diff --git a/pes_to_spec/bnn.py b/pes_to_spec/bnn.py
index 32461c28a8c9313fe8adea470240d397f12e5627..7c0b3411199f5337f3bf60f3d1ac40af3d8f2ac2 100644
--- a/pes_to_spec/bnn.py
+++ b/pes_to_spec/bnn.py
@@ -2,13 +2,259 @@ from sklearn.base import BaseEstimator, RegressorMixin
from typing import Any, Dict, Optional, Union, Tuple
import numpy as np
+import math
from scipy.special import gamma
import torch
import torch.nn as nn
-import torchbnn as bnn
+import torch.nn.functional as F
from torch.utils.data import TensorDataset, DataLoader
+class BayesLinearEmpiricalPrior(nn.Module):
+ """
+ Applies Bayesian Linear
+
+ Args:
+ prior_mu (Float): mean of prior normal distribution.
+ prior_sigma (Float): sigma of prior normal distribution.
+
+ """
+ __constants__ = ['prior_mu', 'prior_sigma', 'bias', 'in_features', 'out_features']
+
+ def __init__(self, prior_mu, prior_sigma, in_features, out_features, bias=True):
+ super(BayesLinearEmpiricalPrior, self).__init__()
+ self.in_features = in_features
+ self.out_features = out_features
+
+ self.prior_mu = prior_mu
+ self.prior_sigma = prior_sigma
+
+ self.prior_log_sigma_w = nn.Parameter(torch.ones((out_features, in_features))*np.log(prior_sigma))
+ self.prior_log_sigma_b = nn.Parameter(torch.ones((out_features,))*np.log(prior_sigma))
+
+ self.weight_mu = nn.Parameter(torch.Tensor(out_features, in_features))
+ self.weight_log_sigma = nn.Parameter(torch.Tensor(out_features, in_features))
+ self.register_buffer('weight_eps', None)
+
+ if bias is None or bias is False :
+ self.bias = False
+ else :
+ self.bias = True
+
+ if self.bias:
+ self.bias_mu = nn.Parameter(torch.Tensor(out_features))
+ self.bias_log_sigma = nn.Parameter(torch.Tensor(out_features))
+ self.register_buffer('bias_eps', None)
+ else:
+ self.register_parameter('bias_mu', None)
+ self.register_parameter('bias_log_sigma', None)
+ self.register_buffer('bias_eps', None)
+
+ self.reset_parameters()
+
+ def reset_parameters(self):
+ stdv = 1. / np.sqrt(self.weight_mu.size(1))
+ self.weight_mu.data.uniform_(-stdv, stdv)
+ self.weight_log_sigma.data.fill_(np.log(self.prior_sigma))
+ if self.bias :
+ self.bias_mu.data.uniform_(-stdv, stdv)
+ self.bias_log_sigma.data.fill_(np.log(self.prior_sigma))
+
+ def freeze(self) :
+ self.weight_eps = torch.randn_like(self.weight_log_sigma)
+ if self.bias :
+ self.bias_eps = torch.randn_like(self.bias_log_sigma)
+
+ def unfreeze(self) :
+ self.weight_eps = None
+ if self.bias :
+ self.bias_eps = None
+
+ def forward(self, input):
+ r"""
+ Overriden.
+ """
+ if self.weight_eps is None :
+ weight = self.weight_mu + torch.exp(self.weight_log_sigma) * torch.randn_like(self.weight_log_sigma)
+ else :
+ weight = self.weight_mu + torch.exp(self.weight_log_sigma) * self.weight_eps
+
+ if self.bias:
+ if self.bias_eps is None :
+ bias = self.bias_mu + torch.exp(self.bias_log_sigma) * torch.randn_like(self.bias_log_sigma)
+ else :
+ bias = self.bias_mu + torch.exp(self.bias_log_sigma) * self.bias_eps
+ else :
+ bias = None
+
+ return F.linear(input, weight, bias)
+
+ def extra_repr(self):
+ r"""
+ Overriden.
