def run(args):
# Seed and load parameters
hyperparams = json.load(open(args.hyperparams))
torchfunc.seed(hyperparams["seed"], cuda=True)
# Setup appropriate model and data
# Datasets specified by user
datasets = get_datasets(args, train=args.train)
print(
f"Datasets used: {[type(dataset).__name__ for dataset in datasets]}\n")
dataset = data.datasets.get(args, *datasets)
models = _dev_utils.get_models(args.models)
writer = SummaryWriter(log_dir=args.tensorboard)
for task, model in enumerate(models):
print(
f"==================================TASK {task}======================================"
)
print(f"Model:\n {model}")
# Model is needed to register it's params for transport & proximity loss functions
loss = nn.loss.get(args, model)
# New, better logging of results into tensorboard
single_pass = nn.passes.Validation(model, loss, args.cuda)
loop = nn.train.get_loop(single_pass, dataset, hyperparams)
gatherer = nn.metrics.get(writer,
dataset,
stage=f"SingleTaskModel{task}",
tasks=len(datasets),
input_type=args.input)
# Run training and validation
_dev_utils.run(loop, gatherer)
print("\n")
def test_seed_context_manager():
first_seed = torch.initial_seed()
with torchfunc.seed(0):
assert 0 == torch.initial_seed()
assert torch.initial_seed() == first_seed
def test_seed_representation():
assert repr(torchfunc.seed(0)) == "torchfunc.seed(value=0, cuda=False)"
def test_seed_str():
assert str(torchfunc.seed(0)) == "torchfunc.seed"
def test_seed():
torchfunc.seed(0)
assert 0 == torch.initial_seed()
def run(args):
# Seed and load parameters
hyperparams = json.load(open(args.hyperparams))
torchfunc.seed(hyperparams["seed"], cuda=True)
# Setup appropriate model and data
model = nn.models.get(args)
# Datasets specified by user
train_datasets = get_datasets(args, train=True)
validation_datasets = get_datasets(args, train=False)
print(
f"Datasets used: {[type(dataset).__name__ for dataset in train_datasets]}\n"
)
# Appropriate dataloader constructed from those datasets
train, validation = (
data.datasets.get(args, *train_datasets),
data.datasets.get(args, *validation_datasets),
)
# Dummy pass with data to infer input shapes (torchlayers dependency)
model(train.shape) # Unsqueeze to add batch
print(f"Model to be trained:\n{model}\n")
# Model is needed to register it's params for transport & proximity loss functions
loss = nn.loss.get(args, model)
optimizer = nn.optimizer.get(hyperparams, model)
print(
"Per epoch model accuracy and loss will be logged using Tensorboard.\n"
f"Log directory: {args.tensorboard}\n")
writer = SummaryWriter(log_dir=args.tensorboard)
train_gatherer = nn.metrics.get(
writer,
train,
stage="Train",
tasks=len(train_datasets),
input_type=args.input,
)
validation_gatherer = nn.metrics.get(writer,
validation,
stage="Validation",
tasks=len(validation_datasets),
input_type=args.input)
# Save best model
checkpointer = None
train_pass = nn.passes.Train(model, loss, optimizer, args.cuda)
validation_pass = nn.passes.Validation(model, loss, args.cuda)
train_loop = nn.train.get_loop(train_pass, train, hyperparams)
validation_loop = nn.train.get_loop(validation_pass, validation,
hyperparams)
# Run training and validation
for epoch in range(hyperparams["epochs"]):
print(
f"=============================== TRAINING {epoch} ==================================="
)
_dev_utils.run(train_loop, train_gatherer, epoch, True, checkpointer)
print(
f"============================== VALIDATION {epoch} =================================="
)
_dev_utils.run(validation_loop, validation_gatherer, epoch, False,
checkpointer)
print(f"Final model, saving at: {args.save}")
torch.save(model, args.save)