def create_experiment(config):
# Fetch dataset
dataset = get_dataset(config)
# Create experiment and set indexes for train, val and test
exp = Experiment(config=config)
exp.define_runs_splits(dataset)
return exp, dataset
def run(experiment_name, idx_k=0):
# Restore configuration and get experiment run
exp = Experiment(load_exp_name=experiment_name)
config = exp.config
exp_run = exp.get_experiment_run(idx_k=idx_k)
# Fetch dataset
dataset = get_dataset(config)
# Create task and class mappings
nr_tasks = 5
task_splitter = ClassTaskSplitter(dataset=dataset,
save_path=exp_run.paths['obj'],
nr_tasks=nr_tasks)
# Get PyTorch datasets
splits = ['train', 'val', 'test']
batch_size = config.get('batch_size', 128)
pytorch_datasets = [{split: None
for split in splits} for task_ix in range(nr_tasks)]
for task_ix in range(nr_tasks):
for split in splits:
index_list = task_splitter.get_task_ixs(
exp_ixs=exp_run.dataset_ixs[split], task_ix=task_ix)
pytorch_datasets[task_ix][split] = TorchDS(dataset_obj=dataset,
index_list=index_list)
print('Got datasets')
# Apply oracles
#task_oracle = get_class('src.continual_learning.oracles.task_oracle.TaskOracle')(
# pytorch_datasets=pytorch_datasets, save_path=exp_run.paths['root'])
oracle = get_class(
'src.continual_learning.oracles.autoencoder_oracle.AutoencoderOracle'
)(pytorch_datasets=pytorch_datasets,
save_path=exp_run.paths['root'],
autoencoder_path=
'src.models.autoencoding.pretrained_autoencoder.PretrainedAutoencoder',
feature_model_name='AlexNet')
print('Initialized oracles')
# Plot confusion matrix
for split in splits:
cm = oracle.get_domain_confusion(split=split)
cm.plot(exp_run.paths['results'],
oracle.name + '_' + split + '_domain_confusion',
label_predicted='Selected Model (predicted)',
label_actual='Data Task (actual)',
figure_size=(7, 5))
def run(experiment_name, idx_k=0):
# Restore configuration and get experiment run
exp = Experiment(load_exp_name=experiment_name)
config = exp.config
exp_run = exp.get_experiment_run(idx_k=idx_k)
# Set GPU
set_gpu(config.get('gpu', 0))
# Fetch dataset
dataset = get_dataset(config)
# Create task and class mappings
nr_tasks = 5
task_splitter = ClassTaskSplitter(dataset=dataset, save_path=exp_run.paths['obj'], nr_tasks=nr_tasks)
# Get PyTorch datasets and dataloaders
splits = ['train', 'val', 'test']
batch_size = config.get('batch_size', 128)
pytorch_datasets = [{split: None for split in splits} for task_ix in range(nr_tasks)]
dataloaders = [{split: None for split in splits} for task_ix in range(nr_tasks)]
for task_ix in range(nr_tasks):
for split in splits:
index_list = task_splitter.get_task_ixs(exp_ixs=exp_run.dataset_ixs[split], task_ix=task_ix)
pytorch_datasets[task_ix][split] = TorchDS(dataset_obj=dataset, index_list=index_list, transform='pretrained')
shuffle = True if split == 'train' else False
dataloaders[task_ix][split] = torch.utils.data.DataLoader(pytorch_datasets[task_ix][split], batch_size=batch_size, shuffle=shuffle)
# Get model and agent
model = get_class(config['model_class_path'])(config, in_channels=dataset.nr_channels, img_size=dataset.img_shape)
model.cuda()
optimizer = torch.optim.Adam(model.parameters(), weight_decay=1e-4)
criterion = nn.MSELoss()
agent = get_class(config['agent_class_path'])(config, model=model, optimizer=optimizer, criterion=criterion)
# Train with data from the first task
for task_ix in range(nr_tasks):
agent.train_model(dataloaders=dataloaders[task_ix], nr_epochs=20)
agent.save_state(path=exp_run.paths['agent_states'], name='task_'+str(task_ix))
for data, _ in dataloader:
b, c, h, w = data.shape
nb_pixels = b * h * w
sum_ = torch.sum(data, dim=[0, 2, 3])
sum_of_square = torch.sum(data**2, dim=[0, 2, 3])
mean = (count * mean + sum_) / (count + nb_pixels)
std = (count * std + sum_of_square) / (count + nb_pixels)
count += nb_pixels
return mean, torch.sqrt(std - mean**2)
#%%
# segChallengeProstate
dataset = get_dataset({"dataset_name": "segChallengeProstate"})
transform = transforms.Compose([
transforms.ToTensor(),
transforms.ToPILImage(),
transforms.Resize(size=(320, 320)),
transforms.ToTensor()
])
pytorch_dataset = TorchSegmentationDataset(dataset_obj=dataset,
index_list=list(
range(len(dataset.instances))),
transform=transform)
print(normalization_values(pytorch_dataset))
# %%
dataset = get_dataset({
"dataset_name": "medcom",