hyperparams['VOCAB_SIZE'] = self.vocab.size()
self.hyperparams['FEAT_VOCAB_SIZE'] = self.feat_vocab.size()
self.build_model(pc, best_model_path)
print 'Model Hypoparameters:'
for k, v in self.hyperparams.items():
print '{:20} = {}'.format(k, v)
print
lm_model_params = []
## loading the language models
for i,(lm_type,path,order) in enumerate(\
zip(\
arguments['--lm_predictors'].split(','),\
arguments["--lm_paths"].split(','),\
[int(o) for o in arguments["--lm_orders"].split(',')]\
)):
lm_model_folder = check_path(path, 'LM_MODEL_FOLDER_{}'.format(i), is_data_path=False)
if lm_type=="srilm_char":
print '...Loading lm model {} from path {}'.format(i,lm_model_folder)
lm_model = SRILM_char_lm_loader(lm_model_folder, order)
elif lm_type=="srilm_morph":
lm_model = SRILM_morpheme_lm_loader(lm_model_folder,order)
assert arguments['--morph_vocab'] != None
lm_model.vocab = Vocab.from_file(check_path(arguments['--morph_vocab'], 'morph_vocab', is_data_path=False))
else:
print "WARNING -- Could not load language model. Unknown type",lm_type,". Use 'srilm_char' or 'srilm_morph'"
lm_models.append(lm_model)
lm_number = len(lm_models)
output_file_path = os.path.join(model_folder,arguments['--pred_path'])
# save best dev model parameters and predictions
print 'Evaluating on test..'
t = time.clock()
if arguments['--indices']:
indices = [int(ind) for ind in arguments['--indices'].split(',')]
accuracy, test_results = evaluate_syncbeam(test_data.iter(indices), nmt_models, lm_models, weights, int(arguments['--beam']), int(arguments['--output_format']), verbose=arguments['--verbose'], exclude_eow=arguments['--exclude_eow'])
else:
accuracy, test_results = evaluate_syncbeam(test_data.iter(), nmt_models, lm_models, weights, int(arguments['--beam']), int(arguments['--output_format']), verbose=arguments['--verbose'], exclude_eow=arguments['--exclude_eow'])