def wer_fun(model, feat, normalizeBias):
global args
# Tranform the formate of KaldiDict feature data in order to forward network
temp = E.KaldiDict()
utts = feat.utts
with torch.no_grad():
for index, utt in enumerate(utts):
data = torch.Tensor(feat[utt][:, np.newaxis, :])
data = torch.autograd.Variable(data)
if args.gpu >= 0:
data = data.cuda(args.gpu)
out1, out2 = model(data, is_training=False, device=args.gpu)
out = out1.cpu().detach().numpy() - normalizeBias
temp[utt] = out
print("(testing) Forward network {}/{}".format(
index, len(utts)),
end=" " * 20 + '\r')
# Tansform KaldiDict to KaldiArk format
print('(testing) Transform to ark', end=" " * 20 + '\r')
amp = temp.ark
# Decode and obtain lattice
hmm = args.TIMITpath + '/exp/dnn4_pretrain-dbn_dnn_ali_test/final.mdl'
hclg = args.TIMITpath + '/exp/tri3/graph/HCLG.fst'
lexicon = args.TIMITpath + '/exp/tri3/graph/words.txt'
print('(testing) Generate Lattice', end=" " * 20 + '\r')
lattice = E.decode_lattice(amp, hmm, hclg, lexicon, args.minActive,
args.maxActive, args.maxMemory, args.beam,
args.latBeam, args.acwt)
# Change language weight from 1 to 10, get the 1best words.
print('(testing) Get 1-best words', end=" " * 20 + '\r')
outs = lattice.get_1best(lmwt=args.minLmwt,
maxLmwt=args.maxLmwt,
outFile=args.outDir + '/outRaw')
# If reference file is not existed, make it.
phonemap = args.TIMITpath + '/conf/phones.60-48-39.map'
outFilter = args.TIMITpath + '/local/timit_norm_trans.pl -i - -m {} -from 48 -to 39'.format(
phonemap)
if not os.path.isfile(args.outDir + '/test_filt.txt'):
refText = args.TIMITpath + '/data/test/text'
cmd = 'cat {} | {} > {}/test_filt.txt'.format(
refText, outFilter, args.outDir)
(_, _) = E.run_shell_cmd(cmd)
# Score WER and find the smallest one.
print('(testing) Score', end=" " * 20 + '\r')
minWER = None
for k in range(args.minLmwt, args.maxLmwt + 1, 1):
cmd = 'cat {} | {} > {}/test_prediction_filt.txt'.format(
outs[k], outFilter, args.outDir)
(_, _) = E.run_shell_cmd(cmd)
os.remove(outs[k])
score = E.wer('{}/test_filt.txt'.format(args.outDir),
"{}/test_prediction_filt.txt".format(args.outDir),
mode='all')
if minWER == None or score['WER'] < minWER:
minWER = score['WER']
os.remove("{}/test_prediction_filt.txt".format(args.outDir))
return minWER
def wer_fun(model, testFeat, normalizeBias):
global args
# Use decode test data to forward network
temp = E.KaldiDict()
print('(testing) Forward network', end=" " * 20 + '\r')
with chainer.using_config('train', False), chainer.no_backprop_mode():
for utt in testFeat.keys():
data = [cp.array(testFeat[utt], dtype=cp.float32)]
out1, out2 = model(data)
out = F.log_softmax(out1, axis=1)
out.to_cpu()
temp[utt] = out.array - normalizeBias
# Tansform KaldiDict to KaldiArk format
print('(testing) Transform to ark', end=" " * 20 + '\r')
amp = temp.ark
# Decode and obtain lattice
hmm = args.TIMITpath + '/exp/dnn4_pretrain-dbn_dnn_ali_test/final.mdl'
hclg = args.TIMITpath + '/exp/tri3/graph/HCLG.fst'
lexicon = args.TIMITpath + '/exp/tri3/graph/words.txt'
print('(testing) Generate Lattice', end=" " * 20 + '\r')
lattice = E.decode_lattice(amp, hmm, hclg, lexicon, args.minActive,
args.maxActive, args.maxMemory, args.beam,
args.latBeam, args.acwt)
