def _create_cp_logger(fname, ex_name, store_frequency, X, rank, init="random"):
init_factors, _ = cp.initialize_factors(X, rank, init)
init_loss = cp.cp_loss(init_factors, X)
X_norm = np.linalg.norm(X.ravel())
del init_factors
def loss(factors):
return cp.cp_loss(factors, X)
def gradient(factors):
return np.linalg.norm(np.concatenate(cp.cp_grad(factors, X)))
def cp_opt_fit(factors):
return 1 - cp.cp_loss(factors, X) / init_loss
def cp_als_fit(factors):
err = X - base.ktensor(*factors)
return 1 - np.linalg.norm(err.ravel()) / X_norm
args = tuple()
log_metrics = {
'loss': loss,
'gradient': gradient,
'cp_opt_fit': cp_opt_fit,
'cp_als_fit': cp_als_fit,
}
return HDF5Logger(fname, ex_name, store_frequency, args, **log_metrics)
def _cp_opt_final_eval(X, experiment_params, outputs):
num_factors = len(X.shape)
factors = [outputs[f'factor_mode_{i}'] for i in range(num_factors)]
init_factors = [outputs[f'init_factor_mode_{i}'] for i in range(num_factors)]
rank = experiment_params['rank']
init = experiment_params['init']
init_loss = cp.cp_loss(init_factors, X)
X_norm = np.linalg.norm(X.ravel())
err = X - base.ktensor(*factors)
return {
'final_loss': cp.cp_loss(factors, X),
'final_gradient': np.linalg.norm(np.concatenate(cp.cp_grad(factors, X))),
'final_cp_opt_fit': 1 - cp.cp_loss(factors, X)/init_loss,
'final_cp_als_fit': 1 - np.linalg.norm(err.ravel())/X_norm,
}
def _cp_als_final_eval(X, experiment_params, outputs):
num_factors = len(outputs) - 1
weights = outputs['weights']
factors = [outputs[f'factor_mode_{i}'] for i in range(num_factors)]
factors = [f * weights[np.newaxis]**(1 / len(factors)) for f in factors]
rank = experiment_params['rank']
init = experiment_params['init']
init_factors, _ = cp.initialize_factors(X, rank, init)
init_loss = cp.cp_loss(init_factors, X)
X_norm = np.linalg.norm(X.ravel())
err = X - base.ktensor(*factors)
return {
'final_loss': cp.cp_loss(factors, X),
'final_gradient': np.linalg.norm(np.concatenate(cp.cp_grad(factors,
X))),
'final_cp_opt_fit': 1 - cp.cp_loss(factors, X) / init_loss,
'final_cp_als_fit': 1 - np.linalg.norm(err.ravel()) / X_norm,
}
def _create_cp_opt_logger(fname, ex_name, store_frequency, X, rank, init="random"):
init_factors, _ = cp.initialize_factors(X, rank, init)
init_loss = cp.cp_loss(init_factors, X)
X_norm = np.linalg.norm(X.ravel())
del init_factors
def loss(parameters):
return cp._cp_loss_scipy(parameters, rank, X.shape, X )
def gradient(parameters):
return np.linalg.norm(np.ravel(cp._cp_grad_scipy(parameters, rank, X.shape, X )))
args = tuple()
log_metrics = {
'loss': loss,
'gradient': gradient,
}
return HDF5Logger(
fname, ex_name, store_frequency, args, **log_metrics
)
def cp_opt_fit(factors):
return 1 - cp.cp_loss(factors, X) / init_loss
def loss(factors):
return cp.cp_loss(factors, X)