def err_per_phone(flags_fname, scores_fname, threshold, text_fname, phn_map,
misp_trend):
flags_dict = {k: v for k, v in kio.read_vec_int_ark(flags_fname)}
text_dict = {k: v for k, v in kio.read_vec_int_ark(text_fname)}
far = np.zeros((len(phn_map), len(threshold)), dtype=np.int32)
frr = np.zeros((len(phn_map), len(threshold)), dtype=np.int32)
cnt_mp = np.zeros(len(phn_map), dtype=np.int32)
cnt_cp = np.zeros(len(phn_map), dtype=np.int32)
for utt_id, scores in kio.read_vec_flt_ark(scores_fname):
flags = flags_dict[utt_id]
phones = text_dict[utt_id]
for i in range(len(flags)):
if flags[i] > 2:
continue
phn = phones[i]
mp = misp_fn[misp_trend](scores[i], threshold)
if flags[i] == 0:
cnt_cp[phn] += 1
frr[phn, mp == True] = frr[phn, mp == True] + 1
elif flags[i] > 0:
cnt_mp[phn] += 1
far[phn, mp == False] = far[phn, mp == False] + 1
cnt_mp = cnt_mp.astype(np.float)
cnt_cp = cnt_cp.astype(np.float)
cnt_mp[cnt_mp == 0.] = float('nan')
cnt_cp[cnt_cp == 0.] = float('nan')
cnt_mp = np.reshape(cnt_mp, (-1, 1))
cnt_cp = np.reshape(cnt_cp, (-1, 1))
return (far / cnt_mp), (frr / cnt_cp)
def testInt32VectorReadWrite(self):
"""
Test read/write for int32 vectors.
"""
# read,
i32_vec = { k:v for k,v in kaldi_io.read_vec_int_ark('tests/data/ali.ark') } # binary,
i32_vec2 = { k:v for k,v in kaldi_io.read_vec_int_ark('tests/data/ali_ascii.ark') } # ascii,
# re-save the data,
with kaldi_io.open_or_fd('tests/data_re-saved/ali.ark','wb') as f:
for k,v in i32_vec.items(): kaldi_io.write_vec_int(f, v, k)
# read and make sure it is the same,
for k,v in kaldi_io.read_vec_int_ark('tests/data_re-saved/ali.ark'):
self.assertTrue(np.array_equal(v,i32_vec[k]), msg="int32 vector same after re-saving")
def get_vad_dict(ipath2vad_ark):
vad_dict = dict()
for uttid, vad_vec in kaldi_io.read_vec_int_ark(ipath2vad_ark):
assert (uttid not in vad_dict
), "Duplicated utterance %s in %s" % (uttid, ipath2vad_ark)
vad_dict[uttid] = vad_vec
return vad_dict
def main():
assert len(sys.argv) == 2, "Improper number of arguments."
ipath2vad_scp = sys.argv[1]
# for debugging
# ipath2vad_scp="ark:copy-vector scp:/home/jerry/research/ap17_olr/lsid/_vad/vad_dev_1s.1.scp ark,t:-|"
# use dict rather than list for further investigation of some specific utterance
# but there is no need to realize this so far
numvoicedframes = dict()
for uttid, vad_vec in kaldi_io.read_vec_int_ark(ipath2vad_scp):
assert (uttid not in numvoicedframes
), "Duplicated utterance %s in %s" % (uttid, ipath2vad_scp)
numvoicedframes[uttid] = (np.sum(vad_vec), vad_vec.shape[0])
# draw histogram for voiced frames
draw_histogram(numvoicedframes, 0,
"distribution of number of voiced frames per utterance")
# draw histogram for all frames
draw_histogram(numvoicedframes, 1,
"distrbution of number of frames per utterance")
# get voiced/whole frames ratio
print("The nvoiced/nwhole ratio is %.4f" % get_ratio(numvoicedframes))
def load_dataset(fea_scp,fea_opts,lab_folder,lab_opts,left,right):
fea= { k:m for k,m in kaldi_io.read_mat_ark('ark:copy-feats scp:'+fea_scp+' ark:- |'+fea_opts) }
lab= { k:v for k,v in kaldi_io.read_vec_int_ark('gunzip -c '+lab_folder+'/ali*.gz | '+lab_opts+' '+lab_folder+'/final.mdl ark:- ark:-|') if k in fea} # Note that I'm copying only the aligments of the loaded fea
fea={k: v for k, v in fea.items() if k in lab} # This way I remove all the features without an aligment (see log file in alidir "Did not Succeded")
count=0
end_snt=0
end_index=[]
snt_name=[]
for k in sorted(fea.keys(), key=lambda k: len(fea[k])):
if count==0:
count=1
fea_conc=fea[k]
lab_conc=lab[k]
end_snt=end_snt+fea[k].shape[0]-left
else:
fea_conc=np.concatenate([fea_conc,fea[k]],axis=0)
lab_conc=np.concatenate([lab_conc,lab[k]],axis=0)
