def predict(config):
model = getattr(models, config.model.name).GNNModel(
**dict(config.model.items()))
_create_if_not_exist(config.output_path)
load_model(config.output_path, model)
model.eval()
pred_temp = []
dataset = MAG240M(config.data_dir, seed=123)
evaluator = MAG240MEvaluator()
dataset.prepare_data()
test_iter = DataGenerator(
dataset=dataset,
samples=[160] * len(config.samples),
batch_size=16,
num_workers=config.num_workers,
data_type="test")
for batch in test_iter.generator():
graph_list, x, y = batch
x = paddle.to_tensor(x, dtype='float32')
y = paddle.to_tensor(y, dtype='int64')
graph_list = [(item[0].tensor(), paddle.to_tensor(item[2]))
for item in graph_list]
out = model(graph_list, x)
pred_temp.append(out.numpy())
pred_temp = np.concatenate(pred_temp, axis=0)
y_pred = pred_temp.argmax(axis=-1)
res = {'y_pred': y_pred}
evaluator.save_test_submission(res, 'results')
def test(args, dataset, g, feats, paper_offset):
print('Loading masks and labels...')
valid_idx = torch.LongTensor(dataset.get_idx_split('valid')) + paper_offset
test_idx = torch.LongTensor(dataset.get_idx_split('test')) + paper_offset
label = dataset.paper_label
print('Initializing data loader...')
sampler = dgl.dataloading.MultiLayerNeighborSampler([160, 160])
valid_collator = ExternalNodeCollator(g, valid_idx, sampler, paper_offset,
feats, label)
valid_dataloader = torch.utils.data.DataLoader(
valid_collator.dataset,
batch_size=16,
shuffle=False,
drop_last=False,
collate_fn=valid_collator.collate,
num_workers=2)
test_collator = ExternalNodeCollator(g, test_idx, sampler, paper_offset,
feats, label)
test_dataloader = torch.utils.data.DataLoader(
test_collator.dataset,
batch_size=16,
shuffle=False,
drop_last=False,
collate_fn=test_collator.collate,
num_workers=4)
print('Loading model...')
model = RGAT(dataset.num_paper_features, dataset.num_classes, 1024, 5, 2,
4, 0.5, 'paper').cuda()
model.load_state_dict(torch.load(args.model_path))
model.eval()
correct = total = 0
for i, (input_nodes, output_nodes,
mfgs) in enumerate(tqdm.tqdm(valid_dataloader)):
with torch.no_grad():
mfgs = [g.to('cuda') for g in mfgs]
x = mfgs[0].srcdata['x']
y = mfgs[-1].dstdata['y']
y_hat = model(mfgs, x)
correct += (y_hat.argmax(1) == y).sum().item()
total += y_hat.shape[0]
acc = correct / total
print('Validation accuracy:', acc)
evaluator = MAG240MEvaluator()
y_preds = []
for i, (input_nodes, output_nodes,
mfgs) in enumerate(tqdm.tqdm(test_dataloader)):
with torch.no_grad():
mfgs = [g.to('cuda') for g in mfgs]
x = mfgs[0].srcdata['x']
y = mfgs[-1].dstdata['y']
y_hat = model(mfgs, x)
y_preds.append(y_hat.argmax(1).cpu())
evaluator.save_test_submission({'y_pred': torch.cat(y_preds)},
args.submission_path)
def infer(config, do_eval=False):
if paddle.distributed.get_world_size() > 1:
paddle.distributed.init_parallel_env()
dataset = MAG240M(config)
evaluator = MAG240MEvaluator()
dataset.prepare_data()
model = getattr(models, config.model.name).GNNModel(**dict(config.model.items()))
if paddle.distributed.get_world_size() > 1:
model = paddle.DataParallel(model)
loss_func = F.cross_entropy
_create_if_not_exist(config.output_path)
load_model(config.output_path, model)
if paddle.distributed.get_rank() == 0:
file = 'model_result_temp'
sudo_label = np.memmap(file, dtype=np.float32, mode='w+',
shape=(121751666, 153))
if do_eval:
valid_iter = DataGenerator(
dataset=dataset,
samples=[200] * len(config.samples),
batch_size=64,
num_workers=config.num_workers,
data_type="eval")
r = evaluate(valid_iter, model, loss_func, config, evaluator, dataset)
log.info("finish eval")
test_iter = DataGenerator(
dataset=dataset,
samples=[200] * len(config.samples),
batch_size=64,
num_workers=config.num_workers,
data_type="test")
r = evaluate(test_iter, model, loss_func, config, evaluator, dataset)
log.info("finish test")
def get_result(config, eval_all=False):
dataset = MAG240MDataset(config.data_dir)
evaluator = MAG240MEvaluator()
file = 'model_result_temp'
sudo_label = np.memmap(file,
dtype=np.float32,
mode='r',
shape=(121751666, 153))
file = "ck_result.txt"
wf = open(file, 'a', encoding='utf-8')
label = dataset.all_paper_label
if eval_all:
valid_idx = dataset.get_idx_split('valid')
pred = sudo_label[valid_idx]
