def train(self,
order,
write_model,
output_res='',
discount='-kndiscount',
cutoff=[]):
write_count = self.workdir + os.path.split(write_model)[1] + '.count'
cmd = self.bindir + 'ngram-count '
cmd += ' -text {0}train.no -vocab {0}vocab'.format(self.workdir)
cmd += ' -order {} -write {} '.format(order, write_count)
for i in range(len(cutoff)):
cmd += ' -gt{}min {}'.format(i + 1, cutoff[i])
os.system(cmd)
cmd = self.bindir + 'ngram-count '
cmd += ' -vocab {}vocab'.format(self.workdir)
cmd += ' -read {}'.format(write_count)
cmd += ' -order {} -lm {} '.format(order, write_model)
cmd += discount + ' -interpolate -gt1min 0 -gt2min 0 -gt3min 0 -gt4min 0 -gt5min 0'
os.system(cmd)
# get ppl
if output_res != '':
PPL = [0] * 3
testno = [
self.workdir + i + '.no' for i in ['train', 'valid', 'test']
]
for i in range(3):
cmd = self.bindir + 'ngram -order {} -lm {} -ppl {}'.format(
order, write_model, testno[i])
res = os.popen(cmd).read()
PPL[i] = float(res[res.find('ppl='):].split()[1])
res_file = wb.FRes(output_res)
res_file.AddPPL('KN{}'.format(order), PPL, testno)
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 wer_all(lmpaths, lmtypes, outlog):
fres = wb.FRes(outlog, True)
fres.Clean()
[read_nbest, read_trans, read_acscore] = data()[3:6]
lmscale_vec = np.linspace(0.1, 0.9, 9)
weight_vec = np.linspace(0.5, 0.5, 1)
for type in lmtypes:
exist_multiple_run = False
a = type.split('+')
for lm in a:
if lmpaths[lm].find('<run>') != -1:
exist_multiple_run = True
break
run_vec = [0]
run_name = type
if exist_multiple_run:
run_vec = range(0, 10)
run_name = type + ':<run>'
for run in run_vec:
run_str = 'run{}'.format(run)
name = run_name.replace('<run>', run_str)
opt_wer_vec = [100, 1.0, 1.0]
opt_weight = 1.0
if len(a) == 1:
lmscore = wb.LoadScore(lmpaths[a[0]].replace('<run>', run_str))
opt_wer_vec = wb.TuneWER(read_nbest, read_trans, lmscore,
read_acscore, lmscale_vec)
opt_weight = 1.0
else:
lmscore1 = np.array(
wb.LoadScore(lmpaths[a[0]].replace('<run>', run_str)))
lmscore2 = np.array(
wb.LoadScore(lmpaths[a[1]].replace('<run>', run_str)))
for w in weight_vec:
lmscore = w * lmscore1 + (1 - w) * lmscore2
[wer, lmscale,
acscale] = wb.TuneWER(read_nbest, read_trans, lmscore,
read_acscore, lmscale_vec)
if wer < opt_wer_vec[0]:
opt_wer_vec = [wer, lmscale, acscale]
opt_weight = w
fres.Add(name, ['wer', 'lmscale', 'acscale', 'weight'],
opt_wer_vec + [opt_weight])
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])
def main():
if len(sys.argv) == 1:
print('\"python run.py -train\" train LSTM\n',
'\"python run.py -rescore\" rescore nbest\n',
'\"python run.py -wer\" compute WER'
)
run_times = range(0, 10) # for multiple run
bindir = '../../tools/trf/bin/'
workdir = 'trflm/'
fres = wb.FRes('models_ppl.txt')
model = trf.model(bindir, workdir)
class_num = 200
train = workdir + 'train.id'
valid = workdir + 'valid.id'
test = workdir + 'test.id'
vocab = workdir + 'vocab_c{}.list'.format(class_num)
order = 4
feat = 'g4_w_c_ws_cs_wsh_csh_tied.fs'
#feat = 'g4_w_c_ws_cs_cpw.fs'
maxlen = 0
tmax = 20000
t0 = 2000
minibatch = 100
gamma_lambda = '3000,0'
gamma_zeta = '0,0.6'
reg = 4e-5
thread = 8
if '-res' in sys.argv:
fres.Read()
for i in range(1,len(fres.head)):
value = []
for runnum in run_times:
write_name = 'trf_c{}_{}.run{}'.format(class_num, feat[0:-3], runnum)
line = fres.Get(write_name)
value.append(line[i])
fres.Add('trf_c{}_{}.runavg'.format(class_num, feat[0:-3]), [fres.head[i]],
['{:.2f}+{:.2f}'.format(np.mean(value), np.std(value))] )
for runnum in run_times:
write_model = workdir + 'trf_c{}_{}.run{}'.format(class_num, feat[0:-3], runnum)
if '-train' in sys.argv or '-all' in sys.argv:
config = '-vocab {} -train {} -valid {} -test {} '.format(vocab, train, valid, test)
config += ' -order {} -feat {} '.format(order, feat)