end_snt=end_snt+fea[k].shape[0]
end_index.append(end_snt)
snt_name.append(k)
end_index[-1]=end_index[-1]-right
return [snt_name,fea_conc,lab_conc,end_index]
def load_dataset(fea_scp,fea_opts,lab_folder,lab_opts,left,right):
fea= { k:m for k,m in kaldi_io.read_mat_ark('ark:copy-feats scp:'+fea_scp+' ark:- |'+fea_opts) }
lab= { k:v for k,v in kaldi_io.read_vec_int_ark('gunzip -c '+lab_folder+'/ali*.gz | '+lab_opts+' '+lab_folder+'/final.mdl ark:- ark:-|') if k in fea} # Note that I'm copying only the aligments of the loaded fea
fea={k: v for k, v in fea.items() if k in lab} # This way I remove all the features without an aligment (see log file in alidir "Did not Succeded")
count=0
end_snt=0
end_index=[]
snt_name=[]
for k in sorted(fea.keys(), key=lambda k: len(fea[k])):
if count==0:
count=1
fea_conc=fea[k]
lab_conc=lab[k]
end_snt=end_snt+fea[k].shape[0]-left
else:
fea_conc=np.concatenate([fea_conc,fea[k]],axis=0)
lab_conc=np.concatenate([lab_conc,lab[k]],axis=0)
end_snt=end_snt+fea[k].shape[0]
end_index.append(end_snt)
snt_name.append(k)
end_index[-1]=end_index[-1]-right
return [snt_name,fea_conc,lab_conc,end_index]
def run(config):
# Load the nnet model
nnet = torch.load(config.model, map_location=lambda storage, loc: storage)
model = nnetFeedforward(nnet['feature_dim'] * nnet['num_frames'],
nnet['num_layers'], nnet['hidden_dim'],
nnet['num_classes'])
model.load_state_dict(nnet['model_state_dict'])
# Load alignment
ali_files = [
os.path.join(config.ali_dir, f) for f in listdir(config.ali_dir)
if f.startswith('ali.')
]
pdf_ali_dict = {}
for file in ali_files:
if config.ali_type == "pdf":
pdf_ali_file = "ark:ali-to-pdf {} ark:'gunzip -c {} |' ark:- |".format(
os.path.join(config.ali_dir, "final.mdl"), file)
else:
pdf_ali_file = "ark:ali-to-phones --per-frame {} ark:'gunzip -c {} |' ark:- |".format(
os.path.join(config.ali_dir, "final.mdl"), file)
pdf_ali_dict.update(
{u: d - 1
for u, d in kaldi_io.read_vec_int_ark(pdf_ali_file)})
# Load feature stuff
feats_config = pickle.load(open(config.egs_config, 'rb'))
if feats_config['feat_type']:
feat_type = feats_config['feat_type'].split(',')[0]
trans_path = feats_config['feat_type'].split(',')[1]
if config.override_trans_path is not None:
trans_path = config.override_trans_path
if feat_type == "pca":
cmd = "transform-feats {} scp:{} ark:- |".format(
trans_path, config.scp)
elif feat_type == "cmvn":
cmd = "apply-cmvn {} scp:{} ark:- |".format(trans_path, config.scp)
else:
cmd = config.scp
if feats_config['concat_feats']:
context = feats_config['concat_feats'].split(',')
cmd += " splice-feats --left-context={:s} --right-context={:s} ark:- ark:- |".format(
context[0], context[1])
# Get the posterior
fer_dict = {}
for utt_id, mat in kaldi_io.read_mat_ark(cmd):
out = model(Variable(torch.FloatTensor(mat)))
out = softmax(out[1].data.numpy())
als = pdf_ali_dict[utt_id]
preds = np.argmax(out, axis=1)
err = (float(preds.shape[0]) - float(np.sum(np.equal(
preds, als)))) * 100 / float(preds.shape[0])
fer_dict[utt_id] = err
return fer_dict
def err(flags_fname, scores_fname, threshold, misp_trend):
flags_dict = {k: v for k, v in kio.read_vec_int_ark(flags_fname)}
far = np.zeros_like(threshold, dtype=np.int32)
frr = np.zeros_like(threshold, dtype=np.int32)
cnt_mp = 0
cnt_cp = 0
for uttid, scores in kio.read_vec_flt_ark(scores_fname):
flags = flags_dict[uttid]
for i in range(len(flags)):
if flags[i] > 2:
continue
mp = misp_fn[misp_trend](scores[i], threshold)
if flags[i] == 0:
cnt_cp += 1
# correct[mp == False] = correct[mp == False] + 1
frr[mp == True] += 1
elif flags[i] > 0:
cnt_mp = cnt_mp + 1
# correct[mp == True] = correct[mp == True] + 1
far[mp == False] = far[mp == False] + 1
else:
print("Imposible?!")
