def test_createall(self):
for type_file in self.model_type_files:
associated_task = type_file.split("/")[-2]
models_config = OmegaConf.load(type_file)
models_config = OmegaConf.merge(models_config, self.data_config)
models_config.update("data.task", associated_task)
for model_name in models_config.models.keys():
print(model_name)
if model_name not in ["MyTemplateModel"]:
models_config.update("model_name", model_name)
instantiate_model(models_config, MockDatasetGeometric(6))
def test_pointnet2ms(self):
params = load_model_config("segmentation", "pointnet2", "pointnet2ms")
dataset = MockDatasetGeometric(5)
model = instantiate_model(params, dataset)
model.set_input(dataset[0])
model.forward()
model.backward()
def test_kpconv(self):
params = load_model_config("segmentation", "kpconv", "SimpleKPConv")
dataset = MockDatasetGeometric(5)
model = instantiate_model(params, dataset)
model.set_input(dataset[0])
model.forward()
model.backward()
def test_model_ckpt_using_pointnet2ms(self,):
# Create a checkpt
name = "model"
self.run_path = os.path.join(DIR, "checkpt")
print(self.run_path)
if not os.path.exists(self.run_path):
os.makedirs(self.run_path)
model_checkpoint = ModelCheckpoint(self.run_path, name, "test", run_config=self.config, resume=False)
dataset = MockDatasetGeometric(5)
model = instantiate_model(self.config, dataset)
model.set_input(dataset[0], "cpu")
model.instantiate_optimizers(self.config)
mock_metrics = {"current_metrics": {"acc": 12}, "stage": "test", "epoch": 10}
model_checkpoint.save_best_models_under_current_metrics(model, mock_metrics)
# Load checkpoint and initialize model
model_checkpoint = ModelCheckpoint(self.run_path, name, "test", self.config, resume=True)
model2 = model_checkpoint.create_model(dataset, weight_name="acc")
self.assertEqual(str(model.optimizer.__class__.__name__), str(model2.optimizer.__class__.__name__))
self.assertEqual(model.optimizer.defaults, model2.optimizer.defaults)
self.assertEqual(model.schedulers["lr_scheduler"].state_dict(), model2.schedulers["lr_scheduler"].state_dict())
self.assertEqual(model.schedulers["bn_scheduler"].state_dict(), model2.schedulers["bn_scheduler"].state_dict())
shutil.rmtree(self.run_path)
remove(os.path.join(ROOT, "{}.pt".format(name)))
remove(os.path.join(DIR, "{}.pt".format(name)))
def create_model(self, dataset, weight_name=Checkpoint._LATEST):
if not self.is_empty:
run_config = copy.deepcopy(self._checkpoint.run_config)
model = instantiate_model(run_config, dataset)
self._initialize_model(model, weight_name)
return model
else:
raise ValueError("Checkpoint is empty")
def test_largekpconv(self):
params = load_model_config("segmentation", "kpconv", "KPConvPaper")
params.update("data.use_category", True)
params.update("data.first_subsampling", 0.02)
dataset = MockDatasetGeometric(5)
model = instantiate_model(params, dataset)
model.set_input(dataset[0])
model.forward()
model.backward()
def test_kpconvpretransform(self):
params = load_model_config("segmentation", "kpconv", "SimpleKPConv")
dataset = MockDatasetGeometric(5)
model = instantiate_model(params, dataset)
model.eval()
dataset_transform = MockDatasetGeometric(5)
dataset_transform.set_strategies(model)
model.set_input(dataset[0])
model.forward()
model.get_output()
torch.testing.assert_allclose(dataset_transform[0].pos, dataset[0].pos)
def main(cfg):
OmegaConf.set_struct(
cfg,
False) # This allows getattr and hasattr methods to function correctly
if cfg.pretty_print:
print(cfg.pretty())
set_debugging_vars_to_global(cfg.debugging)
# Get device
device = torch.device("cuda" if (
torch.cuda.is_available() and cfg.training.cuda) else "cpu")
log.info("DEVICE : {}".format(device))
# Enable CUDNN BACKEND
torch.backends.cudnn.enabled = cfg.training.enable_cudnn
dataset = instantiate_dataset(cfg.data)
model = instantiate_model(cfg, dataset)
log.info(model)
log.info(
"Model size = %i",
sum(param.numel() for param in model.parameters()
if param.requires_grad))
# Set dataloaders
dataset.create_dataloaders(
model,
cfg.training.batch_size,
cfg.training.shuffle,
cfg.training.num_workers,
cfg.training.precompute_multi_scale,
)
log.info(dataset)
# Run training / evaluation
model = model.to(device)
measurement_name = "{}_{}".format(cfg.model_name,