config += ' -len {} '.format(maxlen)
config += ' -write {0}.model -log {0}.log '.format(write_model)
config += ' -t0 {} -iter {}'.format(t0, tmax)
config += ' -gamma-lambda {} -gamma-zeta {}'.format(gamma_lambda, gamma_zeta)
config += ' -L2 {} '.format(reg)
config += ' -mini-batch {} '.format(minibatch)
config += ' -thread {} '.format(thread)
config += ' -print-per-iter 10 '
config += ' -write-at-iter [{}:1000:{}]'.format(tmax-5000, tmax) # output the intermediate models
model.prepare(data()[0], data()[1], data()[2], class_num)
model.train(config)
if '-plot' in sys.argv:
baseline = fres.Get('KN5')
trf.PlotLog([write_model], [baseline])
if '-rescore' in sys.argv or '-all' in sys.argv:
config = ' -vocab {} '.format(vocab)
config += ' -read {}.model '.format(write_model)
config += ' -nbest {} '.format(data()[3])
config += ' -lmscore {0}.lmscore -lmscore-test-id {0}.test-id '.format(write_model)
model.use(config)
if '-wer' in sys.argv or '-all' in sys.argv:
[read_nbest, read_templ, read_acscore, read_lmscore] = data()[3:7]
read_lmscore = write_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
write_templ_id = workdir + os.path.split(read_templ)[1] + '.id'
v = trf.ReadVocab(vocab)
trf.NbestToID(read_templ, write_templ_id, v)
config = ' -vocab {} '.format(vocab)
config += ' -read {}.model '.format(write_model)
config += ' -test {} '.format(write_templ_id)
LL_templ = model.use(config)
PPL_templ = wb.LL2PPL(-LL_templ, write_templ_id)
LL = model.get_last_value(write_model + '.log')
# output the result
name = os.path.split(write_model)[1]
fres.AddLL(name, LL, data()[0:3])
fres.Add(name, ['LL-wsj', 'PPL-wsj'], [LL_templ, PPL_templ])
fres.AddWER(name, wer)
if '-stat' in sys.argv:
# calculate the mean and var of wers of the intermediate models
inte_wer = []
inte_model = []
# find model
for file_name in os.listdir(os.path.split(write_model)[0]):
file_path = os.path.split(write_model)[0] + os.sep + file_name
if not os.path.isfile(file_path):
continue
if file_name.find(os.path.split(write_model)[1]) == 0 and \
file_path.split('.')[-1] == 'model' and \
file_path.split('.')[-2][0] == 'n':
inte_model.append(file_path)
# compute wer
flog = open(workdir + 'inte_model_wer.log', 'wt')
for file_path in sorted(inte_model):
print(file_path)
t = int(file_path.split('.')[-2][1:])
# lmscore
write_lmscore = os.path.splitext(file_path)[0] + '.lmscore'
config = ' -vocab {} '.format(vocab)
config += ' -read {} '.format(file_path)
config += ' -nbest {} '.format(data()[3])
config += ' -lmscore {0} '.format(write_lmscore)
model.use(config, False)
# wer
[wer, lmscale, acscale] = wb.TuneWER(data()[3], data()[4],
wb.LoadScore(write_lmscore),
wb.LoadScore(data()[5]), np.linspace(0.1, 0.9, 9))
print('t={} wer={}'.format(t, wer))
flog.write('{} \t wer={}\n'.format(file_path, wer))
inte_wer.append([t, wer])
flog.close()
# plot wer
inte_wer = sorted(inte_wer, key=lambda d: d[0])
t_list = [i[0] for i in inte_wer]
wer_list = [i[1] for i in inte_wer]
wer_mean = np.mean(wer_list[-20:])
wer_std = np.std(wer_list[-20:])
print('wer_mean={} wer_std={}'.format(wer_mean, wer_std))
plt.figure()
plt.plot(t_list, wer_list)
plt.xlabel('t')
plt.ylabel('wer')
plt.show()
if '-ais' in sys.argv:
[read_nbest, read_templ, read_acscore] = data()[3:6]
write_templ_id = workdir + os.path.split(read_templ)[1] + '.id'
v = trf.ReadVocab(vocab)
trf.NbestToID(read_templ, write_templ_id, v)
# run asi to calculate the normalization constnts of models
ais_chain = 10
ais_inter = 10000
ais_model = '{}.ais{}_{}.model'.format(write_model, ais_chain, ais_inter)
if not os.path.exists(ais_model):
config = ' -vocab {0} -read {1}.model -write {2}'.format(vocab, write_model, ais_model)
config += ' -norm-method AIS -AIS-chain {} -AIS-inter {} -thread {} '.format(ais_chain, ais_inter, thread)