return (far / cnt_mp), (frr / cnt_cp)
def read_data(self):
feat_path = os.path.join(self.recipe_dir, 'data', self.dset, 'feats.scp')
if self.dset == 'train':
label_path = os.path.join(self.recipe_dir, 'exp', self.cfg['ali_dir'])
else:
label_path = os.path.join(self.recipe_dir, 'exp', self.cfg['ali_dir'] + '_' + self.dset)
feat_opts = "apply-cmvn --utt2spk=ark:{0} ark:{1} ark:- ark:- |". \
format(os.path.join(self.recipe_dir, 'data', self.dset, 'utt2spk'),
os.path.join(self.recipe_dir, 'data', self.dset,
self.dset + '_cmvn_speaker.ark'))
if self.cfg['feature_deltas']:
feat_opts += " add-deltas --delta-order=2 ark:- ark:- |"
if self.cfg['feature_context']:
feat_opts += " splice-feats --left-context={0} --right-context={0} ark:- ark:- |". \
format(str(self.cfg['feature_context']))
label_opts = 'ali-to-pdf' if self.cfg['task'] == 'classification' else 'ali-to-phones --per-frame'
feats = {k: m for k, m in kaldi_io.read_mat_ark(
'ark:copy-feats scp:{} ark:- | {}'.format(feat_path, feat_opts))}
lab = {k: v for k, v in kaldi_io.read_vec_int_ark(
'gunzip -c {0}/ali*.gz | {1} {0}/final.mdl ark:- ark:-|'.format(label_path, label_opts))
if k in feats}
feats = {k: v for k, v in feats.items() if k in lab}
return feats, lab
def read_to_vec(self, type='int'):
print("run", self.cmd)
if type == 'int':
generator = kaldi_io.read_vec_int_ark(self.cmd)
if type == 'float':
generator = kaldi_io.read_vec_flt_ark(self.cmd)
result = {utt_id: np.array(vec) for utt_id, vec in generator}
return result
def get_labels(ali_dir, ali_type, config):
ali_files = []
all_ali_dirs = ali_dir.split(',')
for ali_dir in all_ali_dirs:
ali_files.extend([
os.path.join(ali_dir, f) for f in listdir(ali_dir)
if f.startswith('ali.')
])
pdf_ali_dict = {}
for file in ali_files:
if ali_type == "pdf":
pdf_ali_file = "ark:ali-to-pdf {} ark:'gunzip -c {} |' ark:- |".format(
os.path.join(ali_dir, "final.mdl"), file)
if config.notruncpad:
pdf_ali_dict.update({
u + ".pt": torch.FloatTensor(d - 0).long()
for u, d in kaldi_io.read_vec_int_ark(pdf_ali_file)
})
else:
pdf_ali_dict.update({
u + ".pt": F.pad(
torch.FloatTensor(d - 0).long(),
(0, config.max_seq_len - d.shape[0]))
for u, d in kaldi_io.read_vec_int_ark(pdf_ali_file)
})
else:
pdf_ali_file = "ark:ali-to-phones --per-frame {} ark:'gunzip -c {} |' ark:- |".format(
os.path.join(ali_dir, "final.mdl"), file)
if config.notruncpad:
pdf_ali_dict.update({
u + ".pt": torch.FloatTensor(d - 1).long()
for u, d in kaldi_io.read_vec_int_ark(pdf_ali_file)
})
else:
pdf_ali_dict.update({
u + ".pt": F.pad(
torch.FloatTensor(d - 1).long(),
(0, config.max_seq_len - d.shape[0]))
for u, d in kaldi_io.read_vec_int_ark(pdf_ali_file)
})
torch.save(pdf_ali_dict, os.path.join(config.save_dir, 'labels.pkl'))
def __init__(self, input_stream, output_stream, flags_stream, phone_map_fname,
tmp_dir, batch_size):
self.phone_map = read_phone_map(phone_map_fname)
input_it = kio.read_mat_ark(input_stream)
output_dict = {k: v for k, v in kio.read_ali_ark(output_stream)}
self.flags_dict = {k: v for k, v in kio.read_vec_int_ark(flags_stream)}
self.flags = None
gi.DataLoader.__init__(self, input_it, output_dict, tmp_dir, batch_size)
def testInt32VectorReadWrite(self):
"""
Test read/write for int32 vectors.
"""
# read,
i32_vec = {
k: v
for k, v in kaldi_io.read_vec_int_ark('tests/data/ali.ark')
} # binary,
i32_vec2 = {
k: v
for k, v in kaldi_io.read_vec_int_ark('tests/data/ali_ascii.ark')
} # ascii,
# re-save the data,
with kaldi_io.open_or_fd('tests/data_re-saved/ali.ark', 'wb') as f:
for k, v in i32_vec.items():
kaldi_io.write_vec_int(f, v, k)
# read and make sure it is the same,
for k, v in kaldi_io.read_vec_int_ark('tests/data_re-saved/ali.ark'):
self.assertTrue(np.array_equal(v, i32_vec[k]),
msg="int32 vector same after re-saving")
def compute_prior(config):
ali_files = [
os.path.join(config.ali_dir, f) for f in listdir(config.ali_dir)
if f.startswith('ali.')
]
p = np.zeros(config.num_classes)
for file in ali_files:
if config.ali_type == "pdf":
pdf_ali_file = "ark:ali-to-pdf {} ark:'gunzip -c {} |' ark:- |".format(
os.path.join(config.ali_dir, "final.mdl"), file)
for key, ali in kaldi_io.read_vec_int_ark(pdf_ali_file):
for x in range(config.num_classes):
p[x] += len(np.where((ali) == x)[0])
else:
pdf_ali_file = "ark:ali-to-phones --per-frame {} ark:'gunzip -c {} |' ark:- |".format(
os.path.join(config.ali_dir, "final.mdl"), file)
for key, ali in kaldi_io.read_vec_int_ark(pdf_ali_file):
for x in range(config.num_classes):
p[x] += len(np.where((ali - 1) == x)[0])
return np.log(p / np.sum(p))
def read_kaldi_lab(kaldi_ali, kaldi_lab_opts):
"""Read labels in kaldi format.
