def run_model(config,
seed=0,
data_dir='./data',
genotype_class='PCDARTS',
num_epochs=20,
batch_size=get('batch_size'),
init_channels=get('init_channels'),
train_criterion=torch.nn.CrossEntropyLoss,
data_augmentations=None,
save_model_str=None, **kwargs):
"""
Training loop for configurableNet.
:param model_config: network config (dict)
:param data_dir: dataset path (str)
:param num_epochs: (int)
:param batch_size: (int)
:param learning_rate: model optimizer learning rate (float)
:param train_criterion: Which loss to use during training (torch.nn._Loss)
:param model_optimizer: Which model optimizer to use during trainnig (torch.optim.Optimizer)
:param data_augmentations: List of data augmentations to apply such as rescaling.
(list[transformations], transforms.Composition[list[transformations]], None)
If none only ToTensor is used
:return:
"""
# instantiate optimize
if not torch.cuda.is_available():
logging.info('no gpu device available')
sys.exit(1)
gpu = 'cuda:0'
np.random.seed(seed)
torch.cuda.set_device(gpu)
cudnn.benchmark = True
torch.manual_seed(seed)
cudnn.enabled=True
torch.cuda.manual_seed(seed)
logging.info('gpu device = %s' % gpu)
logging.info("config = %s", config)
if data_augmentations is None:
# You can add any preprocessing/data augmentation you want here
data_augmentations = transforms.ToTensor()
elif isinstance(type(data_augmentations), list):
data_augmentations = transforms.Compose(data_augmentations)
elif not isinstance(data_augmentations, transforms.Compose):
raise NotImplementedError
train_dataset = K49(data_dir, True, data_augmentations)
test_dataset = K49(data_dir, False, data_augmentations)
# train_dataset = KMNIST(data_dir, True, data_augmentations)
# test_dataset = KMNIST(data_dir, False, data_augmentations)
# Make data batch iterable
# Could modify the sampler to not uniformly random sample
train_loader = DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)
test_loader = DataLoader(dataset=test_dataset,
batch_size=batch_size,
shuffle=False)
genotype = eval("genotypes.%s" % genotype_class)
model = Network(init_channels, train_dataset.n_classes, config['n_conv_layers'], genotype)
model = model.cuda()
total_model_params = np.sum(p.numel() for p in model.parameters())
logging.info("param size = %fMB", utils.count_parameters_in_MB(model))
criterion = train_criterion
criterion = criterion.cuda()
if config['optimizer'] == 'sgd':
optimizer = torch.optim.SGD(model.parameters(),
lr=config['initial_lr'],
momentum=config['sgd_momentum'],
weight_decay=config['weight_decay'],
nesterov=config['nesterov'])
else:
optimizer = get('opti_dict')[config['optimizer']](model.parameters(), lr=config['initial_lr'], weight_decay=config['weight_decay'])
if config['lr_scheduler'] == 'Cosine':
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, num_epochs)
elif config['lr_scheduler'] == 'Exponential':
lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.1)
logging.info('Generated Network:')
summary(model, (train_dataset.channels,
train_dataset.img_rows,
train_dataset.img_cols),
device='cuda' if torch.cuda.is_available() else 'cpu')
for epoch in range(num_epochs):
lr_scheduler.step()
logging.info('epoch %d lr %e', epoch, lr_scheduler.get_lr()[0])
model.drop_path_prob = config['drop_path_prob'] * epoch / num_epochs
train_acc, train_obj = train(train_loader, model, criterion, optimizer, grad_clip=config['grad_clip_value'])
logging.info('train_acc %f', train_acc)
test_acc, test_obj = infer(test_loader, model, criterion)
logging.info('test_acc %f', test_acc)
if save_model_str:
# Save the model checkpoint, can be restored via "model = torch.load(save_model_str)"
if os.path.exists(save_model_str):
save_model_str += '_'.join(time.ctime())
torch.save(model.state_dict(), save_model_str)
return test_acc
def compute(self, x, budget, config, **kwargs):
"""
Get model with hyperparameters from config generated by get_configspace()
"""
config = get_config_dictionary(x, config)
print("config", config)
if (len(config.keys())<len(x)):
return 100
if not torch.cuda.is_available():
logging.info('no gpu device available')
sys.exit(1)
gpu = 'cuda:0'
np.random.seed(self.seed)
torch.cuda.set_device(gpu)
cudnn.benchmark = True
torch.manual_seed(self.seed)
cudnn.enabled=True
torch.cuda.manual_seed(self.seed)
logging.info('gpu device = %s' % gpu)
logging.info("config = %s", config)
genotype = eval("genotypes.%s" % 'PCDARTS')
model = Network(self.init_channels, self.n_classes, config['n_conv_layers'], genotype)
model = model.cuda()
logging.info("param size = %fMB", utils.count_parameters_in_MB(model))