config += ' -norm-len-max {} '.format(trf.FileMaxLen(read_nbest)-1) # just compute the needed length
model.use(config)
# rescore and compute wer
write_lmscore = os.path.splitext(ais_model)[0] + '.lmscore'
config = ' -vocab {} -read {}'.format(vocab, ais_model)
config += ' -nbest {} -test {} '.format(read_nbest, write_templ_id) # calculate the ppl of test set
config += ' -lmscore {} '.format(write_lmscore)
LL_templ = model.use(config, False)
PPL_templ = wb.LL2PPL(-LL_templ, write_templ_id)
[wer, lmscale, acscale] = wb.TuneWER(read_nbest, read_templ,
write_lmscore, read_acscore, np.linspace(0.1, 0.9, 9))
# calculate the LL of train/valid/test
LL = [0]*3
id_data = [train, valid, test] # are id files
for i in range(3):
config = ' -vocab {} -read {} -test {} '.format(vocab, ais_model, id_data[i])
LL[i] = model.use(config, False)
# write to res file
name = os.path.split(write_model)[1]+":AIS{}-{}".format(ais_chain, ais_inter)
fres.AddLL(name, LL, id_data)
fres.AddWER(name, wer)
fres.Add(name, ['LL-wsj', 'PPL-wsj'], [LL_templ, PPL_templ])
def main():
if len(sys.argv) == 1:
print('\"python run.py -train\" train LSTM\n',
'\"python run.py -rescore\" rescore nbest\n',
'\"python run.py -wer\" compute WER')
run_times = range(0, 1) # for multiple run
bindir = '../../tools/trf/bin/'
workdir = 'trflm/'
fres = wb.FRes('models_ppl.txt')
model = trf.model(bindir, workdir)
class_num = 200
train = workdir + 'train.id'
valid = workdir + 'valid.id'
test = workdir + 'test.id'
vocab = workdir + 'vocab_c{}.list'.format(class_num)
thread = 18
ais_chain = 10
ais_inter = 200000
if '-wer' in sys.argv:
# calculate mean of the WER of 10 TRFs after AIS
res_list = []
for runnum in run_times:
name = 'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run{}.ais{}_{}'.format(
runnum, ais_chain, ais_inter)
res = fres.Get(name)[1:]
if run_times.index(runnum) == 0:
res_list = [[] for i in range(len(res))]
for i in range(len(res)):
res_list[i].append(res[i])
name = 'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.runavg.ais{}_{}'.format(
ais_chain, ais_inter)
head = fres.GetHead()[1:]
for i in range(len(head)):
mean = np.mean(res_list[i])
std = np.std(res_list[i])
fres.Add(name, [head[i]], ['{:.2f}+{:.2f}'.format(mean, std)])
if '-ais' in sys.argv:
for runnum in run_times:
write_model = workdir + 'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run{}'.format(
runnum)
[read_nbest, read_templ, read_acscore] = data()[3:6]
write_templ_id = workdir + os.path.split(read_templ)[1] + '.id'
v = trf.ReadVocab(vocab)
trf.NbestToID(read_templ, write_templ_id, v)
# run asi to calculate the normalization constnts of models
ais_model = '{}.ais{}_{}'.format(write_model, ais_chain, ais_inter)
if not os.path.exists(ais_model + '.model'):
config = ' -vocab {0} -read {1}.model -write {2}.model -log {2}.log'.format(
vocab, write_model, ais_model)
config += ' -norm-method AIS -AIS-chain {} -AIS-inter {} -thread {} '.format(
ais_chain, ais_inter, thread)
config += ' -norm-len-max {} '.format(
trf.FileMaxLen(read_nbest) -
1) # just compute the needed length
model.use(config)
# rescore and compute wer
write_lmscore = ais_model + '.lmscore'
config = ' -vocab {} -read {}.model'.format(vocab, ais_model)
config += ' -nbest {} -test {} '.format(
read_nbest, write_templ_id) # calculate the ppl of test set
config += ' -lmscore {} '.format(write_lmscore)
LL_templ = model.use(config, False)
PPL_templ = wb.LL2PPL(-LL_templ, write_templ_id)
[wer, lmscale, acscale] = wb.TuneWER(read_nbest, read_templ,
write_lmscore, read_acscore,
np.linspace(0.1, 0.9, 9))
# calculate the LL of train/valid/test
LL = [0] * 3
id_data = [train, valid, test] # are id files
for i in range(3):
config = ' -vocab {} -read {}.model -test {} '.format(
vocab, ais_model, id_data[i])