Uses kaldi IO.
Arguments
---------
kaldi_ali : str
Path to directory where kaldi alignents are stored.
kaldi_lab_opts : str
A string that contains the options for reading the kaldi alignments.
Returns
-------
lab : dict
A dictionary contaning the labels.
Note
----
This depends on kaldi-io-for-python. Install it separately.
See: https://github.com/vesis84/kaldi-io-for-python
Example
-------
This example requires kaldi files.
```
lab_folder = '/home/kaldi/egs/TIMIT/s5/exp/dnn4_pretrain-dbn_dnn_ali'
read_kaldi_lab(lab_folder, 'ali-to-pdf')
```
"""
# EXTRA TOOLS
try:
import kaldi_io
except ImportError:
raise ImportError("Could not import kaldi_io. Install it to use this.")
# Reading the Kaldi labels
lab = {
k: v
for k, v in kaldi_io.read_vec_int_ark(
"gunzip -c "
+ kaldi_ali
+ "/ali*.gz | "
+ kaldi_lab_opts
+ " "
+ kaldi_ali
+ "/final.mdl ark:- ark:-|",
)
}
return lab
def run(config, post_dict):
ali_files = [os.path.join(config.ali_dir, f) for f in listdir(config.ali_dir) if f.endswith('.gz')]
fer_dict = {}
for file in ali_files:
pdf_ali_file = "ark:ali-to-pdf {} ark:'gunzip -c {} |' ark:- |".format(
os.path.join(config.ali_dir, "final.mdl"),
file)
pdf_ali_dict = {u: d for u, d in kaldi_io.read_vec_int_ark(pdf_ali_file)}
for key in pdf_ali_dict:
als = pdf_ali_dict[key]
preds = np.argmax(post_dict[key], axis=1)
err = (float(preds.shape[0]) - float(np.sum(np.equal(preds, als)))) * 100 / float(preds.shape[0])
fer_dict[key] = err
return fer_dict
def __init__(self, phone_label=None, feats=None, transform=None):
"""
Args:
phone_label (dict): utt to frame label.
feats (dict): utt to frame features.
transform (callable, optional): Optional transform to be applied
on a sample.
"""
if Phone_cla_Dataset.class_trans_vector == None:
Phone_cla_Dataset.class_trans_vector = np.vectorize(
Phone_cla_Dataset.class_trans)
if Phone_cla_Dataset.maxClassNum == -1:
Phone_cla_Dataset.maxClassNum = max(
list(Data_show.phone2class.values())) + 1
if phone_label == None or feats == None:
self.phone_label = {
u: d
for u, d in kaldi_io.read_vec_int_ark("feats/ali.1.ph")
}
self.feats = {
u: d
for u, d in kaldi_io.read_mat_scp("feats/feats.scp")
}
else:
self.phone_label = phone_label
self.feats = feats
self.feats_list = []
self.phone_label_list = []
self.transform = transform
for utt, feat in feats.items():
if utt in phone_label:
self.feats_list.append(feat)
a = np.zeros(feat.shape[0], int)
for i in range(a.shape[0]):
a[i] = phone_label[utt][(i) // 3]
self.phone_label_list.append(
Phone_cla_Dataset.class_trans_vector(a))
self.feats_nd = np.concatenate(tuple(self.feats_list))
self.phone_label_nd = np.concatenate(tuple(self.phone_label_list))
def get_phoneme_labels(ali_dir):
ali_files = []
all_ali_dirs = ali_dir.split(',')
for ali_dir in all_ali_dirs:
ali_files.extend([
os.path.join(ali_dir, f) for f in os.listdir(ali_dir)
if f.startswith('ali.')
])
pdf_ali_dict = {}
for file in ali_files:
pdf_ali_file = "ark:ali-to-phones --per-frame {} ark:'gunzip -c {} |' ark:- |".format(
os.path.join(ali_dir, "final.mdl"), file)
pdf_ali_dict.update(
{u: d
for u, d in kaldi_io.read_vec_int_ark(pdf_ali_file)})
return pdf_ali_dict
def get_labels(ali_dir, ali_type):
ali_files = [
os.path.join(ali_dir, f) for f in listdir(ali_dir)
if f.startswith('ali.')
]
pdf_ali_dict = {}
for file in ali_files:
if ali_type == "pdf":
pdf_ali_file = "ark:ali-to-pdf {} ark:'gunzip -c {} |' ark:- |".format(
os.path.join(ali_dir, "final.mdl"), file)
else:
pdf_ali_file = "ark:ali-to-phones --per-frame {} ark:'gunzip -c {} |' ark:- |".format(
os.path.join(ali_dir, "final.mdl"), file)
pdf_ali_dict.update(
{u: d - 1
for u, d in kaldi_io.read_vec_int_ark(pdf_ali_file)})
return pdf_ali_dict
def testPipeReadWrite(self):
"""
Test read/write for pipes.
