def _set_config_and_seed(self, **kwargs: Any) -> None:
self.train_config: TrainConfig = merge_config( # type:ignore
TrainConfig(),
kwargs # type:ignore
)
# Update base config so that it includes all parameters
random.seed(self.train_config.seed)
def fit(
self,
L_train: np.ndarray,
Y_dev: Optional[np.ndarray] = None,
class_balance: Optional[List[float]] = None,
**kwargs: Any,
) -> None:
"""Train label model.
Train label model to estimate mu, the parameters used to combine LFs.
Parameters
----------
L_train
An [n,m] matrix with values in {-1,0,1,...,k-1}
Y_dev
Gold labels for dev set for estimating class_balance, by default None
class_balance
Each class's percentage of the population, by default None
**kwargs
Arguments for changing train config defaults
Raises
------
Exception
If loss in NaN
Examples
--------
>>> L = np.array([[0, 0, -1], [-1, 0, 1], [1, -1, 0]])
>>> Y_dev = [0, 1, 0]
>>> label_model = LabelModel(verbose=False)
>>> label_model.fit(L)
>>> label_model.fit(L, Y_dev=Y_dev)
>>> label_model.fit(L, class_balance=[0.7, 0.3])
"""
# Set random seed
self.train_config: TrainConfig = merge_config( # type:ignore
TrainConfig(), kwargs # type:ignore
)
# Update base config so that it includes all parameters
random.seed(self.train_config.seed)
np.random.seed(self.train_config.seed)
torch.manual_seed(self.train_config.seed)
L_shift = L_train + 1 # convert to {0, 1, ..., k}
if L_shift.max() > self.cardinality:
raise ValueError(
f"L_train has cardinality {L_shift.max()}, cardinality={self.cardinality} passed in."
)
self._set_constants(L_shift)
self._set_class_balance(class_balance, Y_dev)
self._create_tree()
lf_analysis = LFAnalysis(L_train)
self.coverage = lf_analysis.lf_coverages()
# Compute O and initialize params
if self.config.verbose: # pragma: no cover
logging.info("Computing O...")
self._generate_O(L_shift)
self._init_params()
# Estimate \mu
if self.config.verbose: # pragma: no cover
logging.info("Estimating \mu...")
# Set model to train mode
self.train()
# Move model to GPU
self.mu_init = self.mu_init.to(self.config.device)
if self.config.verbose and self.config.device != "cpu": # pragma: no cover
logging.info("Using GPU...")
self.to(self.config.device)
# Set training components
self._set_logger()
self._set_optimizer()
self._set_lr_scheduler()
# Restore model if necessary
start_iteration = 0
# Train the model
metrics_hist = {} # The most recently seen value for all metrics
for epoch in range(start_iteration, self.train_config.n_epochs):
self.running_loss = 0.0
self.running_examples = 0
# Zero the parameter gradients
self.optimizer.zero_grad()
# Forward pass to calculate the average loss per example
loss = self._loss_mu(l2=self.train_config.l2)
if torch.isnan(loss):
msg = "Loss is NaN. Consider reducing learning rate."
raise Exception(msg)
# Backward pass to calculate gradients
# Loss is an average loss per example
loss.backward()
# Perform optimizer step
self.optimizer.step()
# Calculate metrics, log, and checkpoint as necessary
metrics_dict = self._execute_logging(loss)
metrics_hist.update(metrics_dict)
# Update learning rate
self._update_lr_scheduler(epoch)
# Post-processing operations on mu
self._clamp_params()
self._break_col_permutation_symmetry()
# Return model to eval mode
self.eval()
# Print confusion matrix if applicable
if self.config.verbose: # pragma: no cover
logging.info("Finished Training")
def __init__(self, name: Optional[str] = None, **kwargs: Any) -> None:
self.config: TrainerConfig = merge_config( # type:ignore
TrainerConfig(),
kwargs # type:ignore
)
self.name = name if name is not None else type(self).__name__
def fit(
self,
L_train: np.ndarray,
Y_dev: Optional[np.ndarray] = None,
class_balance: Optional[List[float]] = None,
progress_bar: bool = True,
**kwargs: Any,
) -> None:
"""Train label model.
