'posteriors':
posteriorgrams
})
if __name__ == "__main__":
usage = "python scores_ABX.py directory input_hmm [input_dbn dbn_dict]"
if len(sys.argv) != 3 and len(sys.argv) != 5:
print(usage)
sys.exit(-1)
with open(sys.argv[2]) as ihmmf:
n_states, transitions, gmms = parse_hmm(ihmmf)
gmms_ = precompute_det_inv(gmms)
map_states_to_phones = phones_mapping(gmms)
likelihoods_computer = functools.partial(compute_likelihoods, gmms_)
depth_1_computer = None
depth_2_computer = None
dbn = None
if len(sys.argv) == 5:
from DBN_Gaussian_timit import DBN # not Gaussian if no GRBM
with open(sys.argv[3]) as idbnf:
dbn = pickle.load(idbnf)
with open(sys.argv[4]) as idbndtf:
dbn_to_int_to_state_tuple = pickle.load(idbndtf)
dbn_phones_to_states = dbn_to_int_to_state_tuple[0]
depth_1_computer = functools.partial(compute_likelihoods_dbn,
dbn,
depth=1)
def process(ofname, iscpfname, ihmmfname,
ilmfname=None, iwdnetfname=None, unibifname=None,
idbnfname=None, idbndictstuple=None):
with open(ihmmfname) as ihmmf:
n_states, transitions, gmms = parse_hmm(ihmmf)
gmms_ = precompute_det_inv(gmms)
map_states_to_phones = phones_mapping(gmms)
likelihoods_computer = functools.partial(compute_likelihoods, gmms_)
gmm_likelihoods_computer = functools.partial(compute_likelihoods, gmms_) #TODO REMOVE
dbn = None
dbn_to_int_to_state_tuple = None
if idbnfname != None:
with open(idbnfname) as idbnf:
dbn = cPickle.load(idbnf)
with open(idbndictstuple) as idbndtf:
dbn_to_int_to_state_tuple = cPickle.load(idbndtf)
dbn_phones_to_states = dbn_to_int_to_state_tuple[0]
likelihoods_computer = functools.partial(compute_likelihoods_dbn, dbn)
# like that = for GRBM first layer (normalize=True, unit=False)
# TODO correct the normalize/unit to work on full test dataset
if iwdnetfname != None:
with open(iwdnetfname) as iwdnf:
transitions = parse_wdnet(transitions, iwdnf) # parse wordnet
elif ilmfname != None:
with open(ilmfname) as ilmf:
if MATRIX_BIGRAM:
transitions = parse_lm_matrix(transitions, ilmf) # parse bigram LM in matrix format in ilmf
else:
transitions = parse_lm(transitions, ilmf) # parse bigram LM in ARPA-MIT in ilmf
elif unibifname != None: # our own unigram and bigram counts,
# c.f. src/produce_LM.py
with open(unibifname) as ubf:
transitions = initialize_transitions(transitions, ubf,
unigrams_only=UNIGRAMS_ONLY)
else:
# uniform transitions between phones
transitions = initialize_transitions(transitions)
transitions = penalty_scale(transitions,
insertion_penalty=INSERTION_PENALTY, scale_factor=SCALE_FACTOR)
dummy = np.ndarray((2,2)) # to force only 1 compile of Viterbi's C
viterbi(dummy, [None, dummy], {}) # also for this compile's debug purposes
if dbn != None:
input_n_frames_mfcc = dbn.rbm_layers[0].n_visible / 39 # TODO generalize
print "this is a DBN with", input_n_frames_mfcc, "MFCC frames"
input_n_frames_arti = dbn.rbm_layers[1].n_visible / 59 # 60 # TODO generalize
print "this is a DBN with", input_n_frames_arti, "articulatory frames"
input_file_name = 'tmp_input_mocha.npy'
map_input_file_name = 'tmp_map_file_to_start_end_mocha.pickle'
try: # TODO remove?
