def evaulate_trf(model, vocab, read_model, tsize, fres):
res_name = '{}:'.format(int(tsize)) + os.path.split(read_model)[-1]
tskdir = '{}/'.format(tsize)
# rescore
config = ' -vocab {} '.format(vocab)
config += ' -read {}.model '.format(read_model)
config += ' -nbest {} '.format(data(tskdir)[3])
config += ' -lmscore {0}.lmscore'.format(read_model)
model.use(config)
# WER
[read_nbest, read_templ, read_acscore, read_lmscore] = data(tskdir)[3:7]
read_lmscore = read_model + '.lmscore'
[wer, lmscale, acscale] = wb.TuneWER(read_nbest, read_templ,
wb.LoadScore(read_lmscore),
wb.LoadScore(read_acscore),
np.linspace(0.1, 0.9, 9))
print('wer={:.4f} lmscale={:.2f} acscale={:.2f}'.format(
wer, lmscale, acscale))
# calculate the ppl on wsj test
templ_txt = model.workdir + os.path.split(read_templ)[-1] + '.rmlabel'
wb.file_rmlabel(read_templ, templ_txt)
PPL_templ = model.ppl(vocab, read_model + '.model', templ_txt)
LL_templ = -wb.PPL2LL(PPL_templ, templ_txt)
# output the result
fres.Add(res_name, ['LL-wsj', 'PPL-wsj'], [LL_templ, PPL_templ])
fres.AddWER(res_name, wer)
def main():
print(sys.argv)
if len(sys.argv) == 1:
print('\"python run_ngram.py -train\" train \n',
'\"python run_ngram.py -rescore\" rescore nbest\n',
'\"python run_ngram.py -wer\" compute WER')
bindir = '../../tools/srilm/'
fres = wb.FRes('result.txt') # the result file
order_reg = [5]
for tsize in [1, 2, 4]:
tskdir = '{}/'.format(tsize)
workdir = tskdir + 'ngramlm/'
model = ngram.model(bindir, workdir)
for order in order_reg:
write_model = workdir + '{}gram.lm'.format(order)
write_name = '{}:KN{}'.format(tsize, order)
print(write_model)
if '-train' in sys.argv or '-all' in sys.argv:
if order_reg.index(order) == 0:
model.prepare(
data(tskdir)[0],
data(tskdir)[1],
data(tskdir)[2])
model.train(order, write_model)
if '-test' in sys.argv or '-all' in sys.argv:
PPL = [0] * 3
PPL[0] = model.ppl(write_model, order, data(tskdir)[0])
PPL[1] = model.ppl(write_model, order, data(tskdir)[1])
PPL[2] = model.ppl(write_model, order, data(tskdir)[2])
fres.AddPPL(write_name, PPL, data(tskdir)[0:3])
if '-rescore' in sys.argv or '-all' in sys.argv:
model.rescore(write_model, order,
data(tskdir)[3], write_model[0:-3] + '.lmscore')
if '-wer' in sys.argv or '-all' in sys.argv:
[nbest, templ] = data(tskdir)[3:5]
lmscore = wb.LoadScore(write_model[0:-3] + '.lmscore')
acscore = wb.LoadScore(data(tskdir)[5])
[wer, lmscale, acscale] = wb.TuneWER(nbest, templ, lmscore,
acscore,
np.linspace(0.1, 0.9, 9))
print('wer={} lmscale={} acscale={}'.format(
wer, lmscale, acscale))
fres.AddWER(write_name, wer)
trans_txt = workdir + os.path.split(templ)[-1] + '.txt'
wb.file_rmlabel(templ, trans_txt)
PPL_temp = model.ppl(write_model, order, trans_txt)
LL_temp = -wb.PPL2LL(PPL_temp, trans_txt)
fres.Add(write_name, ['LL-wsj', 'PPL-wsj'],
[LL_temp, PPL_temp])
def main():
print(sys.argv)
if len(sys.argv) == 1:
print('\"python run_ngram.py -train\" train \n',
'\"python run_ngram.py -rescore\" rescore nbest\n',
'\"python run_ngram.py -wer\" compute WER')
absdir = os.getcwd() + '/'
bindir = absdir + '../../tools/srilm/'
workdir = absdir + 'ngramlm/'
wb.mkdir(workdir)
datas = [absdir + i for i in data()]
result_file = absdir + 'models_ppl.txt' # the result file
model = ngram.model(bindir, workdir)
order_reg = [2, 3, 4, 5]
for order in order_reg:
write_model = workdir + '{}gram.lm'.format(order)
print(write_model)
if '-train' in sys.argv:
if order_reg.index(order) == 0:
model.prepare(datas[0], datas[1], datas[2])
model.train(order, write_model, absdir + 'models_ppl.txt')
if '-rescore' in sys.argv:
model.rescore(write_model, order, datas[3],
write_model[0:-3] + '.lmscore')
if '-wer' in sys.argv:
[nbest, templ] = datas[3:5]
lmscore = wb.LoadScore(write_model[0:-3] + '.lmscore')
acscore = wb.LoadScore(datas[5])
[wer, lmscale, acscale] = wb.TuneWER(nbest, templ, lmscore,
acscore,
np.linspace(0.1, 0.9, 9))
print('wer={} lmscale={} acscale={}'.format(wer, lmscale, acscale))
fres = wb.FRes(result_file)
fres.AddWER('KN{}'.format(order), wer)
trans_txt = workdir + os.path.split(templ)[-1] + '.txt'
wb.file_rmlabel(templ, trans_txt)
PPL_temp = model.ppl(write_model, order, trans_txt)
LL_temp = -wb.PPL2LL(PPL_temp, trans_txt)
fres.Add('KN{}'.format(order), ['LL-wsj', 'PPL-wsj'],
[LL_temp, PPL_temp])
model.prepare(data(tskdir)[0], data(tskdir)[1], data(tskdir)[2], data(tskdir)[3])
model.train(config)
else:
print('exist model: ' + write_model)
if '-test' in sys.argv or '-all' in sys.argv:
PPL = [0]*3
PPL[0] = model.ppl(write_model, data(tskdir)[0])
PPL[1] = model.ppl(write_model, data(tskdir)[1])
PPL[2] = model.ppl(write_model, data(tskdir)[2])
fres.AddPPL(write_name, PPL, data(tskdir)[0:3])
if '-rescore' in sys.argv or '-all' in sys.argv:
write_lmscore = write_model[0:-4] + '.lmscore'
model.rescore(write_model, data(tskdir)[3], write_lmscore)
if '-wer' in sys.argv or '-all' in sys.argv:
[read_nbest, read_templ, read_acscore, read_lmscore] = data(tskdir)[3:7]
read_lmscore = write_model[0:-4] + '.lmscore'
[wer, lmscale, acscale] = wb.TuneWER(read_nbest, read_templ,
read_lmscore, read_acscore, np.linspace(0.1, 0.9, 9))
print('wer={:.4f} lmscale={:.2f} acscale={:.2f}'.format(wer, lmscale, acscale))
fres.AddWER(write_name, wer)
write_templ_rm = workdir + os.path.split(read_templ)[1] + '.rmlabel'
rnn.Nbest_rmlabel(read_templ, write_templ_rm)
PPL_templ = model.ppl(write_model, write_templ_rm)
LL_templ = -wb.PPL2LL(PPL_templ, write_templ_rm)
fres.Add(write_name, ['LL-wsj', 'PPL-wsj'], [LL_templ, PPL_templ])