def train(self):
X_train, y_train, _ = self.load_results_from_result_paths(self.train_paths)
X_val, y_val, _ = self.load_results_from_result_paths(self.val_paths)
base_learner_config = self.parse_config("base:")
param_config = self.parse_config("param:")
# train
base_learner = DecisionTreeRegressor(criterion='friedman_mse', random_state=None, splitter='best',
**base_learner_config)
self.model = NGBRegressor(Dist=Normal, Base=base_learner, Score=LogScore, verbose=True, **param_config)
self.model = self.model.fit(X_train, y_train, X_val=X_val, Y_val=y_val,
early_stopping_rounds=self.model_config["early_stopping_rounds"])
train_pred, var_train = self.model.predict(X_train), None
val_pred, var_val = self.model.predict(X_val), None
# self.save()
fig_train = utils.scatter_plot(np.array(train_pred), np.array(y_train), xlabel='Predicted', ylabel='True',
title='')
fig_train.savefig(os.path.join(self.log_dir, 'pred_vs_true_train.jpg'))
plt.close()
fig_val = utils.scatter_plot(np.array(val_pred), np.array(y_val), xlabel='Predicted', ylabel='True', title='')
fig_val.savefig(os.path.join(self.log_dir, 'pred_vs_true_val.jpg'))
plt.close()
train_metrics = utils.evaluate_metrics(y_train, train_pred, prediction_is_first_arg=False)
valid_metrics = utils.evaluate_metrics(y_val, val_pred, prediction_is_first_arg=False)
logging.info('train metrics: %s', train_metrics)
logging.info('valid metrics: %s', valid_metrics)
return valid_metrics
def train(self):
X_train, y_train, _ = self.load_results_from_result_paths(
self.train_paths)
X_val, y_val, _ = self.load_results_from_result_paths(self.val_paths)
logging.info(
"LGBOOST TRAIN: Careful categoricals not specified in dataset conversion"
)
dtrain = lgb.Dataset(X_train, label=y_train)
dval = lgb.Dataset(X_val, label=y_val)
param_config = self.parse_param_config()
param_config["seed"] = self.seed
self.model = lgb.train(
param_config,
dtrain,
early_stopping_rounds=self.model_config["early_stopping_rounds"],
verbose_eval=1,
valid_sets=[dval])
train_pred, var_train = self.model.predict(X_train), None
val_pred, var_val = self.model.predict(X_val), None
# self.save()
fig_train = utils.scatter_plot(np.array(train_pred),
np.array(y_train),
xlabel='Predicted',
ylabel='True',
title='')
fig_train.savefig(os.path.join(self.log_dir, 'pred_vs_true_train.jpg'))
plt.close()
fig_val = utils.scatter_plot(np.array(val_pred),
np.array(y_val),
xlabel='Predicted',
ylabel='True',
title='')
fig_val.savefig(os.path.join(self.log_dir, 'pred_vs_true_val.jpg'))
plt.close()
train_metrics = utils.evaluate_metrics(y_train,
train_pred,
prediction_is_first_arg=False)
valid_metrics = utils.evaluate_metrics(y_val,
val_pred,
prediction_is_first_arg=False)
logging.info('train metrics: %s', train_metrics)
logging.info('valid metrics: %s', valid_metrics)
return valid_metrics
def train(self):
X_train, y_train, _ = self.load_results_from_result_paths(
self.train_paths)
X_val, y_val, _ = self.load_results_from_result_paths(self.val_paths)
dtrain = xgb.DMatrix(X_train, label=y_train)
dval = xgb.DMatrix(X_val, label=y_val)
param_config = self.parse_param_config()
param_config["seed"] = self.seed
self.model = xgb.train(
param_config,
dtrain,
num_boost_round=self.model_config["param:num_rounds"],
early_stopping_rounds=self.model_config["early_stopping_rounds"],
verbose_eval=1,
evals=[(dval, 'val')])
train_pred, var_train = self.model.predict(dtrain), None
val_pred, var_val = self.model.predict(dval), None
