def __init__(self, dset, conf, save=False):
# Set batches
train, val, test = dset.build_batches("relevant")
# Build model, optimizer, loss and scheduler
model = self.build_model(conf["model"], dset.vocab, dset.char_vocab)
opt = Optimizer(model.parameters(), conf["optim"])
loss = BCE()
lr_sch = LR_scheduler(opt.opt, conf["optim"])
# To track early stopping
self.best = {"val": {"f1": 0}, "test": {}}
step, stop = 0, 0
# For max epochs
for ep in range(conf["train"]["max_epochs"]):
print("\n\tEpoch %d" % ep)
for batch in train:
# set the in training mode.
model.train()
# advance step
step += 1
# forward pass
x, y, mask = self.fw_pass(model, batch)
# measure error
fw_loss = loss(x, y)
# backward pass
opt.train_op(fw_loss)
# validation
if step % conf["train"]["val_steps"] == 0:
# Set the in testing mode
model.eval()
# Eval on val set
val_metrics = utils.bin_fw_eval(model, self.fw_pass, val,
step)
if val_metrics["f1"] > self.best["val"]["f1"]:
# reset Early stop
stop = 0
# Eval on test set
test_metrics = utils.bin_fw_eval(
model, self.fw_pass, test, step)
self.best = {"val": val_metrics, "test": test_metrics}
if save:
model.save(step, conf, self.best, opt, lr_sch,
"bin")
else:
if stop == conf["train"]["patience"]:
return
stop += 1
# maybe update lr
lr_sch.step()
def __init__(self, dset, conf, save=False):
# Set batches
train, val, test = dset.build_batches("tok_tags")
# Build model
model = self.build_model(conf["model"], dset.vocab, dset.char_vocab,
dset.tag_vocab)
opt = Optimizer(model.parameters(), conf["optim"])
if conf["model"]["use_crf"]:
loss = CustomLoss(model.crf)
else:
loss = WeightedCEL(dset.tag_vocab)
lr_sch = LR_scheduler(opt.opt, conf["optim"])
# To track early stopping
self.best = {"val": {"f1": 0}, "test": {}}
step, stop = 0, 0
# Tags to ignore in metrics
ign_tok = [
dset.tag_vocab["<p>"], dset.tag_vocab["<s>"],
dset.tag_vocab["</s>"]
]
for ep in range(conf["train"]["max_epochs"]):
print("\n\tEpoch %d" % ep)
for batch in train:
# set the in training mode.
model.train()
# advance step
step += 1
# forward pass
x, y, mask = self.fw_pass(model, batch)
# measure error
fw_loss = loss(x, y, mask)
# backward pass
opt.train_op(fw_loss)
# validation
if step % conf["train"]["val_steps"] == 0:
# Set the in testing mode
model.eval()
# Eval on val set
val_metrics = utils.ner_fw_eval(model, self.fw_pass, val,
step, ign_tok)
if val_metrics["f1"] > self.best["val"]["f1"]:
# reset Early stop
stop = 0
# Eval on test set
test_metrics = utils.ner_fw_eval(
model, self.fw_pass, test, step, ign_tok)
self.best = {"val": val_metrics, "test": test_metrics}
if save:
model.save(step, conf, self.best, opt, lr_sch,
"ner")
else:
if stop == conf["train"]["patience"]:
return
stop += 1
# maybe update lr
lr_sch.step()