def test_invalid_aux_transducer_loss_enc_layers():
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args(use_aux_transducer_loss=True)
with pytest.raises(ValueError):
E2E(idim, odim, train_args)
train_args = get_default_train_args(
use_aux_transducer_loss=True,
aux_transducer_loss_enc_output_layers="foo")
with pytest.raises(ValueError):
E2E(idim, odim, train_args)
train_args = get_default_train_args(
use_aux_transducer_loss=True,
aux_transducer_loss_enc_output_layers=[0, 4])
with pytest.raises(ValueError):
E2E(idim, odim, train_args)
train_args = get_default_train_args(
use_aux_transducer_loss=True,
use_symm_kl_div_loss=True,
aux_transducer_loss_enc_output_layers=[0],
elayers=3,
etype="blstmp",
subsample="1_2_1",
)
with pytest.raises(ValueError):
E2E(idim, odim, train_args)
def test_no_block_arch():
_, idim, odim, _, _ = get_default_scope_inputs()
args = make_train_args(enc_block_arch=None)
with pytest.raises(ValueError):
E2E(idim, odim, args)
args = make_train_args(dec_block_arch=None)
with pytest.raises(ValueError):
E2E(idim, odim, args)
def test_pytorch_transducer_trainable_and_decodable(train_dic, recog_dic):
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args(**train_dic)
recog_args = get_default_recog_args(**recog_dic)
model = E2E(idim, odim, train_args)
batch = prepare_inputs(idim, odim, ilens, olens)
loss = model(*batch)
loss.backward()
beam_search = BeamSearchTransducer(
decoder=model.dec,
beam_size=recog_args.beam_size,
lm=recog_args.rnnlm,
lm_weight=recog_args.lm_weight,
search_type=recog_args.search_type,
max_sym_exp=recog_args.max_sym_exp,
u_max=recog_args.u_max,
nstep=recog_args.nstep,
prefix_alpha=recog_args.prefix_alpha,
score_norm=recog_args.score_norm_transducer,
)
with torch.no_grad():
in_data = np.random.randn(20, idim)
model.recognize(in_data, beam_search)
def prepare(backend, args):
bs, idim, odim, ilens, olens = get_default_scope_inputs()
n_token = odim - 1
model = E2E(idim, odim, args)
x = torch.randn(bs, max(ilens), idim)
y = (torch.rand(bs, max(olens)) * n_token % n_token).long()
for i in range(bs):
x[i, ilens[i]:] = -1
y[i, olens[i]:] = model.ignore_id
data = []
for i in range(bs):
data.append((
"utt%d" % i,
{
"input": [{
"shape": [ilens[i], idim]
}],
"output": [{
"shape": [olens[i]]
}],
},
))
return model, x, torch.tensor(ilens), y, data
def prepare(args):
bs, idim, odim, ilens, olens = get_default_scope_inputs()
n_token = odim - 1
model = E2E(idim, odim, args)
feats = torch.randn(bs, max(ilens), idim)
labels = (torch.rand(bs, max(olens)) * n_token % n_token).long()
for i in range(bs):
feats[i, ilens[i]:] = -1
labels[i, olens[i]:] = model.ignore_id
data = {}
uttid_list = []
for i in range(bs):
data["utt%d" % i] = {
"input": [{
"shape": [ilens[i], idim]
}],
"output": [{
"shape": [olens[i]]
}],
}
uttid_list.append("utt%d" % i)
return model, feats, torch.tensor(ilens), labels, data, uttid_list
def test_subsampling(train_dic, subsample):
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args(**train_dic)
model = E2E(idim, odim, train_args)
assert model.get_total_subsampling_factor() == subsample
def test_pytorch_transducer_gpu_trainable(backend, trans_type):
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args(trans_type=trans_type)
if trans_type == "warp-rnnt" and torch.version.cuda != "10.0":
with pytest.raises(ImportError):
model = E2E(idim, odim, train_args)
return
model = E2E(idim, odim, train_args)
model.cuda()
batch = prepare_inputs(backend, idim, odim, ilens, olens, is_cuda=True)
loss = model(*batch)
loss.backward()
def test_invalid_aux_task_layer_list():
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args(aux_task_type="default")
with pytest.raises(ValueError):
E2E(idim, odim, train_args)
train_args = get_default_train_args(aux_task_type="default",
aux_task_layer_list="foo")
with pytest.raises(ValueError):
E2E(idim, odim, train_args)
train_args = get_default_train_args(aux_task_type="default",
aux_task_layer_list=[0, 4])
with pytest.raises(ValueError):
E2E(idim, odim, train_args)
def test_calculate_plot_attention():
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args()
model = E2E(idim, odim, train_args)
batch = prepare_inputs(idim, odim, ilens, olens, is_cuda=False)
assert model.calculate_all_attentions(*batch) == []
def test_pytorch_calculate_attentions(atype, backend="pytorch"):
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args(rnnt_mode="rnnt-att", atype=atype)
model = E2E(idim, odim, train_args)
batch = prepare_inputs(backend, idim, odim, ilens, olens, is_cuda=False)
att_ws = model.calculate_all_attentions(*batch)[0]
print(att_ws.shape)
def test_dynamic_quantization(train_dic, recog_dic, quantize_dic):
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args(**train_dic)
recog_args = get_default_recog_args(**recog_dic)
model = E2E(idim, odim, train_args)
if not is_torch_1_5_plus and (
torch.nn.Linear in quantize_dic["mod"]
and quantize_dic["dtype"] == torch.float16
):
