def test():
model_file = "/disc1/ARNet/exp/libri/res18_gru/012.h5"
FEATS_DEV_CLEAN = us.kaldiio.load_scp(dev_clean_file)
FEATS_DEV_OTHER = us.kaldiio.load_scp(dev_other_file)
FEATS_TEST_CLEAN = us.kaldiio.load_scp(test_clean_file)
FEATS_TEST_OTHER = us.kaldiio.load_scp(test_other_file)
dev_clean_lst = us.scp2key(us.read_lines(dev_clean_file))
dev_other_lst = us.scp2key(us.read_lines(dev_other_file))
test_clean_lst = us.scp2key(us.read_lines(test_clean_file))
test_other_lst = us.scp2key(us.read_lines(test_other_file))
dev_clean_data = us.load_ctc(dev_clean_lst, FEATS_DEV_CLEAN,
encoder_len=ENCODER_LEN,
max_input_len=MAX_INPUT_LEN,
max_label_len=MAX_LABEL_LEN,
trans_ids=us.LIBRI_TRANS_IDS)
dev_other_data = us.load_ctc(dev_other_lst, FEATS_DEV_OTHER,
encoder_len=ENCODER_LEN,
max_input_len=MAX_INPUT_LEN,
max_label_len=MAX_LABEL_LEN,
trans_ids=us.LIBRI_TRANS_IDS)
test_clean_data = us.load_ctc(test_clean_lst, FEATS_TEST_CLEAN,
encoder_len=ENCODER_LEN,
max_input_len=MAX_INPUT_LEN,
max_label_len=MAX_LABEL_LEN,
trans_ids=us.LIBRI_TRANS_IDS)
test_other_data = us.load_ctc(test_other_lst, FEATS_TEST_OTHER,
encoder_len=ENCODER_LEN,
max_input_len=MAX_INPUT_LEN,
max_label_len=MAX_LABEL_LEN,
trans_ids=us.LIBRI_TRANS_IDS)
model = mdl.model_res_gru_ctc(shapes=(MAX_INPUT_LEN, FEAT_DIM, 1),
bpe_classes=BPE_CLASSES,
max_label_len=MAX_LABEL_LEN,
cnn=CNN,
raw_model=model_file)
ctc_decode_model = mdl.sub_model(model, 'inputs', 'ctc_pred')
dev_clean_pred = mdl.ctc_pred(ctc_decode_model, dev_clean_data[0], input_len=ENCODER_LEN, batch_size=BATCH_SIZE)
dev_other_pred = mdl.ctc_pred(ctc_decode_model, dev_other_data[0], input_len=ENCODER_LEN, batch_size=BATCH_SIZE)
test_clean_pred = mdl.ctc_pred(ctc_decode_model, test_clean_data[0], input_len=ENCODER_LEN, batch_size=BATCH_SIZE)
test_other_pred = mdl.ctc_pred(ctc_decode_model, test_other_data[0], input_len=ENCODER_LEN, batch_size=BATCH_SIZE)
print("dev-clean-wer:",
us.ctc_eval(dev_clean_data[0]["ctc_labels"],dev_clean_data[0]["ctc_label_len"], dev_clean_pred, True))
print("dev-other-wer:",
us.ctc_eval(dev_other_data[0]["ctc_labels"], dev_other_data[0]["ctc_label_len"], dev_other_pred, True))
print("test-clean-wer:",
us.ctc_eval(test_clean_data[0]["ctc_labels"], test_clean_data[0]["ctc_label_len"], test_clean_pred, True))
print("test-other-wer:",
us.ctc_eval(test_other_data[0]["ctc_labels"], test_other_data[0]["ctc_label_len"], test_other_pred, True))
%(RES_TYPE, ASR_EN, MANY_TO_ONE, METRIC_LOSS, MARGIN,BN_DIM,1 if RAW_MODEL else 0)
MODEL_DIR = "exp/aesrc/%s/" % task