LL[i] = model.use(config, False)
# write to res file
name = os.path.split(ais_model)[-1]
fres.AddLL(name, LL, id_data)
fres.AddWER(name, wer)
fres.Add(name, ['LL-wsj', 'PPL-wsj'], [LL_templ, PPL_templ])
if '-cmp' in sys.argv:
# compare the variance of exp(logz) with the variance of AIS weight
# Load the logz of 10 independent runs
multi_run = 10
logzs = []
for i in range(multi_run):
logz = trf.LoadLogz(
workdir +
'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run0.ais10_20000.run{}.model'
.format(i))
logzs.append(logz[0:33])
mat_logzs = np.matrix(logzs).T
# Load the weight of each length
logws = []
with open(workdir +
'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run0.ais10_20000.log'
) as f:
for line in f:
idx = line.find('logw=')
if idx != -1:
a = line[idx:].split()[1:]
logws.append([float(i) for i in a])
mat_logws = np.matrix(logws)
w_var = mat_var(mat_logws)
z_var = mat_var(mat_logzs)
for i in range(len(w_var)):
rate = np.exp(w_var[i] - z_var[i])
print('len={} w_var={} z_var={} rate={}'.format(
i + 1, w_var[i], z_var[i], rate))
if '-cmp2' in sys.argv:
# compare the logz of AIS and the SAMS
write_model = workdir + 'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run0'
logz_sams = trf.LoadLogz(write_model + '.model')
logz_ais = trf.LoadLogz('{}.ais{}_{}.model'.format(
write_model, ais_chain, ais_inter))
plt.figure()
plt.plot(logz_sams[0:33], 'r-', label='sams')
logz_ais10 = []
for n in range(10):
logz_ais10.append(
trf.LoadLogz('{}.ais10_20000.run{}.model'.format(
write_model, n)))
plt.plot(logz_ais10[-1][0:33], 'g--')
logz_ais_m = [0] * 33
for i in range(33):
for n in range(10):
logz_ais_m[i] += logz_ais10[n][i]
logz_ais_m[i] /= 10
plt.plot(logz_ais_m[0:33], 'r--')
plt.plot(logz_ais[0:33], 'b--', label='ais 10-200K')
#plt.legend()
plt.show()
if '-cmp3' in sys.argv:
trf_model = workdir + 'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run0'
# revise the logz of the trf model to the mean of results of 10 (10-20k) runs
logz_sams = trf.LoadLogz(trf_model + '.model')
logz_ais10 = []
for n in range(10):
logz_ais10.append(
trf.LoadLogz('{}.ais10_20000.run{}.model'.format(trf_model,
n)))
logz_ais_m = [0] * 33
for i in range(33):
for n in range(10):
logz_ais_m[i] += logz_ais10[n][i]
logz_ais_m[i] /= 10
print(logz_ais_m)
ais_model = workdir + 'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run0.ais10_20000.runavg.model'
print('write -> ' + ais_model)
revise_logz(trf_model + '.model', ais_model, logz_ais_m)
# compute WER
print('computer WER')
wer = model.wer(vocab, ais_model, data()[3], data()[4], data()[5])
print('WER={}'.format(wer))
# compute PPL
print('computer PPL')
ppl = model.ppl(vocab, ais_model, data()[4], True)
print('PPL={}'.format(ppl))
# plot the logzs
plt.figure()
for n in range(10):
plt.plot(logz_ais10[n][0:33], 'g-')
plt.plot(logz_ais_m[0:33], 'r', label='ais10-20K-mean')
plt.plot(logz_sams[0:33], 'b', label='sams')
plt.legend()
plt.show()
if '-wer3' in sys.argv:
# smooth zeta
wer_ais = []
wer_smooth = []
ppl_ais = []
ppl_smooth = []
ll_ais = []
ll_smooth = []
for n in range(10):
ais_name = workdir + 'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run{}.ais10_20000'.format(
n)
print(ais_name)
logz_ais = trf.LoadLogz(ais_name + '.model')[0:33]
z = np.polyfit(np.linspace(1, 33, 33), logz_ais, 1)
logz_ais_smooth = z[0] * np.linspace(1, 33, 33) + z[1]
revise_logz(ais_name + '.model', ais_name + '.smooth.model',
logz_ais_smooth.tolist())
print(logz_ais)
print(logz_ais_smooth.tolist())
if n == 0:
logz_sams = trf.LoadLogz(
workdir +
'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run{}.model'.format(n))
logw = load_ais_weight(ais_name + '.log')
plt.figure()
for i in range(len(logw)):
plt.plot((i + 1) * np.ones(len(logw[i])), logw[i], 'k.')