+ """
+ return 'prior_mu={}, prior_sigma={}, in_features={}, out_features={}, bias={}'.format(self.prior_mu, self.prior_sigma, self.in_features, self.out_features, self.bias is not None)
+
+class BayesLinear(nn.Module):
+ r"""
+ Applies Bayesian Linear
+
+ Arguments:
+ prior_mu (Float): mean of prior normal distribution.
+ prior_sigma (Float): sigma of prior normal distribution.
+
+ """
+ __constants__ = ['prior_mu', 'bias', 'in_features', 'out_features']
+
+ def __init__(self, prior_mu, prior_sigma, in_features, out_features, bias=True):
+ super(BayesLinear, self).__init__()
+ self.in_features = in_features
+ self.out_features = out_features
+
+ self.prior_mu = prior_mu
+ self.prior_log_sigma = nn.Parameter(torch.ones(1)*np.log(prior_sigma), requires_grad=True)
+
+ self.weight_mu = nn.Parameter(torch.Tensor(out_features, in_features))
+ self.weight_log_sigma = nn.Parameter(torch.Tensor(out_features, in_features))
+ self.register_buffer('weight_eps', None)
+
+ if bias is None or bias is False:
+ self.bias = False
+ else :
+ self.bias = True
+
+ if self.bias:
+ self.bias_mu = nn.Parameter(torch.Tensor(out_features))
+ self.bias_log_sigma = nn.Parameter(torch.Tensor(out_features))
+ self.register_buffer('bias_eps', None)
+ else:
+ self.register_parameter('bias_mu', None)
+ self.register_parameter('bias_log_sigma', None)
+ self.register_buffer('bias_eps', None)
+
+ self.reset_parameters()
+
+ def reset_parameters(self):
+ stdv = 1. / np.sqrt(self.weight_mu.size(1))
+ self.weight_mu.data.uniform_(-stdv, stdv)
+ self.weight_log_sigma.data.fill_(self.prior_log_sigma.detach()[0])
+ if self.bias :
+ self.bias_mu.data.uniform_(-stdv, stdv)
+ self.bias_log_sigma.data.fill_(self.prior_log_sigma.detach()[0])
+
+ def freeze(self) :
+ self.weight_eps = torch.randn_like(self.weight_log_sigma)
+ if self.bias :
+ self.bias_eps = torch.randn_like(self.bias_log_sigma)
+
+ def unfreeze(self) :
+ self.weight_eps = None
+ if self.bias:
+ self.bias_eps = None
+
+ def forward(self, input):
+ r"""
+ Overriden.
+ """
+ if self.weight_eps is None :
+ weight = self.weight_mu + torch.exp(self.weight_log_sigma) * torch.randn_like(self.weight_log_sigma)
+ else :
+ weight = self.weight_mu + torch.exp(self.weight_log_sigma) * self.weight_eps
+
+ if self.bias:
+ if self.bias_eps is None :
+ bias = self.bias_mu + torch.exp(self.bias_log_sigma) * torch.randn_like(self.bias_log_sigma)
+ else :
+ bias = self.bias_mu + torch.exp(self.bias_log_sigma) * self.bias_eps
+ else :
+ bias = None
+
+ return F.linear(input, weight, bias)
+
+ def extra_repr(self):
+ r"""
+ Overriden.
+ """
+ return 'prior_mu={}, prior_sigma={}, in_features={}, out_features={}, bias={}'.format(self.prior_mu, self.prior_sigma, self.in_features, self.out_features, self.bias is not None)
+
+def _kl_loss(mu_0, log_sigma_0, mu_1, log_sigma_1) :
+ """
+ An method for calculating KL divergence between two Normal distribtuion.
+
+ Arguments:
+ mu_0 (Float) : mean of normal distribution.
+ log_sigma_0 (Float): log(standard deviation of normal distribution).
+ mu_1 (Float): mean of normal distribution.
+ log_sigma_1 (Float): log(standard deviation of normal distribution).