Note: make sure the "os.environ['KALDI_ROOT']" in "kaldi_io/kaldi_io.py" is correct.
"""
# the following line disables 'stderr' forwarding, comment it for DEBUG,
with open("/dev/null","w") as sys.stderr:
# read,
flt_mat4 = { k:m for k,m in kaldi_io.read_mat_ark('ark:copy-feats ark:tests/data/feats.ark ark:- |') }
# write to pipe,
with kaldi_io.open_or_fd('ark:| copy-feats ark:- ark:tests/data_re-saved/mat_pipe.ark','wb') as f:
for k,m in flt_mat4.items(): kaldi_io.write_mat(f, m, k)
# read it again and compare,
for k,m in kaldi_io.read_mat_ark('tests/data_re-saved/mat_pipe.ark'):
self.assertTrue(np.array_equal(m, flt_mat4[k]),"flt. matrix same after read/write via pipe")
# read some other formats from pipe,
i32_vec3 = { k:v for k,v in kaldi_io.read_vec_int_ark('ark:copy-int-vector ark:tests/data/ali.ark ark:- |') }
flt_vec4 = { k:v for k,v in kaldi_io.read_vec_flt_ark('ark:copy-vector ark:tests/data/conf.ark ark:- |') }
def read_custom_feats(self, custom_feats_ark):
feat_path = custom_feats_ark
if self.dset == 'train':
label_path = os.path.join(self.recipe_dir, 'exp', self.cfg['ali_dir'])
else:
label_path = os.path.join(self.recipe_dir, 'exp', self.cfg['ali_dir'] + '_' + self.dset)
feat_opts = ''
if self.cfg['feature_deltas']:
feat_opts += " add-deltas --delta-order=2 ark:- ark:- |"
if self.cfg['feature_context']:
feat_opts += " splice-feats --left-context={0} --right-context={0} ark:- ark:- |". \
format(str(self.cfg['feature_context']))
label_opts = 'ali-to-pdf' if self.cfg['task'] == 'classification' else 'ali-to-phones --per-frame'
feats = {k: m for k, m in kaldi_io.read_mat_ark(
'ark:copy-feats ark:{} ark:- | {}'.format(feat_path, feat_opts))}
lab = {k: v for k, v in kaldi_io.read_vec_int_ark(
'gunzip -c {0}/ali*.gz | {1} {0}/final.mdl ark:- ark:-|'.format(label_path, label_opts))
if k in feats}
feats = {k: v for k, v in feats.items() if k in lab}
return feats, lab
def testPipeReadWrite(self):
"""
Test read/write for pipes.
Note: make sure the "os.environ['KALDI_ROOT']" in "kaldi_io/kaldi_io.py" is correct.
"""
# the following line disables 'stderr' forwarding, comment it for DEBUG,
with open("/dev/null", "w") as sys.stderr:
# read,
flt_mat4 = {
k: m
for k, m in kaldi_io.read_mat_ark(
'ark:copy-feats ark:tests/data/feats.ark ark:- |')
}
# write to pipe,
with kaldi_io.open_or_fd(
'ark:| copy-feats ark:- ark:tests/data_re-saved/mat_pipe.ark',
'wb') as f:
for k, m in flt_mat4.items():
kaldi_io.write_mat(f, m, k)
# read it again and compare,
for k, m in kaldi_io.read_mat_ark(
'tests/data_re-saved/mat_pipe.ark'):
self.assertTrue(np.array_equal(m, flt_mat4[k]),
"flt. matrix same after read/write via pipe")
# read some other formats from pipe,
i32_vec3 = {
k: v
for k, v in kaldi_io.read_vec_int_ark(
'ark:copy-int-vector ark:tests/data/ali.ark ark:- |')
}
flt_vec4 = {
k: v
for k, v in kaldi_io.read_vec_flt_ark(
'ark:copy-vector ark:tests/data/conf.ark ark:- |')
}
def save_smoothed_feats(self):
feat_path = os.path.join(self.recipe_dir, 'data', self.dset, 'feats.scp')
if self.dset == 'train':
label_path = os.path.join(self.recipe_dir, 'exp', self.cfg['ali_dir'])
else:
label_path = os.path.join(self.recipe_dir, 'exp', self.cfg['ali_dir'] + '_' + self.dset)
feat_opts = "apply-cmvn --utt2spk=ark:{0} ark:{1} ark:- ark:- |". \
format(os.path.join(self.recipe_dir, 'data', self.dset, 'utt2spk'),
os.path.join(self.recipe_dir, 'data', self.dset,
self.dset + '_cmvn_speaker.ark'))
label_opts = 'ali-to-pdf' if self.cfg['task'] == 'classification' else 'ali-to-phones --per-frame'
feats = {k: m for k, m in kaldi_io.read_mat_ark(
'ark:copy-feats scp:{} ark:- | {}'.format(feat_path, feat_opts))}
lab = {k: v for k, v in kaldi_io.read_vec_int_ark(
'gunzip -c {0}/ali*.gz | {1} {0}/final.mdl ark:- ark:-|'.format(label_path, label_opts))
if k in feats}
feats = {k: v for k, v in feats.items() if k in lab}
fname = os.path.join(self.recipe_dir, 'data', self.dset, 'smoothed.ark')
f = kaldi_io.open_or_fd(fname, 'wb')