plt.plot(np.linspace(1, 33, 33),
logz_ais,
'r-',
label='standard AIS')
plt.plot(np.linspace(1, 33, 33),
logz_ais_smooth,
'g-',
label='smoothed AIS')
plt.plot(np.linspace(1, 33, 33),
logz_sams[0:33],
'b-',
label='SAMS')
plt.legend()
plt.xlim(1, 33)
plt.xlabel('length')
plt.ylabel('logZ')
plt.show()
wer = model.wer(vocab, ais_name + '.model',
data()[3],
data()[4],
data()[5])
[ppl, LL] = model.ppl(vocab, ais_name + '.model', data()[4], True)
wer_ais.append(wer)
ppl_ais.append(ppl)
ll_ais.append(LL)
fres.Add(
os.path.split(ais_name)[-1], ['WER', 'LL-wsj', 'PPL-wsj'],
[wer, LL, ppl])
wer = model.wer(vocab, ais_name + '.smooth.model',
data()[3],
data()[4],
data()[5])
[ppl, LL] = model.ppl(vocab, ais_name + '.smooth.model',
data()[4], True)
wer_smooth.append(wer)
ppl_smooth.append(ppl)
ll_smooth.append(LL)
fres.Add(
os.path.split(ais_name)[-1] + '.smooth',
['WER', 'LL-wsj', 'PPL-wsj'], [wer, LL, ppl])
for label, d in zip(['WER', 'LL-wsj', 'PPL-wsj'],
[wer_ais, ll_ais, ppl_ais]):
cur_mean = np.mean(d)
cur_std = np.std(d)
fres.Add('trf_c200_g4_w_c_ws_cs_wsh_csh_tied.runavg.ais10_20000',
[label], ['{:.2f}+{:.2f}'.format(cur_mean, cur_std)])
for label, d in zip(['WER', 'LL-wsj', 'PPL-wsj'],
[wer_smooth, ll_smooth, ppl_smooth]):
cur_mean = np.mean(d)
cur_std = np.std(d)
fres.Add(
'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.runavg.ais10_20000.smooth',
[label], ['{:.2f}+{:.2f}'.format(cur_mean, cur_std)])
if '-wer2' in sys.argv:
# perform adjust-AIS and evaluate the WER and PPL
results = []
for n in range(10):
ais_name = workdir + 'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run{}.ais10_20000'.format(
n)
print(ais_name)
logw = load_ais_weight(ais_name + '.log')
logz = [np.mean(a) for a in logw]
revise_logz(ais_name + '.model', ais_name + '.adjust.model', logz)
print(' wer')
wer = model.wer(vocab, ais_name + '.adjust.model',
data()[3],
data()[4],
data()[5])
print(' ppl')
[ppl, LL] = model.ppl(vocab, ais_name + '.adjust.model',
data()[4], True)
fres.Add(
os.path.split(ais_name)[-1] + '.ad',
['WER', 'LL-wsj', 'PPL-wsj'], [wer, LL, ppl])
results.append([wer, LL, ppl])
res_mean = []
res_std = []
for i in range(3):
a = [b[i] for b in results]
res_mean.append(np.mean(a))
res_std.append(np.std(a))
fres.Add('trf_c200_g4_w_c_ws_cs_wsh_csh_tied.runavg.ais10_20000.ad', [
'WER', 'LL-wsj', 'PPL-wsj'
], ['{:.2f}+{:.2f}'.format(res_mean[i], res_std[i]) for i in range(3)])
def main():
if len(sys.argv) == 1:
print('\"python run.py -train\" train LSTM\n',
'\"python run.py -rescore\" rescore nbest\n',
'\"python run.py -wer\" compute WER')
run_times = range(3, 10) # for multiple run
bindir = '../../tools/trf/bin/'
workdir = 'trflm/'
fres = wb.FRes('models_ppl.txt')
model = trf.model(bindir, workdir)
class_num = 200
train = workdir + 'train.id'
valid = workdir + 'valid.id'
test = workdir + 'test.id'
vocab = workdir + 'vocab_c{}.list'.format(class_num)
thread = 8
ais_chain = 10
ais_inter = 20000
if '-wer' in sys.argv:
res_list = []
for runnum in run_times:
name = 'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run0.ais{}_{}.run{}'.format(