Train label model to estimate mu, the parameters used to combine LFs.
Parameters
----------
L_train
An [n,m] matrix with values in {-1,0,1,...,k-1}
Y_dev
Gold labels for dev set for estimating class_balance, by default None
class_balance
Each class's percentage of the population, by default None
progress_bar
To display a progress bar, by default True
**kwargs
Arguments for changing train config defaults.
n_epochs
The number of epochs to train (where each epoch is a single
optimization step), default is 100
lr
Base learning rate (will also be affected by lr_scheduler choice
and settings), default is 0.01
l2
Centered L2 regularization strength, default is 0.0
optimizer
Which optimizer to use (one of ["sgd", "adam", "adamax"]),
default is "sgd"
optimizer_config
Settings for the optimizer
lr_scheduler
Which lr_scheduler to use (one of ["constant", "linear",
"exponential", "step"]), default is "constant"
lr_scheduler_config
Settings for the LRScheduler
prec_init
LF precision initializations / priors, default is 0.7
seed
A random seed to initialize the random number generator with
log_freq
Report loss every this many epochs (steps), default is 10
mu_eps
Restrict the learned conditional probabilities to
[mu_eps, 1-mu_eps], default is None
Raises
------
Exception
If loss in NaN
Examples
--------
>>> L = np.array([[0, 0, -1], [-1, 0, 1], [1, -1, 0]])
>>> Y_dev = [0, 1, 0]
>>> label_model = LabelModel(verbose=False)
>>> label_model.fit(L)
>>> label_model.fit(L, Y_dev=Y_dev, seed=2020, lr=0.05)
>>> label_model.fit(L, class_balance=[0.7, 0.3], n_epochs=200, l2=0.4)
"""
# Set random seed
self.train_config: TrainConfig = merge_config( # type:ignore
TrainConfig(), kwargs # type:ignore
)
# Update base config so that it includes all parameters
random.seed(self.train_config.seed)
np.random.seed(self.train_config.seed)
torch.manual_seed(self.train_config.seed)
# Set Logger
self._set_logger()
L_shift = L_train + 1 # convert to {0, 1, ..., k}
if L_shift.max() > self.cardinality:
raise ValueError(
f"L_train has cardinality {L_shift.max()}, cardinality={self.cardinality} passed in."
)
self._set_constants(L_shift)
self._set_class_balance(class_balance, Y_dev)
self._create_tree()
lf_analysis = LFAnalysis(L_train)
self.coverage = lf_analysis.lf_coverages()
# Compute O and initialize params
if self.config.verbose: # pragma: no cover
logging.info("Computing O...")
self._generate_O(L_shift)
self._init_params()
# Estimate \mu
if self.config.verbose: # pragma: no cover
logging.info(r"Estimating \mu...")
# Set model to train mode
self.train()
# Move model to GPU
self.mu_init = self.mu_init.to(self.config.device)
if self.config.verbose and self.config.device != "cpu": # pragma: no cover
logging.info("Using GPU...")
self.to(self.config.device)
# Set training components
self._set_optimizer()
self._set_lr_scheduler()
# Restore model if necessary
start_iteration = 0
# Train the model
metrics_hist = {} # The most recently seen value for all metrics
if progress_bar:
epochs = trange(start_iteration, self.train_config.n_epochs, unit="epoch")
else:
epochs = range(start_iteration, self.train_config.n_epochs)
for epoch in epochs:
self.running_loss = 0.0
self.running_examples = 0
# Zero the parameter gradients
self.optimizer.zero_grad()
# Forward pass to calculate the average loss per example
loss = self._loss_mu(l2=self.train_config.l2)
if torch.isnan(loss):
msg = "Loss is NaN. Consider reducing learning rate."
raise Exception(msg)
# Backward pass to calculate gradients
# Loss is an average loss per example
loss.backward()
# Perform optimizer step
self.optimizer.step()
# Calculate metrics, log, and checkpoint as necessary
metrics_dict = self._execute_logging(loss)
metrics_hist.update(metrics_dict)
# Update learning rate
self._update_lr_scheduler(epoch)
# Cleanup progress bar if enabled
if progress_bar:
epochs.close()
# Post-processing operations on mu
self._clamp_params()
self._break_col_permutation_symmetry()
# Return model to eval mode
self.eval()
# Print confusion matrix if applicable
if self.config.verbose: # pragma: no cover
logging.info("Finished Training")
def partial_fit(
self,
L_train: np.ndarray,
alpha: Optional[float] = 0.05,
Y_dev: Optional[np.ndarray] = None,
class_balance: Optional[List[float]] = None,
update_balance: bool = False,
update_tree: bool = False,
threshold: Optional[float] = 1e-16,
**kwargs: Any,
) -> None:
"""Train label model.