# self.save()
fig_train = utils.scatter_plot(np.array(train_pred),
np.array(y_train),
xlabel='Predicted',
ylabel='True',
title='')
fig_train.savefig(os.path.join(self.log_dir, 'pred_vs_true_train.jpg'))
plt.close()
fig_val = utils.scatter_plot(np.array(val_pred),
np.array(y_val),
xlabel='Predicted',
ylabel='True',
title='')
fig_val.savefig(os.path.join(self.log_dir, 'pred_vs_true_val.jpg'))
plt.close()
train_metrics = utils.evaluate_metrics(y_train,
train_pred,
prediction_is_first_arg=False)
valid_metrics = utils.evaluate_metrics(y_val,
val_pred,
prediction_is_first_arg=False)
logging.info('train metrics: %s', train_metrics)
logging.info('valid metrics: %s', valid_metrics)
return valid_metrics
def evaluate(self, result_paths):
X_test, y_test, _ = self.load_results_from_result_paths(result_paths)
test_pred, var_test = self.model.predict(X_test), None
test_metrics = utils.evaluate_metrics(y_test,
test_pred,
prediction_is_first_arg=False)
return test_metrics, test_pred, y_test
def test(self):
preds = []
targets = []
self.model.eval()
test_queue = self.load_results_from_result_paths(self.test_paths)
for step, graph_batch in enumerate(test_queue):
graph_batch = graph_batch.to(self.device)
if self.model_config['model'] == 'gnn_vs_gae_classifier':
pred_bins, pred = self.model(graph_batch=graph_batch)
else:
pred = self.model(graph_batch=graph_batch)
preds.extend(pred.detach().cpu().numpy() * 100)
targets.extend(graph_batch.y.detach().cpu().numpy())
fig = utils.scatter_plot(np.array(preds), np.array(targets), xlabel='Predicted', ylabel='True', title='')
fig.savefig(os.path.join(self.log_dir, 'pred_vs_true_test.jpg'))
plt.close()
test_results = utils.evaluate_metrics(np.array(targets), np.array(preds), prediction_is_first_arg=False)
logging.info('test metrics %s', test_results)
return test_results
def validate(self):
preds = []
targets = []
self.model.eval()
valid_queue = self.load_results_from_result_paths(self.val_paths)
for step, (arch_path_enc, y_true) in enumerate(valid_queue):
arch_path_enc = arch_path_enc.to(self.device).float()
y_true = y_true.to(self.device).float()
pred = self.model(arch_path_enc)
preds.extend(pred.detach().cpu().numpy() * 100)
targets.extend(y_true.detach().cpu().numpy())
fig = utils.scatter_plot(np.array(preds),
np.array(targets),
xlabel='Predicted',
ylabel='True',
title='')
fig.savefig(os.path.join(self.log_dir, 'pred_vs_true_valid.jpg'))
plt.close()
val_results = utils.evaluate_metrics(np.array(targets),
np.array(preds),
prediction_is_first_arg=False)
logging.info('validation metrics %s', val_results)
return val_results
def validate(self):
X_val, y_val, _ = self.load_results_from_result_paths(self.val_paths)
val_pred, var_val = self.model.predict(X_val), None
valid_metrics = utils.evaluate_metrics(y_val, val_pred, prediction_is_first_arg=False)
logging.info('validation metrics %s', valid_metrics)
return valid_metrics
def train(self):
from nasbench301.surrogate_models import utils
X_train, y_train, _ = self.load_results_from_result_paths(
self.train_paths)
X_val, y_val, _ = self.load_results_from_result_paths(self.val_paths)
self.model.fit(X_train, y_train)
train_pred, var_train = self.model.predict(X_train), None
val_pred, var_val = self.model.predict(X_val), None
#self.save()
fig_train = utils.scatter_plot(np.array(train_pred),
np.array(y_train),
xlabel='Predicted',
ylabel='True',
title='')
fig_train.savefig(os.path.join(self.log_dir, 'pred_vs_true_train.jpg'))
plt.close()