ais_chain, ais_inter, runnum)
res = fres.Get(name)[1:]
if run_times.index(runnum) == 0:
res_list = [[] for i in range(len(res))]
for i in range(len(res)):
res_list[i].append(res[i])
name = 'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run0.ais{}_{}.avg'.format(
ais_chain, ais_inter)
head = fres.GetHead()[1:]
for i in range(len(head)):
mean = np.mean(res_list[i])
std = np.std(res_list[i])
fres.Add(name, [head[i]], ['{:.2f}+{:.2f}'.format(mean, std)])
if '-ais' in sys.argv:
for runnum in run_times:
write_model = workdir + 'trf_c200_g4_w_c_ws_cs_wsh_csh_tied.run0'
[read_nbest, read_templ, read_acscore] = data()[3:6]
write_templ_id = workdir + os.path.split(read_templ)[1] + '.id'
v = trf.ReadVocab(vocab)
trf.NbestToID(read_templ, write_templ_id, v)
# run asi to calculate the normalization constnts of models
ais_model = '{}.ais{}_{}.run{}.model'.format(
write_model, ais_chain, ais_inter, runnum)
if not os.path.exists(ais_model):
config = ' -vocab {0} -read {1}.model -write {2}'.format(
vocab, write_model, ais_model)
config += ' -norm-method AIS -AIS-chain {} -AIS-inter {} -thread {} '.format(
ais_chain, ais_inter, thread)
config += ' -norm-len-max {} '.format(
trf.FileMaxLen(read_nbest) -
1) # just compute the needed length
model.use(config)
# rescore and compute wer
write_lmscore = os.path.splitext(ais_model)[0] + '.lmscore'
config = ' -vocab {} -read {}'.format(vocab, ais_model)
config += ' -nbest {} -test {} '.format(
read_nbest, write_templ_id) # calculate the ppl of test set
config += ' -lmscore {} '.format(write_lmscore)
LL_templ = model.use(config, False)
PPL_templ = wb.LL2PPL(-LL_templ, write_templ_id)
[wer, lmscale, acscale] = wb.TuneWER(read_nbest, read_templ,
write_lmscore, read_acscore,
np.linspace(0.1, 0.9, 9))
# calculate the LL of train/valid/test
LL = [0] * 3
id_data = [train, valid, test] # are id files
for i in range(3):
config = ' -vocab {} -read {} -test {} '.format(
vocab, ais_model, id_data[i])
LL[i] = model.use(config, False)
# write to res file
name = os.path.splitext(os.path.split(ais_model)[-1])[0]
fres.AddLL(name, LL, id_data)
fres.AddWER(name, wer)
fres.Add(name, ['LL-wsj', 'PPL-wsj'], [LL_templ, PPL_templ])
def main():
if len(sys.argv) == 1:
print('\"python run.py -train\" train LSTM\n',
'\"python run.py -rescore\" rescore nbest\n',
'\"python run.py -wer\" compute WER')
for tsize in [1]:
bindir = '../../tools/trf/bin/'
tskdir = '{}/'.format(tsize)
workdir = tskdir + 'trflm/'
fres = wb.FRes('result.txt')
model = trf.model(bindir, workdir)
class_num = 200
train = workdir + 'train.id'
valid = workdir + 'valid.id'
test = workdir + 'test.id'
vocab = workdir + 'vocab_c{}.list'.format(class_num)
order = 4
feat = 'g4_w_c_ws_cs_wsh_csh_tied.fs'
#feat = 'g4_w_c_ws_cs_cpw.fs'
maxlen = 100
tmax = 50000
t0 = 2000
minibatch = 100
gamma_lambda = '1000,0'
gamma_zeta = '0,0.6'
reg = 1e-6
thread = 8
write_model = workdir + 'trf_c{}_{}'.format(class_num, feat[0:-3])
write_name = '{}:{}'.format(tsize, os.path.split(write_model)[1])
if '-class' in sys.argv:
# just cluster for each tsks.
model.prepare(
data(tskdir)[0],
data(tskdir)[1],
data(tskdir)[2], class_num)
if '-train' in sys.argv or '-all' in sys.argv:
config = '-vocab {} -train {} -valid {} -test {} '.format(
vocab, train, valid, test)
config += ' -order {} -feat {} '.format(order, feat)
config += ' -len {} '.format(maxlen)
config += ' -write {0}.model -log {0}.log '.format(write_model)
config += ' -t0 {} -iter {}'.format(t0, tmax)
config += ' -gamma-lambda {} -gamma-zeta {}'.format(
gamma_lambda, gamma_zeta)
config += ' -L2 {} '.format(reg)
config += ' -mini-batch {} '.format(minibatch)
config += ' -thread {} '.format(thread)
config += ' -print-per-iter 10 -write-at-iter 10000:10000:{}'.format(
tmax)
model.prepare(
data(tskdir)[0],
data(tskdir)[1],
data(tskdir)[2], class_num)
model.train(config)
# output
LL = model.get_last_value(write_model + '.log')
fres.AddLL(write_name, LL, data(tskdir)[0:3])
if '-plot' in sys.argv:
baseline = fres.Get('{}:KN5'.format(tsize))
trf.PlotLog([write_model], [baseline])
if '-rescore' in sys.argv or '-all' in sys.argv:
config = ' -vocab {} '.format(vocab)
config += ' -read {}.model '.format(write_model)
config += ' -nbest {} '.format(data(tskdir)[3])
config += ' -lmscore {0}.lmscore -lmscore-test-id {0}.test-id '.format(
write_model)
model.use(config)
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 + '.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
write_templ_id = workdir + os.path.split(read_templ)[1] + '.id'
v = trf.ReadVocab(vocab)
trf.NbestToID(read_templ, write_templ_id, v)
config = ' -vocab {} '.format(vocab)
config += ' -read {}.model '.format(write_model)
config += ' -test {} '.format(write_templ_id)
LL_templ = model.use(config)
PPL_templ = wb.LL2PPL(-LL_templ, write_templ_id)
# output the result
fres.Add(write_name, ['LL-wsj', 'PPL-wsj'], [LL_templ, PPL_templ])
fres.AddWER(write_name, wer)
trans = root + 'transcript.txt'
ac = root + '1000best.acscore'
lm = root + '1000best.lmscore'
return data_verfy([nbest, trans, ac, lm])
if __name__ == '__main__':
print(sys.argv)
if len(sys.argv) == 1:
print(' \"python run_rnn.py -train\" train rnn\n '
' \"python run_rnn.py -test\" get the ppl\n '
' \"python run_rnn.py -rescore\" rescore nbest\n'
' \"python run_rnn.py -wer\" compute WER'
)
absdir = os.getcwd() + '/'
fres = wb.FRes('result.txt')
for tsize in [1, 2, 4]:
bindir = '../../tools/rnn/rnnlm-0.3e/'
tskdir = absdir + '{}/'.format(tsize)
workdir = tskdir + 'rnnlm/'
model = rnn.model(bindir, workdir)
hidden = 250
cnum = 1
bptt = 4
write_model = workdir + 'h{}_c{}_bptt{}.rnn'.format(hidden, cnum, bptt)
write_name = '{}:RNN:'.format(tsize) + os.path.split(write_model)[1][0:-4]
if '-train' in sys.argv or '-all' in sys.argv:
if not os.path.exists(write_model):
def main():
if len(sys.argv) == 1:
print('\"python run.py -train\" train LSTM\n',
'\"python run.py -rescore\" rescore nbest\n',
'\"python run.py -wer\" compute WER')
bindir = '../../tools/trf/bin/'
workdir = 'trflm/'
fres = wb.FRes('models_ppl.txt')
model = trf.model(bindir, workdir)
nbest_root = 'data/nbest/'
nbest_type_list = ['nbest_mvdr_single_heq_multi']
class_num = 200
train = workdir + 'train.id'
valid = workdir + 'valid.id'
test = workdir + 'test.id'
vocab = workdir + 'vocab_c{}.list'.format(class_num)
order = 4
feat = 'g4_w_c_ws_cs_wsh_csh_tied.fs'
#feat = 'g4_w_c_ws_cs_wsh_csh.fs'
maxlen = 0
tmax = 20000
t0 = 0
minibatch = 100
gamma_lambda = '3000,0'
gamma_zeta = '0,0.6'
reg = 1e-6
thread = 8
write_model = workdir + 'trf_c{}_{}_2'.format(class_num, feat[0:-3])
if '-train' in sys.argv or '-all' in sys.argv:
config = '-vocab {} -train {} -valid {} -test {} '.format(
vocab, train, valid, test)
config += ' -read {}.model'.format(write_model[0:-2])
config += ' -order {} -feat {} '.format(order, feat)