+
+ """
+ if isinstance(log_sigma_1, float):
+ sigma_1 = np.exp(log_sigma_1)
+ else:
+ sigma_1 = torch.exp(log_sigma_1)
+ kl = log_sigma_1 - log_sigma_0 + \
+ (torch.exp(log_sigma_0)**2 + (mu_0-mu_1)**2)/(2*sigma_1**2) - 0.5
+ return kl.sum()
+
+class BKLLoss(nn.Module):
+ """
+ Loss for calculating KL divergence of baysian neural network model.
+
+ Args:
+ reduction (string, optional): Specifies the reduction to apply to the output:
+ ``'mean'``: the sum of the output will be divided by the number of
+ elements of the output.
+ ``'sum'``: the output will be summed.
+ last_layer_only (Bool): True for return only the last layer's KL divergence.
+ """
+ __constants__ = ['reduction']
+
+ def __init__(self, reduction='mean', last_layer_only=False):
+ super(BKLLoss, self).__init__()
+ self.last_layer_only = last_layer_only
+ self.reduction = reduction
+
+ def forward(self, model):
+ """
+ Args:
+ model (nn.Module): a model to be calculated for KL-divergence.
+ """
+ #return bayesian_kl_loss(model, reduction=self.reduction, last_layer_only=self.last_layer_only)
+ device = torch.device("cuda" if next(model.parameters()).is_cuda else "cpu")
+ kl = torch.Tensor([0]).to(device)
+ kl_sum = torch.Tensor([0]).to(device)
+ n = torch.Tensor([0]).to(device)
+
+ for m in model.modules() :
+ if isinstance(m, (BayesLinearEmpiricalPrior)):
+ kl = _kl_loss(m.weight_mu, m.weight_log_sigma, m.prior_mu, m.prior_log_sigma_w)
+ kl_sum += kl
+ n += len(m.weight_mu.view(-1))
+
+ if m.bias :
+ kl = _kl_loss(m.bias_mu, m.bias_log_sigma, m.prior_mu, m.prior_log_sigma_b)
+ kl_sum += kl
+ n += len(m.bias_mu.view(-1))
+ if isinstance(m, (BayesLinear)):
+ kl = _kl_loss(m.weight_mu, m.weight_log_sigma, m.prior_mu, m.prior_log_sigma)
+ kl_sum += kl
+ n += len(m.weight_mu.view(-1))
+
+ if m.bias :
+ kl = _kl_loss(m.bias_mu, m.bias_log_sigma, m.prior_mu, m.prior_log_sigma)
+ kl_sum += kl
+ n += len(m.bias_mu.view(-1))
+
+ if self.last_layer_only or n == 0 :
+ return kl
+
+ if self.reduction == 'mean':
+ return kl_sum/n
+ elif self.reduction == 'sum':
+ return kl_sum
+ else:
+ raise ValueError(f"{self.reduction} is not valid")
+
class AverageMeter(object):
"""Computes and stores the average and current value"""
def __init__(self, name, fmt=':f'):
@@ -49,6 +295,7 @@ class ProgressMeter(object):
return '[' + fmt + '/' + fmt.format(num_batches) + ']'
+
class BNN(nn.Module):
"""
A model Bayesian Neural network.
@@ -61,13 +308,13 @@ class BNN(nn.Module):
between the prediction and the true value. The standard deviation of the Gaussian is left as a
parameter to be fit: sigma.
"""
- def __init__(self, input_dimension: int=1, output_dimension: int=1):
+ def __init__(self, input_dimension: int=1, output_dimension: int=1, rvm: bool=False):
super(BNN, self).__init__()
hidden_dimension = 50
# controls the aleatoric uncertainty
self.log_isigma2 = nn.Parameter(-torch.ones(1, output_dimension)*np.log(0.1**2), requires_grad=True)
# controls the weight hyperprior
- self.log_ilambda2 = nn.Parameter(-torch.ones(1)*np.log(0.1**2), requires_grad=True)
+ self.log_ilambda2 = -np.log(0.1**2)
# inverse Gamma hyper prior alpha and beta
#
@@ -82,8 +329,8 @@ class BNN(nn.Module):
# and the only regularization is to prevent the weights from becoming > 18 + 3 sqrt(var) ~= 50, making this a very loose regularization.