fig_val = utils.scatter_plot(np.array(val_pred),
np.array(y_val),
xlabel='Predicted',
ylabel='True',
title='')
fig_val.savefig(os.path.join(self.log_dir, 'pred_vs_true_val.jpg'))
plt.close()
train_metrics = utils.evaluate_metrics(y_train,
train_pred,
prediction_is_first_arg=False)
valid_metrics = utils.evaluate_metrics(y_val,
val_pred,
prediction_is_first_arg=False)
logging.info('train metrics: %s', train_metrics)
logging.info('valid metrics: %s', valid_metrics)
return valid_metrics
def infer(self, train_queue, valid_queue, model, criterion, optimizer, lr, epoch):
objs = utils.AvgrageMeter()
# VALIDATION
preds = []
targets = []
model.eval()
for step, graph_batch in enumerate(valid_queue):
graph_batch = graph_batch.to(self.device)
if self.model_config['model'] == 'gnn_vs_gae_classifier':
pred_bins, pred = self.model(graph_batch=graph_batch)
criterion = torch.nn.BCELoss()
criterion_2 = torch.nn.MSELoss()
bins = self.create_bins(lower_bound=0,
width=10,
quantity=9)
binned_weights = []
for value in graph_batch.y.cpu().numpy():
bin_index = self.find_bin(value, bins)
binned_weights.append(bin_index)
bins = torch.FloatTensor(binned_weights)
make_one_hot = lambda index: torch.eye(self.model_config['no_bins'])[index.view(-1).long()]
binns_one_hot = make_one_hot(bins).to(self.device)
loss_1 = criterion(pred_bins, binns_one_hot)
loss_2 = criterion_2(pred, self.normalize_data(graph_batch.y))
alpha = self.model_config['classification_loss']
beta = self.model_config['regression_loss']
loss = alpha * loss_1 + beta * loss_2
else:
pred = self.model(graph_batch=graph_batch)
loss = criterion(self.normalize_data(pred), self.normalize_data(graph_batch.y / 100))
preds.extend(pred.detach().cpu().numpy() * 100)
targets.extend(graph_batch.y.detach().cpu().numpy())
n = graph_batch.num_graphs
objs.update(loss.data.item(), n)
if step % self.data_config['report_freq'] == 0:
logging.info('valid %03d %e ', step, objs.avg)
fig = utils.scatter_plot(np.array(preds), np.array(targets), xlabel='Predicted', ylabel='True', title='')
fig.savefig(os.path.join(self.log_dir, 'pred_vs_true_valid_{}.jpg'.format(epoch)))
plt.close()
val_results = utils.evaluate_metrics(np.array(targets), np.array(preds), prediction_is_first_arg=False)
return objs.avg, val_results
def test(self):
X_test, y_test, _ = self.load_results_from_result_paths(self.test_paths)
test_pred, var_test = self.model.predict(X_test), None
fig = utils.scatter_plot(np.array(test_pred), np.array(y_test), xlabel='Predicted', ylabel='True', title='')
fig.savefig(os.path.join(self.log_dir, 'pred_vs_true_test.jpg'))
plt.close()
test_metrics = utils.evaluate_metrics(y_test, test_pred, prediction_is_first_arg=False)
logging.info('test metrics %s', test_metrics)
return test_metrics
def evaluate(self, result_paths):
# Get evaluation data
eval_queue = self.load_results_from_result_paths(result_paths)
preds = []
targets = []
self.model.eval()
for step, graph_batch in enumerate(eval_queue):
graph_batch = graph_batch.to(self.device)
pred = self.model(graph_batch=graph_batch)
preds.extend(pred.detach().cpu().numpy() * 100)
targets.extend(graph_batch.y.detach().cpu().numpy())
test_metrics = utils.evaluate_metrics(np.array(targets), np.array(preds), prediction_is_first_arg=False)
return test_metrics, np.array(preds), np.array(targets)
def evaluate(self, result_paths):
# Get evaluation data
eval_queue = self.load_results_from_result_paths(result_paths)
preds = []
targets = []
self.model.eval()
