def baseline():
train_ds = init_dataset("./data", stage="train", style="fnn")
val_ds = init_dataset("./data", stage="dev", style="fnn")
train_loader = SenseInstanceLoader(train_ds, batch_size=16)
val_loader = SenseInstanceLoader(val_ds, batch_size=16)
criterion = NLLLoss()
model = SimpleFFN(768, train_loader.num_senses())
optimizer = optim.Adam(model.parameters(), lr=0.001)
decoder = AllwordsEmbeddingDecoder()
n_epochs = 20
trainer = SingleEmbeddingTrainer(criterion, optimizer, train_loader,
val_loader, decoder, n_epochs)
best_net = trainer(model)
def confidence_max4():
train_ds = init_dataset("./data", stage="train", style="fnn")
val_ds = init_dataset("./data", stage="dev", style="fnn")
train_loader = SenseInstanceLoader(train_ds, batch_size=16)
val_loader = SenseInstanceLoader(val_ds, batch_size=16)
criterion = ConfidenceLoss4(0.5)
model = SimpleAbstainingFFN(768,
train_loader.num_senses(),
zone_applicant='max_non_abs')
optimizer = optim.Adam(model.parameters(), lr=0.001)
decoder = AllwordsEmbeddingDecoder()
n_epochs = 20
trainer = SingleEmbeddingTrainer(criterion, optimizer, train_loader,
val_loader, decoder, n_epochs)
best_net = trainer(model)
def init_loader(stage, architecture, style, corpus_id, bsz):
data_dir = allwords_data_dir
filename = join(data_dir, 'raganato.json')
sents = SenseTaggedSentences.from_json(filename, corpus_id)
if architecture == "bem":
ds = BEMDataset(sents)
loader = BEMLoader(ds, bsz)
if architecture == 'simple' or architecture == 'abstaining':
vecmgr = DiskBasedVectorManager(
join(join(data_dir, 'vecs'), corpus_id))
ds = SenseInstanceDataset(sents, vecmgr)
if stage == 'train':
if style == 'single':
loader = SenseInstanceLoader(ds, batch_size=bsz)
if style == 'pairwise':
loader = TwinSenseInstanceLoader(ds, batch_size=bsz)
if stage == 'test':
loader = SenseInstanceLoader(ds, batch_size=bsz)
return loader