Train label model to estimate mu, the parameters used to combine LFs.
Parameters
----------
L_train
An [n,m] matrix with values in {-1,0,1,...,k-1}
alpha
Exponential smoothing factor, by default 0.05
Y_dev
Gold labels for dev set for estimating class_balance, by default None
class_balance
Each class's percentage of the population, by default None
**kwargs
Arguments for changing train config defaults
Raises
------
Exception
If loss in NaN
Examples
--------
>>> L = np.array([[0, 0, -1], [-1, 0, 1], [1, -1, 0]])
>>> Y_dev = [0, 1, 0]
>>> label_model = LabelModel(verbose=False)
>>> label_model.fit(L)
>>> label_model.fit(L, Y_dev=Y_dev)
>>> label_model.fit(L, class_balance=[0.7, 0.3])
"""
if not self.is_trained:
raise RuntimeError(
f"This instance is not fitted yet. Call 'fit' with "
"appropriate arguments before using this method.")
n, m = L_train.shape
if m != self.m:
raise ValueError(f"L_train must have shape[1]={self.m}.")
# Set number of epochs to one
self.train_config: TrainConfig = merge_config( # type:ignore
TrainConfig(),
{
"n_epochs": 1,
**kwargs
} # type:ignore
)
L_shift = L_train + 1 # convert to {0, 1, ..., k}
if L_shift.max() > self.cardinality:
raise ValueError(
f"L_train has cardinality {L_shift.max()}, cardinality={self.cardinality} passed in."
)
self._set_constants(L_shift)
if update_balance:
self._update_balance(class_balance, Y_dev)
if update_tree:
self._update_tree(L_train, alpha, threshold)
# Build the mask over O^{-1}
self._update_mask()
lf_analysis = LFAnalysis(L_train)
self.coverage = lf_analysis.lf_coverages()
# Compute O
if self.config.verbose: # pragma: no cover
logging.info("Computing O...")
self._update_O(L_shift, alpha)
# Estimate \mu
if self.config.verbose: # pragma: no cover
logging.info("Estimating \mu...")
# Set model to train mode
self.train()
# Move model to GPU
if self.config.verbose and self.config.device != "cpu": # pragma: no cover
logging.info("Using GPU...")
self.to(self.config.device)
# Set training components
self._set_optimizer()
# Restore model
start_iteration = 0
# Train the model
metrics_hist = {} # The most recently seen value for all metrics
for epoch in range(start_iteration, self.train_config.n_epochs):
self.running_loss = 0.0
self.running_examples = 0
# Zero the parameter gradients
self.optimizer.zero_grad()
# Forward pass to calculate the loss
loss = self._loss_mu(l2=self.train_config.l2)
if torch.isnan(loss):
msg = "Loss is NaN. Consider reducing learning rate."
raise Exception(msg)
# Backward pass to calculate gradients
# Loss is an average loss per example
loss.backward()
# Perform optimizer step
self.optimizer.step()
# Calculate metrics, log, and checkpoint as necessary
metrics_dict = self._execute_logging(loss)
metrics_hist.update(metrics_dict)
# Update learning rate
self._update_lr_scheduler(epoch)
# Post-processing operations on mu
self._clamp_params()
self._break_col_permutation_symmetry()
# Return model to eval mode
self.eval()
# Print confusion matrix if applicable
if self.config.verbose: # pragma: no cover
logging.info("Finished Training")
self.is_trained = True
def test_merge_config(self):
config_updates = {"a": 2, "foo_config": {"a": 0.75}}
bar_config = merge_config(BarConfig(), config_updates)
self.assertEqual(bar_config.a, 2)
self.assertEqual(bar_config.foo_config.a, 0.75)