for step, (arch_path_enc, y_true) in enumerate(eval_queue):
arch_path_enc = arch_path_enc.to(self.device).float()
y_true = y_true.to(self.device).float()
pred = self.model(arch_path_enc)
preds.extend(pred.detach().cpu().numpy() * 100)
targets.extend(y_true.detach().cpu().numpy())
test_metrics = utils.evaluate_metrics(np.array(targets),
np.array(preds),
prediction_is_first_arg=False)
return test_metrics, preds, targets
def infer(self, train_queue, valid_queue, model, criterion, optimizer, lr,
epoch):
objs = utils.AvgrageMeter()
# VALIDATION
preds = []
targets = []
for step, (arch_path_enc, y_true) in enumerate(valid_queue):
arch_path_enc = arch_path_enc.to(self.device).float()
y_true = y_true.to(self.device).float()
pred = self.model(arch_path_enc)
loss = torch.mean(
torch.abs((self.normalize_data(pred) /
self.normalize_data(y_true / 100)) - 1))
preds.extend(pred.detach().cpu().numpy() * 100)
targets.extend(y_true.detach().cpu().numpy())
objs.update(loss.data.item(), len(arch_path_enc))
if step % self.data_config['report_freq'] == 0:
logging.info('valid %03d %e ', step, objs.avg)
fig = utils.scatter_plot(np.array(preds),
np.array(targets),
xlabel='Predicted',
ylabel='True',
title='')
fig.savefig(
os.path.join(self.log_dir,
'pred_vs_true_valid_{}.jpg'.format(epoch)))
plt.close()
val_results = utils.evaluate_metrics(np.array(targets),
np.array(preds),
prediction_is_first_arg=False)
return objs.avg, val_results
def evaluate_ensemble(self, result_paths, apply_noise):
"""Evaluates the metrics on the result paths using ensemble predicitons"""
preds, targets = [], []
# Collect individuals predictions
for member_model in self.ensemble_members:
member_metrics, member_preds, member_targets = member_model.evaluate(
result_paths)
logging.info("==> Eval member metrics: %s", member_metrics)
if len(targets) == 0:
preds.append(member_preds)
targets = member_targets
continue
if np.any((targets - member_targets) > 1e-5):
raise ValueError("Ensemble members have different targets!")
preds.append(member_preds)
means = np.mean(preds, axis=0)
stddevs = np.std(preds, axis=0)
# Apply noise
if apply_noise:
noisy_predictions = [
np.random.normal(loc=mean, scale=stddev, size=1)[0]
for mean, stddev in zip(means, stddevs)
]
ensemble_predictions = noisy_predictions
else:
ensemble_predictions = means
# Evaluate metrics
metrics = utils.evaluate_metrics(targets,
ensemble_predictions,
prediction_is_first_arg=False)
return metrics, ensemble_predictions, stddevs, targets
def train_epoch(self, train_queue, valid_queue, model, criterion, optimizer, lr, epoch):
objs = utils.AvgrageMeter()
# TRAINING
preds = []
targets = []
model.train()
for step, graph_batch in enumerate(train_queue):
graph_batch = graph_batch.to(self.device)
# print(step)
if self.model_config['model'] == 'gnn_vs_gae_classifier':
pred_bins, pred = self.model(graph_batch=graph_batch)
criterion = torch.nn.BCELoss()
criterion_2 = torch.nn.MSELoss()
bins = self.create_bins(lower_bound=0,
width=10,
quantity=9)
binned_weights = []
for value in graph_batch.y.cpu().numpy():
bin_index = self.find_bin(value, bins)
binned_weights.append(bin_index)
bins = torch.FloatTensor(binned_weights)
make_one_hot = lambda index: torch.eye(self.model_config['no_bins'])[index.view(-1).long()]
binns_one_hot = make_one_hot(bins).to(self.device)
loss_1 = criterion(pred_bins, binns_one_hot)
loss_2 = criterion_2(pred, self.normalize_data(graph_batch.y))
alpha = self.model_config['classification_loss']
beta = self.model_config['regression_loss']