for key in tqdm(feats):
tmp = smooth_acoustic(feats[key])
kaldi_io.write_mat(f, tmp, key)
return fname
root.ark_path = kwargs.pop('ark_path', None)
cmd_gmm_info = lambda: '%s/src/gmmbin/gmm-info --print-args=false %s | grep pdfs' % (
root.kaldi_dir, root.mdl_path)
cmd_ali_pdf = lambda: '%s/src/bin/ali-to-pdf --print-args=false %s "ark:gunzip -c %s|" ark,t:-' % (
root.kaldi_dir, root.mdl_path, root.aligz_path)
show_alignments = lambda: '%s/src/bin/show-alignments --print-args=false %s %s "ark:gunzip -c %s|"' % (
root.kaldi_dir, root.phones_path, root.mdl_path, root.aligz_path)
show_transitions = lambda: '%s/src/bin/show-transitions --print-args=false %s %s' % (
root.kaldi_dir, root.phones_path, root.mdl_path)
copy_int_vector = lambda: '%s/src/bin/copy-int-vector --print-args=false "ark:gunzip -c %s|" ark,t:-' % (
root.kaldi_dir, root.aligz_path)
alignment = lambda: {
k: v
for k, v in kaldi_io.read_vec_int_ark(root.aligz_path)
}
def numpdfs():
numpdfs = 0
with subprocess.Popen(cmd_gmm_info(), stdout=subprocess.PIPE,
shell=True) as proc:
numpdfs = int(proc.stdout.readline().decode().strip().split(' ')[-1])
return numpdfs
def transid2info():
transid2info = {}
with subprocess.Popen(show_transitions(),
stdout=subprocess.PIPE,
#!/usr/bin/env python
import numpy as np
import kaldi_io
print('testing int32-vector i/o')
i32_vec = {k: v
for k, v in kaldi_io.read_vec_int_ark('data/ali.ark')} # binary,
i32_vec2 = {k: v
for k, v in kaldi_io.read_vec_int_ark('data/ali_ascii.ark')
} # ascii,
# - store,
with kaldi_io.open_or_fd('data_re-saved/ali.ark', 'wb') as f:
for k, v in i32_vec.items():
kaldi_io.write_vec_int(f, v, k)
# - read and compare,
for k, v in kaldi_io.read_vec_int_ark('data_re-saved/ali.ark'):
assert (np.array_equal(v, i32_vec[k]))
print('testing float-vector i/o')
flt_vec = {k: v for k, v in kaldi_io.read_vec_flt_scp('data/conf.scp')} # scp,
flt_vec2 = {k: v
for k, v in kaldi_io.read_vec_flt_ark('data/conf.ark')
} # binary-ark,
flt_vec3 = {k: v
for k, v in kaldi_io.read_vec_flt_ark('data/conf_ascii.ark')
} # ascii-ark,
# - store,
with kaldi_io.open_or_fd('data_re-saved/conf.ark', 'wb') as f:
for k, v in flt_vec.items():
kaldi_io.write_vec_flt(f, v, k)
def text_generator(path, name):
generator = kaldi_io.read_vec_int_ark(path)
return prepend_generator(generator, name)
parser.add_argument('eval_dir')
args = parser.parse_args()
idx_to_phn_name_file = os.path.join(args.exp_dir, 'phn_sil_to_idx.txt')
phn_to_idx_file = os.path.join(args.exp_dir, 'phn_sil_to_idx.int')
pfeats_name_to_idx_file = os.path.join(args.exp_dir, 'pfeats_name_to_idx.txt')
out_file = os.path.join(args.eval_dir, 'accuracy.txt')
output = open(out_file, 'w')
# Mappings
idx_to_pfeats_name = read_inv_phone_map(pfeats_name_to_idx_file)
idx_to_phn_name = read_inv_phone_map(idx_to_phn_name_file)
pfeats_map = PFeatsMap(phn_to_idx_file, args.lang)
# Inputs
aligns_it = kio.read_vec_int_ark(args.align)
pfeats_it = kio.read_mat_ark(args.pfeats)
# Counters
phn_cnt = np.zeros(pfeats_map.phn_dim(), dtype=np.int)
phn_correct = np.zeros_like(phn_cnt)
pfeats_correct = np.zeros(pfeats_map.pfeats_dim(), dtype=np.int)
# Evaluate accuracy
for utt_phones, utt_pfeats_real in zip(aligns_it, pfeats_it):
for phone, pfeats_real in zip(utt_phones[1], utt_pfeats_real[1]):
pfeats_real = np.exp(pfeats_real)
if not pfeats_map.is_phn_valid(phone):
continue
pfeats_true = pfeats_map.phn_to_pfeats(phone)
Id2LoglikeAMModel.add_args(parser)
parser.add_argument('ali_dir')
parser.add_argument('--max_pdf_id',
type=int,
default=None,
help="Maximum pdf_id")
args = parser.parse_args()
ali_stretch_model = AliStretchModel.build(args, load_from_cashe=False)
id2ll_model = Id2LoglikeAMModel.build(args, load_from_cashe=False)
logger.info(f"Loading {args.ali_dir}/ali.*.gz")
utt2ali = {
key: ali
for key, ali in tqdm(
kaldi_io.read_vec_int_ark(
f'ark: gunzip -c {args.ali_dir}/ali_pdf.1.gz|'))
}
i = 0
for key, ali in tqdm(utt2ali.items()):