# Change language weight from 1 to 10, get the 1best words.
print('(testing) Get 1-best words', end=" " * 20 + '\r')
outs = lattice.get_1best(lmwt=args.minLmwt,
maxLmwt=args.maxLmwt,
outFile=args.outDir + '/outRaw')
# If reference file is not existed, make it.
phonemap = args.TIMITpath + '/conf/phones.60-48-39.map'
outFilter = args.TIMITpath + '/local/timit_norm_trans.pl -i - -m {} -from 48 -to 39'.format(
phonemap)
if not os.path.isfile(args.outDir + '/test_filt.txt'):
refText = args.TIMITpath + '/data/test/text'
cmd = 'cat {} | {} > {}/test_filt.txt'.format(
refText, outFilter, args.outDir)
(_, _) = E.run_shell_cmd(cmd)
# Score WER and find the smallest one.
print('(testing) Score WER', end=" " * 20 + '\r')
minWER = None
for k in range(args.minLmwt, args.maxLmwt + 1, 1):
cmd = 'cat {} | {} > {}/test_prediction_filt.txt'.format(
outs[k], outFilter, args.outDir)
(_, _) = E.run_shell_cmd(cmd)
os.remove(outs[k])
score = E.wer('{}/test_filt.txt'.format(args.outDir),
"{}/test_prediction_filt.txt".format(args.outDir),
mode='all')
if minWER == None or score['WER'] < minWER:
minWER = score['WER']
os.remove("{}/test_prediction_filt.txt".format(args.outDir))
return minWER
def decode_test(outDimPdf=1968, outDimPho=48):
global args
if args.preModel == '':
raise Exception("Expected Pretrained Model.")
elif not os.path.isfile(args.preModel):
raise Exception("No such file:{}.".format(args.preModel))
print("\n############## Parameters Configure ##############")
# Show configure information and write them to file
def configLog(message, f):
print(message)
f.write(message + '\n')
f = open(args.outDir + '/configure', "w")
configLog(
'Start System Time:{}'.format(
datetime.datetime.now().strftime("%Y-%m-%d %X")), f)
configLog('Host Name:{}'.format(socket.gethostname()), f)
configLog('Fix Random Seed:{}'.format(args.randomSeed), f)
configLog('GPU ID:{}'.format(args.gpu), f)
configLog('Pretrained Model:{}'.format(args.preModel), f)
configLog('Output Folder:{}'.format(args.outDir), f)
configLog('Use CMVN:{}'.format(args.useCMVN), f)
configLog('Splice N Frames:{}'.format(args.splice), f)
configLog('Add N Deltas:{}'.format(args.delta), f)
configLog('Normalize Chunk:{}'.format(args.normalizeChunk), f)
configLog('Normalize AMP:{}'.format(args.normalizeAMP), f)
configLog('Decode Minimum Active:{}'.format(args.minActive), f)
configLog('Decode Maximum Active:{}'.format(args.maxActive), f)
configLog('Decode Maximum Memory:{}'.format(args.maxMemory), f)
configLog('Decode Beam:{}'.format(args.beam), f)
configLog('Decode Lattice Beam:{}'.format(args.latBeam), f)
configLog('Decode Acoustic Weight:{}'.format(args.acwt), f)
configLog('Decode minimum Language Weight:{}'.format(args.minLmwt), f)
configLog('Decode maximum Language Weight:{}'.format(args.maxLmwt), f)
f.close()
print("\n############## Decode Test ##############")
#------------------ STEP 1: Load Pretrained Model ------------------
print('Load Model...')
# Initialize model
featDim = 40
if args.delta > 0:
featDim *= (args.delta + 1)
if args.splice > 0:
featDim *= (2 * args.splice + 1)
model = MLP(featDim, outDimPdf, outDimPho)
chainer.serializers.load_npz(args.preModel, model)
if args.gpu >= 0:
model.to_gpu(args.gpu)
#------------------ STEP 2: Prepare Test Data ------------------
print('Prepare decode test data...')
