def validate(self, dataset, epoch):
self.model.eval()
total_loss = 0.0
batch_size = self.config.batch_size
indices = torch.randperm(len(dataset))
if self.config.cuda:
indices = indices.cuda()
total_batches = math.floor(len(indices) / batch_size) + 1
batches = list(get_batches(indices, batch_size))
for i, batch in tqdm(enumerate(batches),
desc='Testing epoch ' + str(epoch + 1) + '',
total=total_batches):
X, y = dataset.get_batch(batch)
X, y = Var(X, requires_grad=False), Var(y, requires_grad=False)
loss = self.model.forward_train(X, y)
total_loss += loss.item()
logging.debug('Validation batch {}, loss {}'.format(
i, loss.item()))
total_loss /= len(dataset)
return total_loss
def train(self, dataset, epoch, st_batch=None):
self.model.train()
self.optimizer.zero_grad()
total_loss = 0.0
batch_size = self.config.batch_size
indices = torch.randperm(len(dataset))
if self.config.cuda:
indices = indices.cuda()
total_batches = math.floor(len(indices) / batch_size) + 1
batches = list(get_batches(indices, batch_size))
if st_batch:
batches = batches[st_batch:]
for i, batch in tqdm(enumerate(batches),
desc='Training epoch ' + str(epoch + 1) + '',
total=total_batches):
trees, queries, tgt_node_seq, tgt_par_rule_seq, tgt_par_t_seq, \
tgt_action_seq, tgt_action_seq_type = dataset.get_batch(batch)
#print(trees)
loss = self.model.forward_train(trees, queries, tgt_node_seq,
tgt_action_seq, tgt_par_rule_seq,
tgt_par_t_seq, tgt_action_seq_type)
assert loss > 0, "NLL can not be less than zero"
total_loss += loss.data[0]
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
logging.debug('Batch {}, loss {}'.format(i + 1, loss[0]))
return total_loss / len(dataset)
def train(self, dataset, epoch):
self.model.train()
self.optimizer.zero_grad()
total_loss = 0.0
batch_size = self.config.batch_size
indices = torch.randperm(len(dataset))
if self.config.cuda:
indices = indices.cuda()
total_batches = math.floor(len(indices) / batch_size) + 1
batches = list(get_batches(indices, batch_size))
for i, batch in tqdm(enumerate(batches),
desc='Training epoch ' + str(epoch + 1) + '',
total=total_batches):
X, y = dataset.get_batch(batch)
X, y = Var(X, requires_grad=False), Var(y, requires_grad=False)
loss = self.model.forward_train(X, y)
total_loss += loss.item()
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
logging.debug('Batch {}, loss {}'.format(i + 1, loss.item()))
return total_loss / len(dataset)
def visualize(self, dataset, writer):
self.model.train()
self.optimizer.zero_grad()
batch_size = 2
indices = torch.randperm(len(dataset))
batch = next(get_batches(indices, batch_size))
trees, queries, tgt_node_seq, tgt_par_rule_seq, tgt_par_t_seq, \
tgt_action_seq, tgt_action_seq_type = dataset.get_batch(batch)
loss = self.model.forward_train(trees, queries, tgt_node_seq,
tgt_action_seq, tgt_par_rule_seq,
tgt_par_t_seq, tgt_action_seq_type)
assert loss > 0, "NLL can not be less than zero"
loss.backward()
writer.add_graph(self.model, loss)