i += 1
ali_stretch_model.add_utts(ali)
logger.info(f"AliStretchModel processed {i} utterances")
ali_stretch_model.compute()
ali_stretch_model.save_to_file()
logger.info(f"Loaded {len(utt2ali)} alis")
logger.info(f"Loading logprobs and train model")
i = 0
for k, m in tqdm(
kaldi_io.read_mat_ark(f'ark: cat {args.ali_dir}/output.1.ark |'),
total=len(utt2ali)):
maxscore = float("-inf")
targetscore = 0
# pdb.set_trace()
# if 'wrong' == wrong_or_correct:
# for uttid in wrong_uttids:
# print(uttid)
# else:
# for uttid in correct_uttids:
# print(uttid)
numvoicedframes_correct = dict()
numvoicedframes_wrong = dict()
for uttid, vad_vec in kaldi_io.read_vec_int_ark(ipath2vad_scp):
assert (uttid not in numvoicedframes_correct
), "Duplicated utterance %s in %s" % (uttid, ipath2vad_scp)
assert (uttid not in numvoicedframes_wrong
), "Duplicated utterance %s in %s" % (uttid, ipath2vad_scp)
if uttid in correct_uttids:
numvoicedframes_correct[uttid] = (np.sum(vad_vec), vad_vec.shape[0])
if uttid in wrong_uttids:
numvoicedframes_wrong[uttid] = (np.sum(vad_vec), vad_vec.shape[0])
draw_histogram(
numvoicedframes_correct, 0,
"distribution of number of voiced frames per utterance(correct utterance)")
print("The nvoiced/nwhole ratio is %.4f" % get_ratio(numvoicedframes_correct))
draw_histogram(
def main():
#assert len(sys.argv) == 2, "Improper number of arguments."
#ipath2vad_scp=sys.argv[1]
# for debugging
#ipath2vad_scp="ark:copy-vector scp:/home/jerry/research/ap17_olr/lsid/_vad/vad_dev_1s.1.scp ark,t:-|"
ipath2train_vad_scp = "ark:copy-vector scp:/home/jerry/research/ap17_olr/lsid/data/train/vad.scp ark,t:- |"
ipath2dev1s_vad_scp = "ark:copy-vector scp:/home/jerry/research/ap17_olr/lsid/data/dev_1s/vad.scp ark,t:- |"
# use dict rather than list for further investigation of some specific utterance
# but there is no need to realize this so far
train_numvoicedframes = dict()
for uttid, vad_vec in kaldi_io.read_vec_int_ark(ipath2train_vad_scp):
assert (
uttid not in train_numvoicedframes
), "Duplicated utterance %s in %s" % (uttid, ipath2train_vad_scp)
train_numvoicedframes[uttid] = (np.sum(vad_vec), vad_vec.shape[0])
dev1s_numvoicedframes = dict()
for uttid, vad_vec in kaldi_io.read_vec_int_ark(ipath2dev1s_vad_scp):
assert (
uttid not in dev1s_numvoicedframes
), "Duplicated utterance %s in %s" % (uttid, ipath2dev1s_vad_scp)
dev1s_numvoicedframes[uttid] = (np.sum(vad_vec), vad_vec.shape[0])
## draw histogram for voiced frames for train
#draw_histogram(train_numvoicedframes, 0, "distribution of number of voiced frames per utterance")
## draw histogram of voiced frames for dev_1s
idx = 0
train_nframe_list = [
train_numvoicedframes[uttid][idx]
for uttid in train_numvoicedframes.keys()
]
train_num_bins = range(0, 1000)
dev1s_nframe_list = [
dev1s_numvoicedframes[uttid][idx]
for uttid in dev1s_numvoicedframes.keys()
]
dev1s_num_bins = range(0, 1000)
#num_bins = 100
title = "distribution of number of frames per utterance"
plt.hist(train_nframe_list, train_num_bins, facecolor='blue', alpha=0.5)
plt.hist(dev1s_nframe_list, dev1s_num_bins, facecolor='red', alpha=0.5)
plt.xlabel("number of frames per utterance")
plt.ylabel("number of utterance")
plt.title(title)
plt.show()
## draw histogram for all frames
#draw_histogram(numvoicedframes, 1, "distrbution of number of frames per utterance")
#pdb.set_trace()
# get voiced/whole frames ratio
print("The nvoiced/nwhole ratio is %.4f" % get_ratio(numvoicedframes))
parser.add_argument('exp_dir')
parser.add_argument('eval_dir')
args = parser.parse_args()
os.makedirs(args.eval_dir, exist_ok=True)
model_fname = os.path.join(args.exp_dir, 'final.h5')
phn_map_fname = os.path.join(args.exp_dir, 'phn_sil_to_idx.int')
score_fname = os.path.join(args.eval_dir, 'score.ark')
score_txt_fname = os.path.join(args.eval_dir, 'score.txt')
score_txt = open(score_txt_fname, 'w')
# Load phone map
phn_map = read_phone_map(phn_map_fname)
# Load kaldi files
flags_dict = {k: v for k, v in kio.read_vec_int_ark(args.flags)}