criterion = nn.CrossEntropyLoss()
criterion = criterion.cuda()
if config['optimizer'] == 'sgd':
optimizer = torch.optim.SGD(model.parameters(),
lr=config['initial_lr'],
momentum=0.9,
weight_decay=config['weight_decay'],
nesterov=True)
else:
optimizer = settings.opti_dict[config['optimizer']](model.parameters(), lr=config['initial_lr'])
if config['lr_scheduler'] == 'Cosine':
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, int(budget))
elif config['lr_scheduler'] == 'Exponential':
lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.1)
indices = list(range(int(self.split*len(self.train_dataset))))
valid_indices = list(range(int(self.split*len(self.train_dataset)), len(self.train_dataset)))
print("Training size=", len(indices))
training_sampler = SubsetRandomSampler(indices)
valid_sampler = SubsetRandomSampler(valid_indices)
train_queue = torch.utils.data.DataLoader(dataset=self.train_dataset,
batch_size=self.batch_size,
sampler=training_sampler)
valid_queue = torch.utils.data.DataLoader(dataset=self.train_dataset,
batch_size=self.batch_size,
sampler=valid_sampler)
for epoch in range(int(budget)):
lr_scheduler.step()
logging.info('epoch %d lr %e', epoch, lr_scheduler.get_lr()[0])
model.drop_path_prob = config['drop_path_prob'] * epoch / int(budget)
train_acc, train_obj = train(train_queue, model, criterion, optimizer, grad_clip=config['grad_clip_value'])
logging.info('train_acc %f', train_acc)
valid_acc, valid_obj = infer(valid_queue, model, criterion)
logging.info('valid_acc %f', valid_acc)
return valid_obj # Hyperband always minimizes, so we want to minimise the error, error = 1-acc
def create_run_ensemble(model_state_list,
n_layers,
grad_clip_value=5,
seed=0,
num_epochs=20,
learning_rate=0.001,
init_channels=get('init_channels'),
batch_size=get('batch_size'),
genotype_class='PCDARTS'):
if not torch.cuda.is_available():
logging.info('no gpu device available')
sys.exit(1)
gpu = 'cuda:0'
np.random.seed(seed)
torch.cuda.set_device(gpu)
cudnn.benchmark = True
torch.manual_seed(seed)
cudnn.enabled=True
torch.cuda.manual_seed(seed)
logging.info('gpu device = %s' % gpu)
logging.info("config = %s", config)
if data_augmentations is None:
# You can add any preprocessing/data augmentation you want here
data_augmentations = transforms.ToTensor()
elif isinstance(type(data_augmentations), list):
data_augmentations = transforms.Compose(data_augmentations)
elif not isinstance(data_augmentations, transforms.Compose):
raise NotImplementedError
train_dataset = K49(data_dir, True, data_augmentations)
test_dataset = K49(data_dir, False, data_augmentations)
# train_dataset = KMNIST(data_dir, True, data_augmentations)
# test_dataset = KMNIST(data_dir, False, data_augmentations)
# Make data batch iterable
# Could modify the sampler to not uniformly random sample
train_loader = DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)
test_loader = DataLoader(dataset=test_dataset,
batch_size=batch_size,
shuffle=False)
genotype = eval("genotypes.%s" % genotype_class)
dataset = dict()
dims = []
for i, model_state in enumerate(model_state_list):
model = Network(init_channels, train_dataset.n_classes, n_layers, genotype)
model.load_state_dict(torch.load(model_state))
model.cuda()
for p in model.parameters():
p.requires_grad = False
trn_labels = []
trn_features = []
if i == 0:
for d,la in train_loader:
o = model(Variable(d.cuda()))
o = o.view(o.size(0),-1)
trn_labels.extend(la)
trn_features.extend(o.cpu().data)
test_labels = []
test_features = []
for d,la in test_loader:
o = model(Variable(d.cuda()))
o = o.view(o.size(0),-1)
test_labels.extend(la)
test_features.extend(o.cpu().data)
dataset['trn_labels'] = trn_labels
dataset['test_labels'] = test_labels
else:
for d,la in train_loader:
o = model(Variable(d.cuda()))
o = o.view(o.size(0),-1)
trn_features.extend(o.cpu().data)
test_labels = []
test_features = []
for d,la in test_loader:
o = model(Variable(d.cuda()))
o = o.view(o.size(0),-1)
test_features.extend(o.cpu().data)
dataset['trn_features'].extend(trn_features)
dims.extend(dataset['trn_features'][i][0].size(0))
dataset['test_features'].extend(test_features)
trn_feat_dset = FeaturesDataset(dataset['trn_features'][0],dataset['trn_features'][1],dataset['trn_features'][2],dataset['trn_labels'])
test_feat_dset = FeaturesDataset(dataset['test_features'][0],dataset['test_features'][1],dataset['test_features'][2],dataset['test_labels'])
trn_feat_loader = DataLoader(trn_feat_dset,batch_size=64,shuffle=True)
test_feat_loader = DataLoader(val_feat_dset,batch_size=64)
model = EnsembleModel(dims, out_size=train_dataset.n_classes)