save_path = os.path.join(config.valid_path, "all_eval_result")
np.save(save_path, pred)
y_pred = pred.argmax(1)
y_true = label[valid_idx]
valid_acc = evaluator.eval({'y_true': y_true, 'y_pred': y_pred})['acc']
print("all eval result\n")
print(f"valid_acc: {valid_acc}\n")
wf.write("all eval result\n")
wf.write(f"valid_acc: {valid_acc}\n")
else:
valid_path = os.path.join(config.valid_path, config.valid_name)
valid_idx = np.load(valid_path)
test_idx = dataset.get_idx_split('test')
pred = sudo_label[valid_idx]
y_pred = pred.argmax(1)
y_true = label[valid_idx]
valid_acc = evaluator.eval({'y_true': y_true, 'y_pred': y_pred})['acc']
print(f"eval cv {config.valid_name} result\n")
print(f"valid_acc: {valid_acc}\n")
wf.write(f"eval cv {config.valid_name} result\n")
wf.write(f"valid_acc: {valid_acc}\n")
save_path_test = os.path.join(config.valid_path, config.test_name)
pred_test = sudo_label[test_idx]
print(pred_test.shape)
np.save(save_path_test, pred_test)
trainer.fit(model, datamodule=datamodule)
if args.evaluate:
dirs = glob.glob(f'logs/{args.model}/lightning_logs/*')
version = max([int(x.split(os.sep)[-1].split('_')[-1]) for x in dirs])
logdir = f'logs/{args.model}/lightning_logs/version_{version}'
print(f'Evaluating saved model in {logdir}...')
ckpt = glob.glob(f'{logdir}/checkpoints/*')[0]
trainer = Trainer(gpus=args.device, resume_from_checkpoint=ckpt)
model = RGNN.load_from_checkpoint(
checkpoint_path=ckpt, hparams_file=f'{logdir}/hparams.yaml')
datamodule.batch_size = 16
datamodule.sizes = [160] * len(args.sizes) # (Almost) no sampling...
trainer.test(model=model, datamodule=datamodule)
evaluator = MAG240MEvaluator()
loader = datamodule.hidden_test_dataloader()
model.eval()
y_preds = []
for batch in tqdm(loader):
batch = batch.to(int(args.device))
with torch.no_grad():
out = model(batch.x, batch.adjs_t).argmax(dim=-1).cpu()
y_preds.append(out)
res = {'y_pred': torch.cat(y_preds, dim=0)}
evaluator.save_test_submission(res, f'results/{args.model}')
def train(config, do_eval=False):
if paddle.distributed.get_world_size() > 1:
paddle.distributed.init_parallel_env()
dataset = MAG240M(config.data_dir, seed=123)
evaluator = MAG240MEvaluator()
dataset.prepare_data()
train_iter = DataGenerator(dataset=dataset,
samples=config.samples,
batch_size=config.batch_size,
num_workers=config.num_workers,
data_type="train")
valid_iter = DataGenerator(dataset=dataset,
samples=config.samples,
batch_size=config.batch_size,
num_workers=config.num_workers,
data_type="eval")
model = getattr(models,
config.model.name).GNNModel(**dict(config.model.items()))
if paddle.distributed.get_world_size() > 1:
model = paddle.DataParallel(model)
loss_func = F.cross_entropy
opt, lr_scheduler = optim.get_optimizer(
parameters=model.parameters(),
learning_rate=config.lr,
max_steps=config.max_steps,
weight_decay=config.weight_decay,
warmup_proportion=config.warmup_proportion,
clip=config.clip,
use_lr_decay=config.use_lr_decay)
_create_if_not_exist(config.output_path)
load_model(config.output_path, model)
swriter = SummaryWriter(os.path.join(config.output_path, 'log'))
if do_eval and paddle.distributed.get_rank() == 0:
valid_iter = DataGenerator(dataset=dataset,
samples=[160] * len(config.samples),
batch_size=16,
num_workers=config.num_workers,
data_type="eval")
r = evaluate(valid_iter, model, loss_func, config, evaluator, dataset)
log.info(dict(r))
else:
best_valid_acc = -1
for e_id in range(config.epochs):
loss_temp = []
for batch in tqdm.tqdm(train_iter.generator()):
loss = train_step(model, loss_func, batch, dataset)
log.info(loss.numpy()[0])
loss.backward()
opt.step()
opt.clear_gradients()
loss_temp.append(loss.numpy()[0])
if lr_scheduler is not None:
lr_scheduler.step()
loss = np.mean(loss_temp)
log.info("Epoch %s Train Loss: %s" % (e_id, loss))
swriter.add_scalar('loss', loss, e_id)
if e_id >= config.eval_step and e_id % config.eval_per_steps == 0 and \
paddle.distributed.get_rank() == 0:
r = evaluate(valid_iter, model, loss_func, config, evaluator,
dataset)
log.info(dict(r))
for key, value in r.items():
swriter.add_scalar('eval/' + key, value, e_id)
best_valid_acc = max(best_valid_acc, r['acc'])
if best_valid_acc == r['acc']:
save_model(config.output_path, model, e_id, opt,
lr_scheduler)
swriter.close()