if not os.path.isdir(MODEL_DIR): os.mkdir(MODEL_DIR)
# file
train_file = "/disc1/AESRC2020/data/aesrc_fbank/train.scp"
dev_file = "/disc1/AESRC2020/data/aesrc_fbank/dev.scp"
# feats
FEATS = us.kaldiio.load_scp(train_file)
FEATS_DEV = us.kaldiio.load_scp(dev_file)
# list
train_lst = us.scp2key(us.read_lines(train_file))
dev_lst = us.scp2key(us.read_lines(dev_file))
N_BATCHS = len(train_lst) // BATCH_SIZE
lr_reducer = ReduceLROnPlateau(factor=0.3, cooldown=0, patience=1, min_lr=1e-5,
monitor='val_accent_labels_acc', mode='max', min_delta=0.001, verbose=1)
early_stopper = EarlyStopping(patience=3,
monitor='val_accent_labels_acc', mode='max', min_delta=0.001, verbose=1)
csv_logger = CSVLogger('%s/train.csv' % MODEL_DIR)
# generator
generator = us.generator_sarnet(train_lst, FEATS, BATCH_SIZE,
encoder_len=ENCODER_LEN,
max_input_len=MAX_INPUT_LEN,
max_label_len=MAX_LABEL_LEN,
CNN = 'res18'
RAW_MODEL = '/disc1/ARNet/exp/aesrc/res18_gru2/002.h5'
LOG_FILE = '/disc1/ARNet/exp/aesrc/res18_gru2/model.csv'
MODEL_DIR = '/disc1/ARNet/exp/aesrc/res18_gru2/'
# file
train_file = "/disc1/AESRC2020/data/aesrc_fbank_sp/train.scp"
dev_file = "/disc1/AESRC2020/data/aesrc_fbank_sp/dev.scp"
# feats
FEATS = us.kaldiio.load_scp(train_file)
FEATS_DEV = us.kaldiio.load_scp(dev_file)
# list
train_lst = us.limit_time_utts(
us.limit_trans_utts(us.scp2key(us.read_lines(train_file)),
us.AESRC_TRANS_IDS, MAX_LABEL_LEN),
us.AESRC_UTT2FRAMES, MAX_INPUT_LEN)
dev_lst = us.scp2key(us.read_lines(dev_file))
dev_lst = random.sample(dev_lst, 1000)
N_BATCHS = len(train_lst) // BATCH_SIZE
# callbacks
lr_reducer = ReduceLROnPlateau(factor=0.5,
cooldown=0,
patience=0,
min_lr=0.5e-6,
monitor='accent_labels_acc',
mode='max',
min_delta=0.5,
verbose=1)
7: "<RU>"
}
BPE_CLASSES = 1000
ACCENT_CLASSES = 8
MAX_INPUT_LEN = 1200
ENCODER_LEN = 114
MAX_LABEL_LEN = 72
FEAT_DIM = 80
EPOCHS = 10
BATCH_SIZE = 32
RAW_MODEL = "/disc1/ARNet/exp/aesrc/007.h5"
dev_file = "/disc1/AESRC2020/data/aesrc_fbank_sp/dev.scp"
FEATS_DEV = us.kaldiio.load_scp(dev_file)
dev_lst = us.scp2key(us.read_lines(dev_file))
# dev_lst = random.sample(dev_lst,100)
dev_data = us.load_ctc_accent(dev_lst,
FEATS_DEV,
encoder_len=ENCODER_LEN,
max_input_len=MAX_INPUT_LEN,
max_label_len=MAX_LABEL_LEN,
trans_ids=us.AESRC_TRANS_IDS,
accent_classes=ACCENT_CLASSES,
accent_dct=us.AESRC_ACCENT,
accent_ids=us.AESRC_ACCENT2INT)
print("==== test ====")
model = mdl.model_ctc_accent(shapes=(MAX_INPUT_LEN, FEAT_DIM, 1),
accent_classes=ACCENT_CLASSES,
bpe_classes=BPE_CLASSES,