criterion = torch.nn.optim.CrossEntropyLoss
criterion = criterion.cuda()
optimizer = torch.nn.optim.SGD(model.parameters(),
lr=learning_rate,
momentum=0.9)
for epoch in range(num_epochs):
epoch_loss, epoch_accuracy = fit(epoch,model,trn_feat_loader,critierion, training=True)
val_epoch_loss , val_epoch_accuracy = fit(epoch,model, test_feat_loader, criterion, training=False)
if save_model_str:
# Save the model checkpoint, can be restored via "model = torch.load(save_model_str)"
if not os.path.exists(save_model_str):
os.mkdir(save_model_str)
torch.save(model.state_dict(), os.path.join(save_model_str, time.ctime()))
def create_run_ensemble(model_description,
ensemble_config,
seed=get('seed'),
num_epochs=20,
data_dir='./data',
init_channels=get('init_channels'),
batch_size=get('batch_size'),
genotype_class='PCDARTS',
data_augmentations=None,
save_model_str=None):
if not torch.cuda.is_available():
logging.info('no gpu device available')
sys.exit(1)
gpu = 'cuda:0'
np.random.seed(seed)
torch.cuda.set_device(gpu)
cudnn.benchmark = True
torch.manual_seed(seed)
cudnn.enabled = True
torch.cuda.manual_seed(seed)
logging.info('gpu device = %s' % gpu)
if data_augmentations is None:
# You can add any preprocessing/data augmentation you want here
data_augmentations = transforms.ToTensor()
elif isinstance(type(data_augmentations), list):
data_augmentations = transforms.Compose(data_augmentations)
elif not isinstance(data_augmentations, transforms.Compose):
raise NotImplementedError
train_dataset = K49(data_dir, True, data_augmentations)
test_dataset = K49(data_dir, False, data_augmentations)
# train_dataset = KMNIST(data_dir, True, data_augmentations)
# test_dataset = KMNIST(data_dir, False, data_augmentations)
# Make data batch iterable
# Could modify the sampler to not uniformly random sample
train_loader = DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)
test_loader = DataLoader(dataset=test_dataset,
batch_size=batch_size,
shuffle=False)
genotype = eval("genotypes.%s" % genotype_class)
trained_models = []
for i, model_state in enumerate(model_description.keys()):
model = Network(
init_channels, train_dataset.n_classes,
model_description[model_state]['config']['n_conv_layers'],
genotype)
model.load_state_dict(
torch.load(model_description[model_state]['model_path']))
model.cuda()
model.drop_path_prob = model_description[model_state]['config'][
'drop_path_prob']
trained_models.append(model)
ensemble_model = EnsembleModel(trained_models,
dense_units=ensemble_config['dense_units'],
out_size=train_dataset.n_classes)
ensemble_model = ensemble_model.cuda()
summary(ensemble_model, input_size=(1, 28, 28))
criterion = torch.nn.CrossEntropyLoss()
criterion = criterion.cuda()
if ensemble_config['optimizer'] == 'sgd':
optimizer = torch.optim.SGD(
model.parameters(),
lr=ensemble_config['initial_lr'],
momentum=ensemble_config['sgd_momentum'],
weight_decay=ensemble_config['weight_decay'],
nesterov=ensemble_config['nesterov'])
else:
optimizer = get('opti_dict')[ensemble_config['optimizer']](
model.parameters(),
lr=ensemble_config['initial_lr'],
weight_decay=ensemble_config['weight_decay'])
if ensemble_config['lr_scheduler'] == 'Cosine':
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, num_epochs)
elif ensemble_config['lr_scheduler'] == 'Exponential':
lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer,
gamma=0.1)
print('Started Training')
for epoch in range(num_epochs):
logging.info('epoch %d lr %e', epoch, lr_scheduler.get_lr()[0])
model.drop_path_prob = ensemble_config[
'drop_path_prob'] * epoch / num_epochs
for p in ensemble_model.model_1.parameters():
p.requires_grad = False
for p in ensemble_model.model_2.parameters():
p.requires_grad = False
for p in ensemble_model.model_3.parameters():
p.requires_grad = False
for p in ensemble_model.out_classifier.parameters():
p.requires_grad = True
train_acc, train_obj, models_avg = ensemble_train(
train_loader,
ensemble_model,
criterion,
optimizer,
grad_clip=ensemble_config['grad_clip_value'])
logging.info('train_acc %f', train_acc)
logging.info('models_avg {}'.format(models_avg))
lr_scheduler.step()
test_acc, test_obj, models_avg = ensemble_infer(
test_loader, ensemble_model, criterion)
logging.info('test_acc %f', test_acc)
logging.info('models_avg {}'.format(models_avg))
if save_model_str:
# Save the model checkpoint, can be restored via "model = torch.load(save_model_str)"
if not os.path.exists(save_model_str):
os.mkdir(save_model_str)
os.path.join(save_model_str, 'ENSEMBLE')
torch.save(ensemble_model.state_dict(),
os.path.join(save_model_str, time.ctime()))