def decode_dataset(params):
global decoder, saved_output, references, labels
lm_alpha, lm_beta = params
decoder.reset_params(lm_alpha, lm_beta)
hypothesis = ctc_beam_decode_batch(saved_output, decoder, labels)
wer = get_wer(references, hypothesis)
return [lm_alpha, lm_beta, wer * 100]
def merge_transcripts(params):
global data, labels, utterances, references, args
th, service_weight, blank_conf = params
th_blank = 10 * th
new_probs_list = []
for utt in utterances:
ds2_probs = data[utt]['ds2_probs']
ctc_alignment = data[utt]['alignment']
word_confs = data[utt]['word_confs']
frame_confs = get_frame_lvl_cnfs(ctc_alignment, word_confs, labels,
blank_conf)
new_probs = fine_merge(ds2_probs, ctc_alignment, frame_confs, th,
th_blank, service_weight, labels.index('_'))
new_probs_list.append(new_probs)
new_transcipts = ctc_beam_decode(new_probs_list, labels, args.lm_path,
labels.index('_'), args.lm_alpha,
args.lm_beta, args.beam_width)
wer = get_wer(references, new_transcipts)
return th, service_weight, blank_conf, wer
# Initialize dataloader
eval_cfg_dict = dict(type="audio,transcription",
audio=feats_config,
transcription=transcripts_config,
manifest_filename=eval_manifest,
macrobatch_size=be.bsz,
minibatch_size=be.bsz)
eval_set = DataLoader(backend=be, config=eval_cfg_dict)
eval_set = data_transform(eval_set)
# Load the model
model = Model(args.model_file)
# Process data and compute stats
wer, sample_size, results = get_wer(model, be, eval_set,
beam_decoder,
nout,
use_wer=args.use_wer)
print("\n" + "-" * 80)
if args.use_wer:
print("wer = {}".format(wer))
else:
print("cer = {}".format(wer))
print("-" * 80 + "\n")
if args.inference_file:
# Save results in args.inference_file
with open(args.inference_file, 'wb') as f:
pkl.dump((results, wer), f, protocol=0)
print("Saved inference results to {}".format(args.inference_file))
eval_manifest = args.manifest['val']
if not os.path.exists(eval_manifest):
raise IOError("Manifest file {} not found".format(eval_manifest))
# Setup required dataloader parameters
nbands = 13
max_utt_len = 30
max_tscrpt_len = 1300
eval_set = make_loader(eval_manifest, alphabet, nbands, max_tscrpt_len,
max_utt_len, backend_obj=be)
# Load the model
model = Model(args.model_file)
# Process data and compute stats
wer, sample_size, results = get_wer(model, be, eval_set, argmax_decoder, nout,
use_wer=args.use_wer, print_examples=args.print_examples)
print("\n" + "-" * 80)
if args.use_wer:
print("wer = {}".format(wer))
else:
print("cer = {}".format(wer))
print("-" * 80 + "\n")
if args.inference_file:
# Save results in args.inference_file
with open(args.inference_file, 'wb') as f:
pkl.dump((results, wer), f)
print("Saved inference results to {}".format(args.inference_file))