# In recognize(...) from asr.py we raise ValueError however
# AssertionError is originaly raised by torch.
with pytest.raises(AssertionError):
model = torch.quantization.quantize_dynamic(
model,
quantize_dic["mod"],
dtype=quantize_dic["dtype"],
)
pytest.skip("Skip rest of the test after checking AssertionError")
else:
model = torch.quantization.quantize_dynamic(
model,
quantize_dic["mod"],
quantize_dic["dtype"],
)
beam_search = BeamSearchTransducer(
decoder=model.dec,
joint_network=model.transducer_tasks.joint_network,
beam_size=recog_args.beam_size,
lm=recog_args.rnnlm,
lm_weight=recog_args.lm_weight,
search_type=recog_args.search_type,
max_sym_exp=recog_args.max_sym_exp,
u_max=recog_args.u_max,
nstep=recog_args.nstep,
prefix_alpha=recog_args.prefix_alpha,
score_norm=recog_args.score_norm_transducer,
quantization=True,
)
with torch.no_grad():
in_data = np.random.randn(20, idim)
if not is_torch_1_4_plus and torch.nn.LSTM in quantize_dic["mod"]:
# Cf. previous comment
with pytest.raises(AssertionError):
model.recognize(in_data, beam_search)
else:
model.recognize(in_data, beam_search)
def test_pytorch_multi_gpu_trainable(backend):
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args()
ngpu = 2
device_ids = list(range(ngpu))
model = E2E(idim, odim, train_args)
model = torch.nn.DataParallel(model, device_ids)
model.cuda()
batch = prepare_inputs(backend, idim, odim, ilens, olens, is_cuda=True)
loss = 1.0 / ngpu * model(*batch)
loss.backward(loss.new_ones(ngpu))
def test_pytorch_transducer_trainable_and_decodable(train_dic, recog_dic):
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args(**train_dic)
recog_args = get_default_recog_args(**recog_dic)
model = E2E(idim, odim, train_args)
batch = prepare_inputs(idim, odim, ilens, olens)
loss = model(*batch)
loss.backward()
with torch.no_grad():
in_data = np.random.randn(20, idim)
model.recognize(in_data, recog_args, train_args.char_list, recog_args.rnnlm)
def test_auxiliary_task(train_dic):
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args(**train_dic)
recog_args = get_default_recog_args()
model = E2E(idim, odim, train_args)
batch = prepare_inputs(idim, odim, ilens, olens)
loss = model(*batch)
loss.backward()
beam_search = BeamSearchTransducer(
decoder=model.dec,
joint_network=model.joint_network,
beam_size=recog_args.beam_size,
lm=recog_args.rnnlm,
lm_weight=recog_args.lm_weight,
search_type=recog_args.search_type,
max_sym_exp=recog_args.max_sym_exp,
u_max=recog_args.u_max,
nstep=recog_args.nstep,
prefix_alpha=recog_args.prefix_alpha,
score_norm=recog_args.score_norm_transducer,
)
tmpdir = tempfile.mkdtemp(prefix="tmp_", dir="/tmp")
torch.save(model.state_dict(), tmpdir + "/model.dummy.best")
with open(tmpdir + "/model.json", "wb") as f:
f.write(
json.dumps(
(idim, odim, vars(train_args)),
indent=4,
ensure_ascii=False,
sort_keys=True,
).encode("utf_8"))
with torch.no_grad():
in_data = np.random.randn(20, idim)
model, _ = load_trained_model(tmpdir + "/model.dummy.best",
training=False)
model.recognize(in_data, beam_search)
def test_pytorch_transducer_trainable_and_decodable(train_dic, recog_dic):
idim, odim, ilens, olens = get_default_scope_inputs()
train_args = get_default_train_args(**train_dic)
recog_args = get_default_recog_args(**recog_dic)
model = E2E(idim, odim, train_args)
batch = prepare_inputs(idim, odim, ilens, olens)
# to avoid huge training time, cer/wer report
# is only enabled at validation steps
if train_args.report_cer or train_args.report_wer:
model.training = False
loss = model(*batch)
loss.backward()
beam_search = BeamSearchTransducer(
decoder=model.dec,
joint_network=model.joint_network,
beam_size=recog_args.beam_size,
lm=recog_args.rnnlm,
lm_weight=recog_args.lm_weight,
search_type=recog_args.search_type,
max_sym_exp=recog_args.max_sym_exp,
u_max=recog_args.u_max,
nstep=recog_args.nstep,
prefix_alpha=recog_args.prefix_alpha,
score_norm=recog_args.score_norm_transducer,
)
with torch.no_grad():
in_data = np.random.randn(20, idim)
model.recognize(in_data, beam_search)