# An alternative would be to set the (alpha, beta) both to very low values, whichmakes the hyper prior become closer to the non-informative Jeffrey's prior.
# Using this alternative (ie: (0.1, 0.1) for the weights' hyper prior) leads to very large lambda and numerical issues with the fit.
- self.alpha_lambda = 0.001
- self.beta_lambda = 0.001
+ self.alpha_lambda = 0.0001
+ self.beta_lambda = 0.0001
# Hyperprior choice on the likelihood noise level:
# The likelihood noise level is controlled by sigma in the likelihood and it should be allowed to be very broad, but different
@@ -92,20 +339,45 @@ class BNN(nn.Module):
# Making both alpha and beta small makes the gamma distribution closer to the Jeffey's prior, which makes it non-informative
# This seems to lead to a larger training time, though.
# Since, after standardization, we know to expect the variance to be of order (1), we can select also alpha and beta leading to high variance in this range
- self.alpha_sigma = 0.001
- self.beta_sigma = 0.001
-
- self.model = nn.Sequential(
- bnn.BayesLinear(prior_mu=0.0,
- prior_sigma=torch.exp(-0.5*self.log_ilambda2),
+ self.alpha_sigma = 0.0001
+ self.beta_sigma = 0.0001
+
+ if rvm:
+ self.model = nn.Sequential(
+ BayesLinearEmpiricalPrior(prior_mu=0.0,
+ prior_sigma=np.exp(-0.5*self.log_ilambda2),
+ in_features=input_dimension,
+ out_features=hidden_dimension),
+ nn.ReLU(),
+ BayesLinearEmpiricalPrior(prior_mu=0.0,
+ prior_sigma=np.exp(-0.5*self.log_ilambda2),
+ in_features=hidden_dimension,
+ out_features=output_dimension)
+ )
+ else:
+ self.model = nn.Sequential(
+ BayesLinear(prior_mu=0.0,
+ prior_sigma=np.exp(-0.5*self.log_ilambda2),
in_features=input_dimension,
out_features=hidden_dimension),
- nn.ReLU(),
- bnn.BayesLinear(prior_mu=0.0,
- prior_sigma=torch.exp(-0.5*self.log_ilambda2),
+ nn.ReLU(),
+ BayesLinear(prior_mu=0.0,
+ prior_sigma=np.exp(-0.5*self.log_ilambda2),
in_features=hidden_dimension,
out_features=output_dimension)
- )
+ )
+ self.rvm = rvm
+
+ def prune(self):
+ """Prune weights."""
+ with torch.no_grad():
+ for layer in self.model.modules():
+ if isinstance(layer, BayesLinearEmpiricalPrior):
+ log_isigma2 = -2.0*layer.prior_log_sigma_w
+ isigma2 = torch.exp(log_isigma2)
+ keep = isigma2 < 1e4
+ layer.weight_mu[~keep] *= 0.0
+ layer.weight_log_sigma[~keep] = -12.0
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
@@ -138,10 +410,21 @@ class BNN(nn.Module):
# with a standardized input, this hyperprior forces sigma to be
# on avg. 1 and it is broad enough to allow for different sigma
isigma2 = torch.exp(self.log_isigma2)
- neg_log_hyperprior_noise = self.neg_log_gamma(self.log_isigma2, isigma2, self.alpha_sigma, self.beta_sigma)
- ilambda2 = torch.exp(self.log_ilambda2)
- neg_log_hyperprior_weights = self.neg_log_gamma(self.log_ilambda2, ilambda2, self.alpha_lambda, self.beta_lambda)
- return neg_log_hyperprior_noise.sum() + neg_log_hyperprior_weights.sum()
+ neg_log_hyperprior_noise = self.neg_log_gamma(self.log_isigma2, isigma2, self.alpha_sigma, self.beta_sigma).sum()
+ if self.rvm:
+ log_ilambda2 = [-2.0*self.model[0].prior_log_sigma_w,
+ -2.0*self.model[2].prior_log_sigma_w,
+ -2.0*self.model[0].prior_log_sigma_b,
+ -2.0*self.model[2].prior_log_sigma_b
+ ]
+ else:
+ log_ilambda2 = [-2.0*self.model[0].prior_log_sigma,
+ -2.0*self.model[2].prior_log_sigma,
+ ]
+ ilambda2 = [torch.exp(k) for k in log_ilambda2]
+ neg_log_hyperprior_weights = sum(self.neg_log_gamma(log_k, k, self.alpha_lambda, self.beta_lambda).sum()
+ for log_k, k in zip(log_ilambda2, ilambda2))
+ return neg_log_hyperprior_noise + neg_log_hyperprior_weights
def aleatoric_uncertainty(self) -> torch.Tensor:
"""
@@ -154,7 +437,18 @@ class BNN(nn.Module):
"""
Get the weights precision.
