def test_script_module():
module = QuartznetModule(list(ascii_lowercase))
module_script = torch.jit.script(module)
x = torch.randn(10, 1337)
out1 = module.predict(x)[0]
out2 = module_script.predict(x)[0]
assert out1 == out2
def test_change_vocab():
module = QuartznetModule(list(ascii_lowercase))
module.change_vocab(["a", "b", "c"])
assert module.hparams.initial_vocab_tokens == ["a", "b", "c"]
# comparing to 10 to account for the 3 initial tokens plus
# the few special tokens automatically added.
assert len(module.text_pipeline.vocab) < 10
assert module.decoder.out_channels < 10
def test_dev_run_train(sample_manifest):
module = QuartznetModule(list(ascii_lowercase))
data = ManifestDatamodule(
train_manifest=sample_manifest,
val_manifest=sample_manifest,
test_manifest=sample_manifest,
num_workers=0,
)
trainer = pl.Trainer(
fast_dev_run=True, logger=None, checkpoint_callback=None, gpus=-1
)
trainer.fit(module, datamodule=data)
def test_expected_prediction_from_pretrained_model():
# Loading the sample file
try:
folder = get_default_cache_folder()
download_url(
"https://github.com/fastaudio/10_Speakers_Sample/raw/76f365de2f4d282ec44450d68f5b88de37b8b7ad/train/f0001_us_f0001_00001.wav",
download_folder=str(folder),
filename="f0001_us_f0001_00001.wav",
resume=True,
)
# Preparing data and model
module = QuartznetModule.load_from_nemo(
checkpoint_name=NemoCheckpoint.QuartzNet5x5LS_En)
audio, sr = torchaudio.load(folder / "f0001_us_f0001_00001.wav")
assert sr == 16000
output = module.predict(audio)
expected = "the world needs opportunities for new leaders and new ideas"
assert output[0].strip() == expected
except HTTPError:
return
lr_unfrozen=4e-3,
betas=[0.8, 0.5])
wandb.init(config=hyperparameter_defaults,
project="thunder-speech",
entity="madeupmasters")
config = wandb.config
full_dm = ManifestDatamodule(
train_manifest="train/data/manifests/prepared_train_manifest.json",
val_manifest="train/data/manifests/prepared_test_manifest.json",
test_manifest="train/data/manifests/prepared_test_manifest.json",
num_workers=24,
bs=config.bs,
)
model = QuartznetModule.load_from_nemo(
checkpoint_name="QuartzNet15x5Base-En", )
labels = [
" ",
"ɑ",
"d",
"m",
"ɛ",
"v",
"ʒ",
"ð",
"ʃ",
"θ",
"g",
"i",
"ŋ",
def test_try_to_load_without_parameters_raises_error():
with pytest.raises(ValueError):
QuartznetModule.load_from_nemo()