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)
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 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 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 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 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 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 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 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
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 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 ensemble_analysis(nasbench_data, ensemble_parent_dir, data_splits_root):
# Get model directories
member_dirs = get_ensemble_member_dirs(ensemble_parent_dir)
member_dirs_dict = group_dirs_by_surrogate_type(member_dirs)
print("==> Found ensemble member directories:", member_dirs_dict)
# Load an ensemble for each surrogate type
surrogate_ensemble = None
for surrogate_type, member_dirs in member_dirs_dict.items():
# Load config
print("==> Loading %s configs..." % surrogate_type)
model_log_dir = member_dirs[0]
data_config = json.load(
open(os.path.join(model_log_dir, 'data_config.json'), 'r'))
model_config = json.load(
open(os.path.join(model_log_dir, 'model_config.json'), 'r'))
train_paths = json.load(
open(os.path.join(data_splits_root, 'train_paths.json'), 'r'))
val_paths = json.load(
open(os.path.join(data_splits_root, 'val_paths.json'), 'r'))
test_paths = json.load(
open(os.path.join(data_splits_root, 'test_paths.json'), 'r'))
# Load ensemble
print("==> Loading %s ensemble..." % surrogate_type)
surrogate_ensemble_single = Ensemble(
member_model_name=model_config['model'],
data_root='None',
log_dir=ensemble_parent_dir,
starting_seed=data_config['seed'],
model_config=model_config,
data_config=data_config,
ensemble_size=len(member_dirs),
init_ensemble=False)
surrogate_ensemble_single.load(model_paths=member_dirs,
train_paths=train_paths,
val_paths=val_paths,
test_paths=test_paths)
# Combine different model types
if surrogate_ensemble is None:
surrogate_ensemble = surrogate_ensemble_single
else:
for member_model in surrogate_ensemble_single.ensemble_members:
surrogate_ensemble.add_member(member_model)
print("==> Ensemble creation completed.")
# Set the same seed as used during the training.
np.random.seed(data_config['seed'])
# Evaluate ensemble
print("==> Evaluating ensemble performance...")
train_metrics, train_preds, train_stddevs, train_targets = surrogate_ensemble.evaluate_ensemble(
train_paths, apply_noise=False)
val_metrics, val_preds, val_stddevs, val_targets = surrogate_ensemble.validate_ensemble(
apply_noise=False)
test_metrics, test_preds, test_stddevs, test_targets = surrogate_ensemble.test_ensemble(
apply_noise=False)
print('==> Ensemble train metrics', train_metrics)
print('==> Ensemble val metrics', val_metrics)
print('==> Ensemble test metrics', test_metrics)
train_metrics_with_noise, train_preds_with_noise, _, _ = surrogate_ensemble.evaluate_ensemble(
train_paths, apply_noise=True)
val_metrics_with_noise, val_preds_with_noise, _, _ = surrogate_ensemble.validate_ensemble(
apply_noise=True)
test_metrics_with_noise, test_preds_with_noise, _, _ = surrogate_ensemble.test_ensemble(
apply_noise=True)
print('==> Ensemble train metrics (noisy)', train_metrics_with_noise)
print('==> Ensemble val metrics (noisy)', val_metrics_with_noise)
print('==> Ensemble test metrics (noisy)', test_metrics_with_noise)
train_mean_stddev = np.mean(train_stddevs)
val_mean_stddev = np.mean(val_stddevs)
test_mean_stddev = np.mean(test_stddevs)
print("==> Mean ensemble stddev on train set %f" % train_mean_stddev)
print("==> Mean ensemble stddev on validation set %f" % val_mean_stddev)
print("==> Mean ensemble stddev on test set %f" % test_mean_stddev)
# Plots
fig_val = utils.scatter_plot(np.array(train_preds),
np.array(train_targets),
xlabel='Predicted',
ylabel='True',
title='')
fig_val.savefig(os.path.join(ensemble_parent_dir,
'pred_vs_true_train.jpg'))
plt.close()
fig_val = utils.scatter_plot(np.array(val_preds),
np.array(val_targets),
xlabel='Predicted',
ylabel='True',
title='')
fig_val.savefig(os.path.join(ensemble_parent_dir, 'pred_vs_true_val.jpg'))
plt.close()
fig_test = utils.scatter_plot(np.array(test_preds),
np.array(test_targets),
xlabel='Predicted',
ylabel='True',
title='')
fig_test.savefig(os.path.join(ensemble_parent_dir,
'pred_vs_true_test.jpg'))
plt.close()
fig_val = utils.scatter_plot(np.array(train_preds_with_noise),
np.array(train_targets),
xlabel='Predicted',
ylabel='True',
title='')
fig_val.savefig(
os.path.join(ensemble_parent_dir, 'pred_vs_true_train_noisy.jpg'))
plt.close()
fig_val_noise = utils.scatter_plot(np.array(val_preds_with_noise),
np.array(val_targets),
xlabel='Predicted',
ylabel='True',
title='')
fig_val_noise.savefig(
os.path.join(ensemble_parent_dir, 'pred_vs_true_val_noisy.jpg'))
plt.close()
fig_test_noise = utils.scatter_plot(np.array(test_preds_with_noise),
np.array(test_targets),
xlabel='Predicted',
ylabel='True',
title='')
fig_test_noise.savefig(
os.path.join(ensemble_parent_dir, 'pred_vs_true_test_noisy.jpg'))
plt.close()
# Test query_mean method
# print("==> Testing ensemble predictions with query mean...")
# test_results, configs, val_scores, test_scores = get_test_configs(member_dirs[0])
# preds = [surrogate_ensemble.query_mean(config) for config in configs]
# metrics = utils.evaluate_metrics(val_scores, preds, prediction_is_first_arg=False)
# print("==> Metrics on test data (query mean):", metrics)
# Perform cell topology performance analysis
print("==> Checking configs from cell topology analysis...")
test_paths = [
filename for filename in Path(
"/home/user/projects/nasbench_201_2/analysis/nb_301_cell_topology/cell_topology_analysis"
).rglob('*.json')
]
test_metrics_topology, ensemble_predictions_topology, stddevs_topology, targets_topology = surrogate_ensemble.evaluate_ensemble(
test_paths, apply_noise=False)
print("==> Metrics on cell topology analysis:", test_metrics_topology)
# Log
results = {
"train_metrics": train_metrics,
"val_metrics": val_metrics,
"test_metrics": test_metrics,
"train_metrics_with_noise": train_metrics_with_noise,
"val_metrics_with_noise": val_metrics_with_noise,
"test_metrics_with_noise": test_metrics_with_noise,
"train_mean_stddev": train_mean_stddev,
"val_mean_stddev": val_mean_stddev,
"test_mean_stddev": test_mean_stddev,
"test_results_cell_topology": test_metrics_topology
}
for key, val in results.items():
if isinstance(val, dict):
for subkey, subval in val.items():
results[key][subkey] = np.float64(subval)
else:
try:
results[key] = np.float64(val)
except:
pass
json.dump(
results,
open(os.path.join(ensemble_parent_dir, "ensemble_performance.json"),
"w"))