print "loading concat MFCC from pickled file"
with open(input_file_name) as concat:
all_input = np.load(concat)
with open(map_input_file_name) as map_input:
map_file_to_start_end = cPickle.load(map_input)
except:
print "concatenating MFCC and articulatory files" # TODO parallelize + use np.concatenate
all_input = np.ndarray((0, dbn.rbm_layers[0].n_visible + dbn.rbm_layers[1].n_visible), dtype='float32')
map_file_to_start_end = {}
with open(iscpfname) as iscpf:
for line in iscpf:
cline = clean(line)
start = all_input.shape[0]
# get the 1 framed signals
x_mfcc = htkmfc.open(cline).getall()
with open(cline[:-4] + '_ema.npy') as ema:
x_arti = np.load(ema)[:, 2:]
# compute deltas and deltas deltas for articulatory features
_, x_arti = from_mfcc_ema_to_mfcc_arti_tuple(x_mfcc, x_arti)
# add the adjacent frames
if input_n_frames_mfcc > 1:
x_mfcc = padding(input_n_frames_mfcc, x_mfcc)
if input_n_frames_arti > 1:
x_arti = padding(input_n_frames_arti, x_arti)
# do feature transformations if any
# TODO with mocha_timit_params.json params
# concatenate
x_mfcc_arti = np.concatenate((x_mfcc, x_arti), axis=1)
all_input = np.append(all_input, x_mfcc_arti, axis=0)
map_file_to_start_end[cline] = (start, all_input.shape[0])
with open(input_file_name, 'w') as concat:
np.save(concat, all_input)
with open(map_input_file_name, 'w') as map_input:
cPickle.dump(map_file_to_start_end, map_input)
else: # GMM
all_mfcc = np.ndarray((0, 39), dtype='float32') # TODO generalize
print "computing likelihoods"
if dbn != None: # TODO clean
tmp_likelihoods = likelihoods_computer(all_input)
#mean_dbns = np.mean(tmp_likelihoods, 0)
#tmp_likelihoods *= (mean_gmms / mean_dbns)
print tmp_likelihoods
print tmp_likelihoods.shape
columns_remapping = [dbn_phones_to_states[map_states_to_phones[i]] for i in xrange(tmp_likelihoods.shape[1])]
print columns_remapping
likelihoods = (tmp_likelihoods[:, columns_remapping],
map_file_to_start_end)
print likelihoods[0]
print likelihoods[0].shape
else:
likelihoods = (likelihoods_computer(all_mfcc), map_file_to_start_end)
print "computing viterbi paths"
list_mlf_string = []
with open(iscpfname) as iscpf:
il = InnerLoop(likelihoods,
map_states_to_phones, transitions,
using_bigram=(ilmfname != None
or iwdnetfname != None
or unibifname != None))
p = Pool(cpu_count())
list_mlf_string = p.map(il, iscpf)
with open(ofname, 'w') as of:
of.write('#!MLF!#\n')
for line in list_mlf_string:
of.write(line)
print "written", mfcc_file
scipy.io.savemat(mfcc_file[:-4] + APPEND_NAME, mdict={
'likelihoods': self.likelihoods[0][start:end],
'posteriors': posteriorgrams})
if len(sys.argv) != 3 and len(sys.argv) != 5:
print usage
sys.exit(-1)
with open(sys.argv[2]) as ihmmf:
n_states, transitions, gmms = parse_hmm(ihmmf)
gmms_ = precompute_det_inv(gmms)
map_states_to_phones = phones_mapping(gmms)
likelihoods_computer = functools.partial(compute_likelihoods, gmms_)
dbn = None
if len(sys.argv) == 5:
from DBN_Gaussian_timit import DBN # not Gaussian if no GRBM
with open(sys.argv[3]) as idbnf:
dbn = cPickle.load(idbnf)
with open(sys.argv[4]) as idbndtf:
dbn_to_int_to_state_tuple = cPickle.load(idbndtf)
dbn_phones_to_states = dbn_to_int_to_state_tuple[0]
likelihoods_computer = functools.partial(compute_likelihoods_dbn, dbn)
# TODO bigrams
transitions = initialize_transitions(transitions)
#print transitions
def process(ofname,
iscpfname,
ihmmfname,
ilmfname=None,
iwdnetfname=None,
unibifname=None,
idbnfname=None,
idbndictstuple=None):
with open(ihmmfname) as ihmmf:
n_states, transitions, gmms = parse_hmm(ihmmf)
gmms_ = precompute_det_inv(gmms)
map_states_to_phones = phones_mapping(gmms)
likelihoods_computer = functools.partial(compute_likelihoods, gmms_)
gmm_likelihoods_computer = functools.partial(compute_likelihoods,
gmms_) #TODO REMOVE
dbn = None
dbn_to_int_to_state_tuple = None
if idbnfname != None:
with open(idbnfname) as idbnf:
dbn = cPickle.load(idbnf)
with open(idbndictstuple) as idbndtf:
dbn_to_int_to_state_tuple = cPickle.load(idbndtf)
