class LeapWrapper(BaseWrapper):
"""Wrapper around the Leap meta-learner.
Arguments:
model (nn.Module): classifier.
optimizer_cls: optimizer class.
meta_optimizer_cls: meta optimizer class.
optimizer_kwargs (dict): kwargs to pass to optimizer upon construction.
meta_optimizer_kwargs (dict): kwargs to pass to meta optimizer upon
construction.
meta_kwargs (dict): kwargs to pass to meta-learner upon construction.
criterion (func): loss criterion to use.
"""
def __init__(self, model, optimizer_cls, meta_optimizer_cls,
optimizer_kwargs, meta_optimizer_kwargs, meta_kwargs,
criterion):
super(LeapWrapper, self).__init__(criterion, model, optimizer_cls,
optimizer_kwargs)
self.meta = Leap(model, **meta_kwargs)
self.meta_optimizer_cls = \
optim.SGD if meta_optimizer_cls.lower() == 'sgd' else optim.Adam
self.meta_optimizer = self.meta_optimizer_cls(self.meta.parameters(),
**meta_optimizer_kwargs)
def _partial_meta_update(self, l, final):
self.meta.update(l, self.model)
def _final_meta_update(self):
self.meta.normalize()
self.meta_optimizer.step()
self.meta_optimizer.zero_grad()
def run_task(self, task, train, meta_train):
if meta_train:
self.meta.init_task()
if train:
self.meta.to(self.model)
return super(LeapWrapper, self).run_task(task, train, meta_train)
def leap_adapt(model,
source_corpus,
target_corpus,
char2idx,
args,
device,
lang_model_n_words=0):
model = model.to(device)
leap = Leap(model)
meta_optimizer = torch.optim.Adam(leap.parameters(),
lr=args.leap_meta_lr_init)
lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
meta_optimizer,
factor=args.lr_decay,
patience=args.patience,
threshold=args.threshold)
best_score = 3
for meta_epoch in np.arange(args.n_meta_epochs):
source_valid_cosine = []
target_valid_cosine = []
model.train()
for meta_batch in np.arange(args.n_meta_batch):
meta_optimizer.zero_grad()
leap.init_task()
leap.to(model)
inner_optimizer = torch.optim.Adam(model.parameters(),
lr=args.leap_inner_lr_init)
for inner_batch in np.arange(args.n_task_steps):
inner_optimizer.zero_grad()
source_train_contexts, source_train_targets, source_train_vocabs = source_corpus.get_batch(
args.meta_batch_size,
args.n_shot,
char2idx,
device,
fixed=args.fixed_shot)
pred_emb = model.forward(source_train_contexts,
source_train_vocabs)
loss = -nn.functional.cosine_similarity(
pred_emb, source_train_targets).mean()
loss.backward()
leap.update(loss, model)
inner_optimizer.step()
leap.init_task()
leap.to(model)
inner_optimizer = torch.optim.Adam(model.parameters(),
lr=args.leap_inner_lr_init)
for inner_batch in np.arange(args.n_task_steps):
inner_optimizer.zero_grad()
target_train_contexts, target_train_targets, target_train_vocabs = target_corpus.get_batch(
args.meta_batch_size,
args.n_shot,
char2idx,
device,
fixed=args.fixed_shot,
repeat_ctxs=args.meta_repeat_ctxs)
pred_emb = model.forward(target_train_contexts,
target_train_vocabs)
loss = -nn.functional.cosine_similarity(
pred_emb, target_train_targets).mean()
loss.backward()
leap.update(loss, model)
inner_optimizer.step()
leap.normalize()
meta_optimizer.step()
leap.to(model)
model.eval()
with torch.no_grad():
for batch in np.arange(args.n_batch):
source_valid_contexts, source_valid_targets, source_valid_vocabs = source_corpus.get_batch(
args.meta_batch_size,
args.n_shot,
char2idx,
device,
use_valid=True,
fixed=args.fixed_shot)
pred_emb = model.forward(source_valid_contexts,
source_valid_vocabs)
loss = -nn.functional.cosine_similarity(
pred_emb, source_valid_targets).mean()
source_valid_cosine += [loss.cpu().numpy()]
target_valid_contexts, target_valid_targets, target_valid_vocabs = target_corpus.get_batch(
args.meta_batch_size,
args.n_shot,
char2idx,
device,
use_valid=True,
fixed=args.fixed_shot,
repeat_ctxs=args.meta_repeat_ctxs)
pred_emb = model.forward(target_valid_contexts,
target_valid_vocabs)
loss = -nn.functional.cosine_similarity(
pred_emb, target_valid_targets).mean()
target_valid_cosine += [loss.cpu().numpy()]
avg_source_valid, avg_target_valid = np.average(
source_valid_cosine), np.average(target_valid_cosine)
score = avg_target_valid
lr_scheduler.step(score)
print(
f"Average source cosine loss: {avg_source_valid}; Average target cosine loss: {avg_target_valid}"
)
if score < best_score:
best_score = score
torch.save(model.state_dict(),
os.path.join(args.save_dir, 'leap_model.pt'))
if meta_optimizer.param_groups[0]['lr'] < args.leap_lr_early_stop:
print('LR early stop')
break
verbose = 1
##########
meta_steps = 10
leap = Leap(net)
meta_trainer = gluon.Trainer(list(leap.parameters()), 'sgd',
{'learning_rate': 0.0001})
meta_logger = MetaLogger(num_tasks)
log_params = True
##########
for ms in range(meta_steps):
for task in range(num_tasks):
train_data = train_data_all[task]
leap.to(net)
leap.init_task()
trainer = gluon.Trainer(net.collect_params(), 'sgd',
{'learning_rate': 0.0001})
if log_params:
meta_logger.log_initial_params(ms, net)
for e in range(epochs):
cumulative_loss = 0
# inner loop
for i, (data, label) in enumerate(train_data):
data = data.as_in_context(model_ctx)
label = label.as_in_context(model_ctx)
with autograd.record():
output = net(data)