loss = alpha * loss_1 + beta * loss_2
else:
pred = self.model(graph_batch=graph_batch)
if self.model_config['loss:loss_log_transform']:
loss = criterion(self.normalize_data(pred), self.normalize_data(graph_batch.y / 100))
else:
loss = criterion(pred, graph_batch.y / 100)
if self.model_config['loss:pairwise_ranking_loss']:
m = 0.1
'''
y = list(map(lambda y_i: 1 if y_i == True else -1, graph_batch.y[0: -1] > graph_batch.y[1:]))
pairwise_ranking_loss = torch.nn.HingeEmbeddingLoss(margin=m)(pred[0:-1] - pred[1:],
target=torch.from_numpy(np.array(y)))
'''
pairwise_ranking_loss = []
sort_idx = torch.argsort(graph_batch.y, descending=True)
for idx, idx_y_i in enumerate(sort_idx):
for idx_y_i_p1 in sort_idx[idx + 1:]:
pairwise_ranking_loss.append(torch.max(torch.tensor(0.0, dtype=torch.float),
m - (pred[idx_y_i] - pred[idx_y_i_p1])))
pairwise_ranking_loss = torch.mean(torch.stack(pairwise_ranking_loss))
loss += pairwise_ranking_loss
if step % self.data_config['report_freq'] == 0:
logging.info('Pairwise ranking loss {}'.format(pairwise_ranking_loss))
preds.extend(pred.detach().cpu().numpy() * 100)
targets.extend(graph_batch.y.detach().cpu().numpy())
optimizer.zero_grad()
loss.backward()
optimizer.step()
n = graph_batch.num_graphs
objs.update(loss.data.item(), n)
if step % self.data_config['report_freq'] == 0:
logging.info('train %03d %e', step, objs.avg)
fig = utils.scatter_plot(np.array(preds), np.array(targets), xlabel='Predicted', ylabel='True', title='')
fig.savefig(os.path.join(self.log_dir, 'pred_vs_true_train_{}.jpg'.format(epoch)))
plt.close()
train_results = utils.evaluate_metrics(np.array(targets), np.array(preds), prediction_is_first_arg=False)
return objs.avg, train_results
def train_epoch(self, train_queue, valid_queue, model, criterion,
optimizer, lr, epoch):
objs = utils.AvgrageMeter()
# TRAINING
preds = []
targets = []
model.train()
for step, (arch_path_enc, y_true) in enumerate(train_queue):
arch_path_enc = arch_path_enc.to(self.device).float()
y_true = y_true.to(self.device).float()
pred = self.model(arch_path_enc)
if self.model_config['loss:loss_log_transform']:
loss = torch.mean(
torch.abs((self.normalize_data(pred) /
self.normalize_data(y_true / 100)) - 1))
else:
loss = criterion(1 - pred, 1 - y_true / 100)
if self.model_config['loss:pairwise_ranking_loss']:
m = 0.1
pairwise_ranking_loss = []
sort_idx = torch.argsort(y_true, descending=True)
for idx, idx_y_i in enumerate(sort_idx):
for idx_y_i_p1 in sort_idx[idx + 1:]:
pairwise_ranking_loss.append(
torch.max(torch.tensor(0.0, dtype=torch.float),
m - (pred[idx_y_i] - pred[idx_y_i_p1])))
pairwise_ranking_loss = torch.mean(
torch.stack(pairwise_ranking_loss))
loss += pairwise_ranking_loss
if step % self.data_config['report_freq'] == 0:
logging.info('Pairwise ranking loss {}'.format(
pairwise_ranking_loss))
preds.extend(pred.detach().cpu().numpy() * 100)
targets.extend(y_true.detach().cpu().numpy())
optimizer.zero_grad()
loss.backward()
optimizer.step()
objs.update(loss.data.item(), len(arch_path_enc))
if step % self.data_config['report_freq'] == 0:
logging.info('train %03d %e', step, objs.avg)
fig = utils.scatter_plot(np.array(preds),
np.array(targets),
xlabel='Predicted',
ylabel='True',
title='')
fig.savefig(
os.path.join(self.log_dir,
'pred_vs_true_train_{}.jpg'.format(epoch)))
plt.close()
train_results = utils.evaluate_metrics(np.array(targets),
np.array(preds),
prediction_is_first_arg=False)
return objs.avg, train_results