ali_dict = {k: v for k, v in kio.read_ali_ark(args.phones)}
feats_it = kio.read_mat_ark(args.feats)
# Load classifier model
model = MispModel.load(model_fname)
with open(score_fname, 'wb') as f:
for utt, feats in feats_it:
# Workaround to deal with missing alignments
if utt not in ali_dict:
continue
print(utt, end=' ', file=score_txt)
flags = flags_dict[utt]
ali = ali_dict[utt]
args = parser.parse_args()
# Prepere files
os.makedirs(args.res_dir, exist_ok=True)
feats_fd = sys.stdin.buffer
flags_fname = os.path.join(args.data_dir, 'text_ext_flags')
ali_force_frame_fname = os.path.join(args.cbps_dir,
'force_ali_test/ali_frames.gz')
ali_force_fname = os.path.join(args.cbps_dir, 'force_ali_test/ali_pdf.gz')
score_fname = os.path.join(args.res_dir, 'score.ark')
score_txt_fname = os.path.join(args.res_dir, 'score.txt')
score_txt = open(score_txt_fname, 'w')
cmp_fd = open('test/cmp_gop.txt', 'w')
# Load kaldi files
flags_it = kio.read_vec_int_ark(flags_fname)
ali_force_it = kio.read_ali_ark(ali_force_fname)
ali_force_frm_it = kio.read_ali_ark(ali_force_frame_fname)
feats_it = kio.read_mat_ark(feats_fd)
with open(score_fname, 'wb') as f:
for flags_t, ali_force_t, ali_force_frm_t, feats_t in zip(
flags_it, ali_force_it, ali_force_frm_it, feats_it):
# Unpack each tuple
utt, flags = flags_t
_, ali_force = ali_force_t
_, ali_force_frm = ali_force_frm_t
_, feats = feats_t
# Get only features for corresponding states in alignments
probs_force = hlp.np_pick(feats, ali_force)
def load_dataset(fea_scp,
fea_opts,
lab_folder,
lab_opts,
left,
right,
max_sequence_length,
fea_only=False):
fea = {
k: m
for k, m in kaldi_io.read_mat_ark('ark:copy-feats scp:' + fea_scp +
' ark:- |' + fea_opts)
}
if not fea_only:
lab = {
k: v
for k, v in kaldi_io.read_vec_int_ark('gunzip -c ' + lab_folder +
'/ali*.gz | ' + lab_opts +
' ' + lab_folder +
'/final.mdl ark:- ark:-|')
if k in fea
} # Note that I'm copying only the aligments of the loaded fea
fea = {
k: v
for k, v in fea.items() if k in lab
} # This way I remove all the features without an aligment (see log file in alidir "Did not Succeded")
end_snt = 0
end_index = []
snt_name = []
fea_conc = []
lab_conc = []
tmp = 0
for k in sorted(sorted(fea.keys()), key=lambda k: len(fea[k])):
#####
# If the sequence length is above the threshold, we split it with a minimal length max/4
# If max length = 500, then the split will start at 500 + (500/4) = 625.
# A seq of length 625 will be splitted in one of 500 and one of 125
if (len(fea[k]) > max_sequence_length) and max_sequence_length > 0:
fea_chunked = []
lab_chunked = []
for i in range((len(fea[k]) + max_sequence_length - 1) //
max_sequence_length):
if (len(fea[k][i * max_sequence_length:]) >
max_sequence_length + (max_sequence_length / 4)):
fea_chunked.append(fea[k][i * max_sequence_length:(i + 1) *
max_sequence_length])
if not fea_only:
lab_chunked.append(
lab[k][i * max_sequence_length:(i + 1) *
max_sequence_length])
else:
lab_chunked.append(
np.zeros((fea[k][i * max_sequence_length:(i + 1) *
max_sequence_length].shape[0], )))
else:
fea_chunked.append(fea[k][i * max_sequence_length:])
if not fea_only:
lab_chunked.append(lab[k][i * max_sequence_length:])
else:
lab_chunked.append(
np.zeros(
(fea[k][i * max_sequence_length:].shape[0], )))
break
for j in range(0, len(fea_chunked)):
fea_conc.append(fea_chunked[j])
lab_conc.append(lab_chunked[j])
snt_name.append(k + '_split' + str(j))
else:
fea_conc.append(fea[k])
if not fea_only:
lab_conc.append(lab[k])
else:
lab_conc.append(np.zeros((fea[k].shape[0], )))
snt_name.append(k)
tmp += 1
fea_zipped = zip(fea_conc, lab_conc)
fea_sorted = sorted(fea_zipped, key=lambda x: x[0].shape[0])
fea_conc, lab_conc = zip(*fea_sorted)
for entry in fea_conc:
end_snt = end_snt + entry.shape[0]
end_index.append(end_snt)
fea_conc = np.concatenate(fea_conc)
lab_conc = np.concatenate(lab_conc)
return [snt_name, fea_conc, lab_conc, np.asarray(end_index)]