def get_uci_dataset(dataset_name):
data_path = os.path.join(vadam.__path__[0], 'data')
ds = []
for i in range(20):
dataset = Dataset(data_set=dataset_name + str(i), data_folder=data_path)
x_train, y_train = dataset.load_full_train_set(use_cuda=False)
x_test, y_test = dataset.load_full_test_set(use_cuda=False)
y_train = y_train.unsqueeze(1).double().numpy()
y_test = y_test.unsqueeze(1).double().numpy()
x_train = x_train.double().numpy()
x_test = x_test.double().numpy()
scl = StandardScaler(copy=False)
scl.fit_transform(x_train)
scl.transform(x_test)
y_scl = StandardScaler(copy=False)
y_scl.fit_transform(y_train)
y_scl.transform(y_test)
ds.append((None, None, x_train, y_train.squeeze(), x_test, y_test.squeeze(), y_scl.scale_[0]))
return ds
class Experiment():
def __init__(self, data_set, model_params, train_params, optim_params, evals_per_epoch=1, normalize_x=False, normalize_y=False, results_folder="./results", data_folder=DEFAULT_DATA_FOLDER, use_cuda=torch.cuda.is_available()):
# Store parameters
self.data_set = data_set
self.model_params = model_params
self.train_params = train_params
self.optim_params = optim_params
self.evals_per_epoch = evals_per_epoch
self.normalize_x = normalize_x
self.normalize_y = normalize_y
self.data_folder = data_folder
self.results_folder = results_folder
self.use_cuda = use_cuda
# Set random seed
seed = train_params['seed']
torch.manual_seed(seed)
if self.use_cuda:
torch.cuda.manual_seed_all(seed)
# Initialize metric history
self.objective_history = []
# Initialize data
self.data = Dataset(data_set = data_set, data_folder = data_folder)
## All subclasses should override:
# Define folder name for results
self.folder_name = None
# Define prediction function
self.prediction = None
# Define objective
self.objective = None
# Initialize model
self.model = None
# Initialize optimizer
self.optimizer = None
# Initialize metric history
self.metric_history = None
# Initialize final metric
self.final_metric = None
def run(self, log_metric_history=True):
# Prepare
num_epochs = self.train_params['num_epochs']
batch_size = self.train_params['batch_size']
seed = self.train_params['seed']
# Set random seed
torch.manual_seed(seed)
if self.use_cuda:
torch.cuda.manual_seed_all(seed)
# Prepare data loader for training
train_loader = self.data.get_train_loader(batch_size = batch_size)
# Load full data set for evaluation
x_train, y_train = self.data.load_full_train_set(use_cuda = self.use_cuda)
x_test, y_test = self.data.load_full_test_set(use_cuda = self.use_cuda)
# Compute normalization of x
if self.normalize_x:
self.x_means = torch.mean(x_train, dim=0)
self.x_stds = torch.std(x_train, dim=0)
self.x_stds[self.x_stds == 0] = 1
# Compute normalization of y
if self.normalize_y:
self.y_mean = torch.mean(y_train)
self.y_std = torch.std(y_train)
if self.y_std==0:
self.y_std = 1
# Set iterations for evaluation
num_batches = np.ceil(self.data.get_train_size() / batch_size)
if self.evals_per_epoch > num_batches:
evals_per_epoch = num_batches
else:
evals_per_epoch = self.evals_per_epoch
eval_iters = np.round((1 + np.arange(evals_per_epoch)) * (num_batches / evals_per_epoch)).astype(int)
# Train model
for epoch in range(num_epochs):
# Set model in training mode
self.model.train(True)
# Initialize batch objective accumulator
batch_objective = []
for i, (x, y) in enumerate(train_loader):
# Prepare minibatch
if self.use_cuda:
x, y = x.cuda(), y.cuda()
# Normalize x and y
if self.normalize_x:
x = (x-self.x_means)/self.x_stds
if self.normalize_y:
y = (y-self.y_mean)/self.y_std
# Update parameters
def closure():
self.optimizer.zero_grad()
logits = self.prediction(x)
loss = self.objective(logits, y)
loss.backward()
return loss
loss = self.optimizer.step(closure)
if log_metric_history and (i+1) in eval_iters:
# Set model in test mode
self.model.train(False)
# Evaluate model
with torch.no_grad():
self._evaluate_model(self.metric_history, x_train, y_train, x_test, y_test)
# Store batch objective
batch_objective.append(loss.detach().cpu().item())
# Compute and store average objective from last epoch
self.objective_history.append(np.mean(batch_objective))
if log_metric_history:
# Print progress
self._print_progress(epoch)
else:
# Print average objective from last epoch
self._print_objective(epoch)
# Set model in test mode
self.model.train(False)
# Evaluate model
with torch.no_grad():
self._evaluate_model(self.final_metric, x_train, y_train, x_test, y_test)
def _evaluate_model(self, metric_dict, x_train, y_train, x_test, y_test):
## All subclasses should override:
raise NotImplementedError
def _print_progress(self, epoch):
## All subclasses should override:
raise NotImplementedError
def _print_objective(self, epoch):
# Print average objective from last epoch
print('Epoch [{}/{}], Objective: {:.4f}'.format(
epoch+1,
self.train_params['num_epochs'],
self.objective_history[-1]))
def save(self, save_final_metric=True, save_metric_history=True, save_objective_history=True, save_model=True, save_optimizer=True, create_folder=True, folder_path=None):
# Define folder path
if not folder_path:
folder_path = self.folder_name
# Create folder
if create_folder:
os.makedirs(folder_path, exist_ok=True)
# Store state dictionaries for model and optimizer
if save_model:
torch.save(self.model.state_dict(), os.path.join(folder_path, 'model.pt'))
if save_optimizer:
torch.save(self.optimizer.state_dict(), os.path.join(folder_path, 'optimizer.pt'))
# Store history
if save_final_metric:
output = open(os.path.join(folder_path, 'final_metric.pkl'), 'wb')
pickle.dump(self.final_metric, output)
output.close()
if save_metric_history:
output = open(os.path.join(folder_path, 'metric_history.pkl'), 'wb')
pickle.dump(self.metric_history, output)
output.close()
if save_objective_history:
output = open(os.path.join(folder_path, 'objective_history.pkl'), 'wb')
pickle.dump(self.objective_history, output)
output.close()
def load(self, load_final_metric=True, load_metric_history=True, load_objective_history=True, load_model=True, load_optimizer=True, folder_path=None):
# Define folder path
if not folder_path:
folder_path = self.folder_name
# Load state dictionaries for model and optimizer
if load_model:
state_dict = torch.load(os.path.join(folder_path, 'model.pt'))
self.model.load_state_dict(state_dict)
self.model.train(False)
if load_optimizer:
state_dict = torch.load(os.path.join(folder_path, 'optimizer.pt'))
self.optimizer.load_state_dict(state_dict)
# Load history
if load_final_metric:
pkl_file = open(os.path.join(folder_path, 'final_metric.pkl'), 'rb')
self.final_metric = pickle.load(pkl_file)
pkl_file.close()
if load_metric_history:
pkl_file = open(os.path.join(folder_path, 'metric_history.pkl'), 'rb')
self.metric_history = pickle.load(pkl_file)
pkl_file.close()
if load_objective_history:
pkl_file = open(os.path.join(folder_path, 'objective_history.pkl'), 'rb')
self.objective_history = pickle.load(pkl_file)
pkl_file.close()
torch.manual_seed(seed)
if use_cuda:
torch.cuda.manual_seed_all(seed)
###############
## Load data ##
###############
from vadam.datasets import Dataset
data = Dataset(data_set=data_set, data_folder="./../vadam/data")
train_loader = data.get_train_loader(batch_size)
x_train, y_train = data.load_full_train_set(use_cuda=use_cuda)
x_test, y_test = data.load_full_test_set(use_cuda=use_cuda)
##################
## Define model ##
##################
from vadam.models import MLP
model = MLP(input_size=data.num_features,
hidden_sizes=hidden_sizes,
output_size=None,
act_func=act_func)
if use_cuda:
model = model.cuda()
#######################