dataset.__class__.__name__)
run(cfg, model, dataset, device, measurement_name)
def test_runall(self):
def is_known_to_fail(model_name):
forward_failing = [
"MinkUNet_WIP", "pointcnn", "RSConv_4LD", "RSConv_2LD",
"randlanet"
]
for failing in forward_failing:
if failing.lower() in model_name.lower():
return True
return False
def get_dataset(conv_type):
features = 2
if conv_type.lower() == "dense":
return MockDataset(features, num_points=2048)
if conv_type.lower() == "sparse":
return MockDatasetGeometric(features,
transform=ToSparseInput(0.01),
num_points=1024)
return MockDatasetGeometric(features)
for type_file in self.model_type_files:
associated_task = type_file.split("/")[-2]
models_config = OmegaConf.load(type_file)
models_config = OmegaConf.merge(models_config, self.data_config)
models_config.update("data.task", associated_task)
for model_name in models_config.models.keys():
with self.subTest(model_name):
if not is_known_to_fail(model_name):
models_config.update("model_name", model_name)
dataset = get_dataset(
models_config.models[model_name].conv_type)
model = instantiate_model(models_config, dataset)
model.set_input(dataset[0], device)
try:
model.forward()
model.backward()
except Exception as e:
print("Model failing:")
print(model)
raise e
def test_accumulated_gradient(self):
params = load_model_config("segmentation", "pointnet2", "pointnet2ms")
config_training = OmegaConf.load(
os.path.join(DIR, "test_config/training_config.yaml"))
dataset = MockDatasetGeometric(5)
model = instantiate_model(params, dataset)
model.instantiate_optimizers(config_training)
model.set_input(dataset[0])
expected_make_optimizer_step = [
False, False, True, False, False, True, False, False, True, False
]
expected_contains_grads = [
False, True, True, False, True, True, False, True, True, False
]
make_optimizer_steps = []
contains_grads = []
for epoch in range(10):
model.forward()
make_optimizer_step = model.manage_optimizer_zero_grad(
) # Accumulate gradient if option is up
make_optimizer_steps.append(make_optimizer_step)
grad_ = model._modules["lin1"].weight.grad
if grad_ is not None:
contains_grads.append((grad_.sum() != 0).item())
else:
contains_grads.append(False)
model.backward() # calculate gradients
if make_optimizer_step:
model._optimizer.step() # update parameters
self.assertEqual(contains_grads, expected_contains_grads)
self.assertEqual(make_optimizer_steps, expected_make_optimizer_step)
def main(cfg):
OmegaConf.set_struct(
cfg,
False) # This allows getattr and hasattr methods to function correctly
if cfg.pretty_print:
print(cfg.pretty())
# Get device
device = torch.device("cuda" if (
torch.cuda.is_available() and cfg.training.cuda) else "cpu")
log.info("DEVICE : {}".format(device))
# Enable CUDNN BACKEND
torch.backends.cudnn.enabled = cfg.training.enable_cudnn
# Start Wandb if public
launch_wandb(cfg, cfg.wandb.public and cfg.wandb.log)
# Checkpoint
checkpoint = ModelCheckpoint(
cfg.training.checkpoint_dir,
cfg.model_name,
cfg.training.weight_name,
run_config=cfg,
resume=bool(cfg.training.checkpoint_dir),
)
# Create model and datasets
if not checkpoint.is_empty:
dataset = instantiate_dataset(checkpoint.data_config)
model = checkpoint.create_model(dataset,
weight_name=cfg.training.weight_name)
else:
dataset = instantiate_dataset(cfg.data)
model = instantiate_model(cfg, dataset)
model.instantiate_optimizers(cfg)
log.info(model)
model.log_optimizers()
log.info(
"Model size = %i",
sum(param.numel() for param in model.parameters()
if param.requires_grad))
# Set dataloaders
dataset.create_dataloaders(
model,
cfg.training.batch_size,
cfg.training.shuffle,
cfg.training.num_workers,
cfg.training.precompute_multi_scale,
)
log.info(dataset)
# Choose selection stage
selection_stage = getattr(cfg, "selection_stage", "")
checkpoint.selection_stage = dataset.resolve_saving_stage(selection_stage)
tracker: BaseTracker = dataset.get_tracker(model, dataset, cfg.wandb.log,
cfg.tensorboard.log)
launch_wandb(cfg, not cfg.wandb.public and cfg.wandb.log)
# Run training / evaluation
model = model.to(device)
visualizer = Visualizer(cfg.visualization, dataset.num_batches,
dataset.batch_size, os.getcwd())
run(cfg, model, dataset, device, tracker, checkpoint, visualizer)