# Fmllr file
testFilePath = args.TIMITpath + '/data-fmllr-tri3/test/feats.scp'
testFeat = E.load(testFilePath)
# Use CMVN
if args.useCMVN:
testUttSpk = args.TIMITpath + '/data-fmllr-tri3/test/utt2spk'
testCmvnState = args.TIMITpath + '/data-fmllr-tri3/test/cmvn.ark'
testFeat = E.use_cmvn(testFeat, testCmvnState, testUttSpk)
# Add delta
if args.delta > 0:
testFeat = E.add_delta(testFeat, args.delta)
# Splice frames
if args.splice > 0:
testFeat = testFeat.splice(args.splice)
# Transform to array
testFeat = testFeat.array
# Normalize
if args.normalizeChunk:
testFeat = testFeat.normalize()
# Normalize acoustic model output
if args.normalizeAMP:
# Compute pdf counts in order to normalize acoustic model posterior probability.
countFile = args.outDir + '/pdfs_counts.txt'
# Get statistics file
if not os.path.isfile(countFile):
trainAliFile = args.TIMITpath + '/exp/dnn4_pretrain-dbn_dnn_ali/ali.*.gz'
_ = E.analyze_counts(aliFile=trainAliFile, outFile=countFile)
with open(countFile) as f:
line = f.readline().strip().strip("[]").strip()
# Get AMP bias value
counts = np.array(list(map(float, line.split())), dtype=np.float32)
normalizeBias = np.log(counts / np.sum(counts))
else:
normalizeBias = 0
#------------------ STEP 3: Decode ------------------
temp = E.KaldiDict()
print('Compute Test WER: Forward network', end=" " * 20 + '\r')
with chainer.using_config('train', False), chainer.no_backprop_mode():
for utt in testFeat.keys():
data = cp.array(testFeat[utt], dtype=cp.float32)
out1, out2 = model(data)
out = F.log_softmax(out1, axis=1)
out.to_cpu()
temp[utt] = out.array - normalizeBias
# Tansform KaldiDict to KaldiArk format
print('Compute Test WER: Transform to ark', end=" " * 20 + '\r')
amp = temp.ark
# Decode and obtain lattice
hmm = args.TIMITpath + '/exp/dnn4_pretrain-dbn_dnn_ali_test/final.mdl'
hclg = args.TIMITpath + '/exp/tri3/graph/HCLG.fst'
lexicon = args.TIMITpath + '/exp/tri3/graph/words.txt'
print('Compute Test WER: Generate Lattice', end=" " * 20 + '\r')
lattice = E.decode_lattice(amp, hmm, hclg, lexicon, args.minActive,
args.maxActive, args.maxMemory, args.beam,
args.latBeam, args.acwt)
# Change language weight from 1 to 10, get the 1best words.
print('Compute Test WER: Get 1Best', end=" " * 20 + '\r')
outs = lattice.get_1best(lmwt=args.minLmwt,
maxLmwt=args.maxLmwt,
outFile=args.outDir + '/outRaw.txt')
#------------------ STEP 4: Score ------------------
# If reference file is not existed, make it.
phonemap = args.TIMITpath + '/conf/phones.60-48-39.map'
outFilter = args.TIMITpath + '/local/timit_norm_trans.pl -i - -m {} -from 48 -to 39'.format(
phonemap)
if not os.path.isfile(args.outDir + '/test_filt.txt'):
refText = args.TIMITpath + '/data/test/text'
cmd = 'cat {} | {} > {}/test_filt.txt'.format(refText, outFilter,
args.outDir)
(_, _) = E.run_shell_cmd(cmd)
# Score WER and find the smallest one.
print('Compute Test WER: compute WER', end=" " * 20 + '\r')
minWER = (None, None)
for k in range(args.minLmwt, args.maxLmwt + 1, 1):
cmd = 'cat {} | {} > {}/tanslation_{}.txt'.format(
outs[k], outFilter, args.outDir, k)
(_, _) = E.run_shell_cmd(cmd)
os.remove(outs[k])
score = E.wer('{}/test_filt.txt'.format(args.outDir),
"{}/tanslation_{}.txt".format(args.outDir, k),
mode='all')
if minWER[0] == None or score['WER'] < minWER[0]:
minWER = (score['WER'], k)
print("Best WER:{}％ at {}/tanslation_{}.txt".format(
minWER[0], args.outDir, k))
