def forward(self, inputs, supervision):
(mfcc, ivec) = inputs
# FIXME supervisions are distributed over devices. support multi host
device = next(self.parameters()).device
device_id = device.index
# supervision = supervisions[device_id]
n_pdf = supervision.n_pdf
n_batch, n_freq, n_time_in = mfcc.shape
# FIXME do not hardcode here
n_time_out = math.floor((n_time_in - (29 - 1) - 1) / 3 + 1)
lf_mmi_pred, xe_pred = self.model(mfcc.to(device), ivec.to(device))
ref_shape = (n_batch, n_pdf, n_time_out)
assert lf_mmi_pred.shape == ref_shape, "{} != {}".format(
lf_mmi_pred.shape, ref_shape)
io.set_kaldi_device(device_id)
loss, results = chain_loss(
lf_mmi_pred,
self.den_graph,
supervision,
l2_regularize=self.args.l2_regularize,
leaky_hmm_coefficient=self.args.leaky_hmm_coefficient,
xent_regularize=self.args.xent_regularize,
xent_input=xe_pred,
kaldi_way=True)
return pad_chain_loss(loss, results, n_batch)
def forward(data, idx):
(mfcc, ivec), supervision = data
n_batch, n_freq, n_time_in = mfcc.shape
# FIXME do not hardcode here
n_time_out = math.floor((n_time_in - (29 - 1) -1) / 3 + 1)
lf_mmi_pred, xe_pred = model(mfcc.cuda(), ivec.cuda())
ref_shape = (n_batch, n_pdf, n_time_out)
assert lf_mmi_pred.shape == ref_shape, "{} != {}".format(lf_mmi_pred.shape, ref_shape)
return chain_loss(lf_mmi_pred, den_graph, supervision,
l2_regularize=args.l2_regularize,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
xent_regularize=args.xent_regularize,
xent_input=xe_pred, kaldi_way=True)
def test_io():
exp_root = "/data/work70/skarita/exp/chime5/kaldi-22fbdd/egs/chime5/s5/"
den_fst_rs = exp_root + "exp/chain_train_worn_u100k_cleaned/tdnn1a_sp/den.fst"
cmd = "nnet3-chain-copy-egs --frame-shift=1 ark:/data/work49/skarita/repos/torchain/cegs.1.ark ark:- | nnet3-chain-shuffle-egs --buffer-size=5000 --srand=0 ark:- ark:- | nnet3-chain-merge-egs --minibatch-size=128,64,32 ark:- ark:-"
with io.open_example(cmd) as example:
idx = example.indexes
print(idx.shape)
print(idx[0])
(mfcc, ivec), sup = example.value()
print(mfcc.shape)
sup = example.supervision
assert sup.n_frame == idx.shape[1]
for use_xent in [True]:
for use_kaldi_way in [True]:
print("xent: ", use_xent, "kaldi: ", use_kaldi_way)
with io.open_example(cmd) as example:
n_pdf = example.supervision.n_pdf
print(n_pdf)
# n_pdf = 2928
den_graph = io.DenominatorGraph(den_fst_rs, n_pdf)
model = Model(n_pdf)
model.cuda()
print(model)
opt = torch.optim.SGD(model.parameters(), lr=1e-6)
count = 0
start = time.time()
for (mfcc, ivec), supervision in example:
x = Variable(mfcc).cuda()
print("input:", x.shape)
pred, xent = model(x)
if not use_xent:
xent = None
loss, results = chain_loss(pred,
den_graph,
supervision,
l2_regularize=0.01,
xent_regularize=0.01,
xent_input=xent,
kaldi_way=use_kaldi_way)
opt.zero_grad()
loss.backward()
opt.step()
print(count, results)
count += 1
if count > 10:
break
elapsed = time.time() - start
print("took: %f" % (elapsed))
def test_rand_io():
# scp_path = "/data/work49/skarita/repos/torch-backup/100.scp" # example/chime5/exp/scp/egs.scp"
scp_path = "/data/work49/skarita/repos/torch-backup/example/chime5/exp/scp/train.scp"
seed = 1
exp_root = "/data/work70/skarita/exp/chime5/kaldi-22fbdd/egs/chime5/s5/"
den_fst_rs = exp_root + "exp/chain_train_worn_u100k_cleaned/tdnn1a_sp/den.fst"
for use_xent in [True]:
for use_kaldi_way in [True]:
print("xent: ", use_xent, "kaldi: ", use_kaldi_way)
print("shuffling")
io.set_kaldi_device()
example = io.RandExample(scp_path, seed, 128)
n_pdf = example.supervision.n_pdf
print(n_pdf)
# n_pdf = 2928
den_graph = io.DenominatorGraph(den_fst_rs, n_pdf)
model = Model(n_pdf)
model.cuda()
print(model)
opt = torch.optim.SGD(model.parameters(), lr=1e-6)
count = 0
start = time.time()
for (mfcc, ivec), supervision in example:
x = Variable(mfcc).cuda()
print("input:", x.shape)
pred, xent = model(x)
if not use_xent:
xent = None
loss, results = chain_loss(pred,
den_graph,
supervision,
l2_regularize=0.01,
xent_regularize=0.01,
xent_input=xent,
kaldi_way=use_kaldi_way)
opt.zero_grad()
loss.backward()
opt.step()
print(count, results)
count += 1
if count > 10:
break
elapsed = time.time() - start
print("took: %f" % (elapsed))