dbn_phones_to_states = dbn_to_int_to_state_tuple[0]
likelihoods_computer = functools.partial(compute_likelihoods_dbn, dbn)
# like that = for GRBM first layer (normalize=True, unit=False)
# TODO correct the normalize/unit to work on full test dataset
if iwdnetfname != None:
with open(iwdnetfname) as iwdnf:
transitions = parse_wdnet(transitions, iwdnf) # parse wordnet
elif ilmfname != None:
with open(ilmfname) as ilmf:
if MATRIX_BIGRAM:
transitions = parse_lm_matrix(
transitions,
ilmf) # parse bigram LM in matrix format in ilmf
else:
transitions = parse_lm(
transitions, ilmf) # parse bigram LM in ARPA-MIT in ilmf
elif unibifname != None: # our own unigram and bigram counts,
# c.f. src/produce_LM.py
with open(unibifname) as ubf:
transitions = initialize_transitions(transitions,
ubf,
unigrams_only=UNIGRAMS_ONLY)
else:
# uniform transitions between phones
transitions = initialize_transitions(transitions)
transitions = penalty_scale(transitions,
insertion_penalty=INSERTION_PENALTY,
scale_factor=SCALE_FACTOR)
dummy = np.ndarray((2, 2)) # to force only 1 compile of Viterbi's C
viterbi(dummy, [None, dummy], {}) # also for this compile's debug purposes
if dbn != None:
input_n_frames_mfcc = dbn.rbm_layers[
0].n_visible / 39 # TODO generalize
print "this is a DBN with", input_n_frames_mfcc, "MFCC frames"
input_n_frames_arti = dbn.rbm_layers[
1].n_visible / 59 # 60 # TODO generalize
print "this is a DBN with", input_n_frames_arti, "articulatory frames"
input_file_name = 'tmp_input_mocha.npy'
map_input_file_name = 'tmp_map_file_to_start_end_mocha.pickle'
try: # TODO remove?
print "loading concat MFCC from pickled file"
with open(input_file_name) as concat:
all_input = np.load(concat)
with open(map_input_file_name) as map_input:
map_file_to_start_end = cPickle.load(map_input)
except:
print "concatenating MFCC and articulatory files" # TODO parallelize + use np.concatenate
all_input = np.ndarray(
(0, dbn.rbm_layers[0].n_visible + dbn.rbm_layers[1].n_visible),
dtype='float32')
map_file_to_start_end = {}
with open(iscpfname) as iscpf:
for line in iscpf:
cline = clean(line)
start = all_input.shape[0]
# get the 1 framed signals
x_mfcc = htkmfc.open(cline).getall()
with open(cline[:-4] + '_ema.npy') as ema:
x_arti = np.load(ema)[:, 2:]
# compute deltas and deltas deltas for articulatory features
_, x_arti = from_mfcc_ema_to_mfcc_arti_tuple(
x_mfcc, x_arti)
# add the adjacent frames
if input_n_frames_mfcc > 1:
x_mfcc = padding(input_n_frames_mfcc, x_mfcc)
if input_n_frames_arti > 1:
x_arti = padding(input_n_frames_arti, x_arti)
# do feature transformations if any
# TODO with mocha_timit_params.json params
# concatenate
x_mfcc_arti = np.concatenate((x_mfcc, x_arti), axis=1)
all_input = np.append(all_input, x_mfcc_arti, axis=0)
map_file_to_start_end[cline] = (start, all_input.shape[0])
with open(input_file_name, 'w') as concat:
np.save(concat, all_input)
with open(map_input_file_name, 'w') as map_input:
cPickle.dump(map_file_to_start_end, map_input)
else: # GMM
all_mfcc = np.ndarray((0, 39), dtype='float32') # TODO generalize
print "computing likelihoods"
if dbn != None: # TODO clean
tmp_likelihoods = likelihoods_computer(all_input)
#mean_dbns = np.mean(tmp_likelihoods, 0)
#tmp_likelihoods *= (mean_gmms / mean_dbns)
print tmp_likelihoods
print tmp_likelihoods.shape
columns_remapping = [
dbn_phones_to_states[map_states_to_phones[i]]
for i in xrange(tmp_likelihoods.shape[1])
]
print columns_remapping
likelihoods = (tmp_likelihoods[:, columns_remapping],
map_file_to_start_end)
print likelihoods[0]
print likelihoods[0].shape
else:
likelihoods = (likelihoods_computer(all_mfcc), map_file_to_start_end)
print "computing viterbi paths"
list_mlf_string = []
with open(iscpfname) as iscpf:
il = InnerLoop(likelihoods,
map_states_to_phones,
transitions,
using_bigram=(ilmfname != None or iwdnetfname != None
or unibifname != None))
p = Pool(cpu_count())
list_mlf_string = p.map(il, iscpf)
with open(ofname, 'w') as of:
of.write('#!MLF!#\n')
for line in list_mlf_string:
of.write(line)