def test_dataloader(self):
if self.testset is not None:
return GeoLoader(self.testset, batch_size=1, num_workers=1)
else:
return None
def val_dataloader(self):
if self.valset is not None:
num_val_workers = 0 if ("gpus" in self.hparams.keys() and self.hparams["gpus"] > 0) else 8
return GeoLoader(self.valset, batch_size=1, num_workers=num_val_workers)
else:
return None
def val_dataloader(self):
print(self.valset)
if self.valset is not None:
return GeoLoader(self.valset, batch_size=1, num_workers=1)
else:
return None
def train_dataloader(self):
if self.trainset is not None:
return GeoLoader(self.trainset, batch_size=1, num_workers=8)
else:
return None
"acc_top1": [],
"acc_top5": [],
}
}
ce = nn.CrossEntropyLoss()
batch_sampler = data_torch.get_directory_batch_sampler(train_set, batch_size)
from pprint import pprint as pp
update = 0
for batch in batch_sampler:
net.train()
batch_data = train_set.get_multiple(batch)
goals = list(map(lambda x: x[0]['goal_ids'], batch_data))
tactics = list(map(lambda x: x[1]['tac_id'], batch_data))
geo_loader = GeoLoader(goals, batch_size=batch_size)
tactic_tensor = torch.cat(tactics, 0)
# print(tactic_tensor)
for graphs in geo_loader:
graphs.to(device)
tactic_tensor = tactic_tensor.to(device)
update += 1
optimizer.zero_grad()
out = net(graphs)
loss = ce(out, tactic_tensor)
loss.backward()
optimizer.step()
if update % 10 == 1:
print("*"* 100)
print("Train loss", loss.cpu().item())