config += ' -len {} '.format(maxlen)
config += ' -write {0}.model -log {0}.log '.format(write_model)
config += ' -t0 {} -iter {}'.format(t0, tmax)
config += ' -gamma-lambda {} -gamma-zeta {}'.format(
gamma_lambda, gamma_zeta)
config += ' -L2 {} '.format(reg)
config += ' -mini-batch {} '.format(minibatch)
config += ' -thread {} '.format(thread)
config += ' -print-per-iter 10 '
config += ' -write-at-iter [{}:10000:{}]'.format(
tmax - 30000, tmax) # output the intermediate models
model.prepare('data/train', 'data/valid', 'data/valid', class_num)
model.train(config)
if '-plot' in sys.argv:
baseline = fres.Get('KN5')
trf.PlotLog([write_model], [baseline])
if '-rescore' in sys.argv or '-all' in sys.argv:
for nbest_type in nbest_type_list:
nbest_dir = nbest_root + nbest_type + '/'
for tsk in [
'nbestlist_{}_{}'.format(a, b) for a in ['dt05', 'et05']
for b in ['real', 'simu']
]:
write_dir = workdir + nbest_type + '/' + tsk + '/'
wb.mkdir(write_dir)
print('{} : {}'.format(nbest_type, tsk))
print(' write -> {}'.format(write_dir))
write_lmscore = write_dir + os.path.split(write_model)[-1]
# fill the empty lines
process_nbest(nbest_dir + tsk + '/words_text',
write_lmscore + '.nbest')
config = ' -vocab {} '.format(vocab)
config += ' -read {}.model '.format(write_model)
config += ' -nbest {} '.format(write_lmscore + '.nbest')
config += ' -lmscore {0}.lmscore -lmscore-test-id {0}.test-id '.format(
write_lmscore)
model.use(config)
if '-wer' in sys.argv or '-all' in sys.argv:
for nbest_type in nbest_type_list:
nbest_dir = nbest_root + nbest_type + '/'
lmpaths = {
'KN5':
nbest_dir + '<tsk>/lmwt.lmonly',
'RNN':
nbest_dir + '<tsk>/lmwt.rnn',
'LSTM':
'lstm/' + nbest_type + '/<tsk>/lmwt.lstm',
'TRF':
workdir + nbest_type + '/<tsk>/' +
os.path.split(write_model)[-1] + '.lmscore'
}
# 'TRF': nbestdir + '<tsk>/lmwt.trf'}
# lmtypes = ['LSTM', 'KN5', 'RNN', 'TRF', 'RNN+KN5', 'LSTM+KN5', 'RNN+TRF', 'LSTM+TRF']
lmtypes = ['TRF', 'RNN', 'KN5', 'RNN+TRF']
wer_workdir = 'wer/' + nbest_type + '/'
print('wer_workdir = ' + wer_workdir)
wer.wer_all(wer_workdir, nbest_dir, lmpaths, lmtypes)
config = wer.wer_tune(wer_workdir)
wer.wer_print(wer_workdir, config)
lmpaths = {
'KN5': 'ngramlm/5gram.lmscore',
'RNN': 'rnnlm/h250_c1_bptt5.run0.lmscore',
'LSTM': 'lstmlm/h250_dropout0_epoch10.run0.lmscore',
'TRF': 'trflm/trf_c200_g4_w_c_ws_cs_wsh_csh_tied.<run>.lmscore'
}
lmtypes = [
'KN5', 'RNN', 'LSTM', 'TRF', 'RNN+KN5', 'RNN+TRF', 'LSTM+KN5',
'LSTM+TRF'
]
outlog = 'wer.log'
if not os.path.exists(outlog):
wer_all(lmpaths, lmtypes, outlog)
fres = wb.FRes(outlog, True)
lmwers = dict()
with open(outlog, 'rt') as f:
f.readline()
for a in [line.split() for line in f]:
if a[0].find('[all]') != -1:
break
type = a[0].split(':')[0]
wer_vec = lmwers.setdefault(type, [])
wer_vec.append(float(a[1]))
for type in lmtypes:
wer_vec = lmwers[type]
wer_mean = np.mean(wer_vec)
wer_std = np.std(wer_vec)
a = line.lower().split()
f2.write(' '.join(a[skip:]) + '\n')
if __name__ == '__main__':
print(sys.argv)
if len(sys.argv) == 1:
print('\"python run.py -train\" train rnn\n',
'\"python run.py -rescore\" rescore nbest\n',
'\"python run.py -wer\" compute WER')
run_times = range(0, 10) # revise this to control the run times
bindir = '../../tools/rnn/rnnlm-0.3e/'
workdir = 'rnnlm/'
fres = wb.FRes('models_ppl.txt')
model = rnn.model(bindir, workdir)
hidden = 250
cnum = 1
bptt = 5
for runnum in run_times:
write_model = workdir + 'h{}_c{}_bptt{}.run{}.rnn'.format(
hidden, cnum, bptt, runnum) # the write model name
write_name = 'RNN:h{}c{}bptt{}:run{}'.format(
hidden, cnum, bptt, runnum) # the name in res file.
if '-train' in sys.argv or '-all' in sys.argv:
if not os.path.exists(write_model):
config = ' -rnnlm {} '.format(write_model)