"""
- return torch.exp(self.log_ilambda2[0])
+ if self.rvm:
+ log_ilambda2 = [-2.0*self.model[0].prior_log_sigma_w,
+ -2.0*self.model[2].prior_log_sigma_w,
+ -2.0*self.model[0].prior_log_sigma_b,
+ -2.0*self.model[2].prior_log_sigma_b
+ ]
+ else:
+ log_ilambda2 = [-2.0*self.model[0].prior_log_sigma,
+ -2.0*self.model[2].prior_log_sigma,
+ ]
+ ilambda2 = [torch.exp(k) for k in log_ilambda2]
+ return sum(k.mean() for k in ilambda2)/len(ilambda2)
class BNNModel(RegressorMixin, BaseEstimator):
"""
@@ -162,14 +456,15 @@ class BNNModel(RegressorMixin, BaseEstimator):
Args:
"""
- def __init__(self, state_dict=None):
+ def __init__(self, state_dict=None, rvm: bool=False):
if state_dict is not None:
Nx = state_dict["model.0.weight_mu"].shape[1]
Ny = state_dict["model.2.weight_mu"].shape[0]
- self.model = BNN(Nx, Ny)
+ self.model = BNN(Nx, Ny, rvm=rvm)
self.model.load_state_dict(state_dict)
else:
- self.model = BNN()
+ self.model = BNN(rvm=rvm)
+ self.rvm = rvm
self.model.eval()
def state_dict(self) -> Dict[str, Any]:
@@ -197,7 +492,7 @@ class BNNModel(RegressorMixin, BaseEstimator):
torch.from_numpy(weights))
# create model
- self.model = BNN(X.shape[1], y.shape[1])
+ self.model = BNN(X.shape[1], y.shape[1], rvm=self.rvm)
# prepare data loader
B = 50
@@ -218,11 +513,11 @@ class BNNModel(RegressorMixin, BaseEstimator):
weight_prior /= float(B)
# KL loss
- kl_loss = bnn.BKLLoss(reduction='sum', last_layer_only=False)
+ kl_loss = BKLLoss(reduction='sum', last_layer_only=False)
# train
self.model.train()
- epochs = 1000
+ epochs = 250
for epoch in range(epochs):
meter = {k: AverageMeter(k, ':6.3f')
for k in ('loss', '-log(lkl)', '-log(prior)', '-log(hyper)', 'sigma', 'w.prec.')}
@@ -251,6 +546,9 @@ class BNNModel(RegressorMixin, BaseEstimator):
meter['w.prec.'].update(self.model.w_precision().detach().cpu().item(), B)
progress.display(len(loader))
+ if self.rvm:
+ self.model.prune()
+
self.model.eval()
return self
diff --git a/pes_to_spec/model.py b/pes_to_spec/model.py
index 5751510b4a5e98187ffd1081849c4dba0223d2fa..44c14289b3ac237cbdf9ac29275caa7cfe89bb6e 100644
--- a/pes_to_spec/model.py
+++ b/pes_to_spec/model.py
@@ -574,7 +574,7 @@ class Model(TransformerMixin, BaseEstimator):
Set to None to perform no selection.
validation_size: Fraction (number between 0 and 1) of the data to take for
validation and systematic uncertainty estimate.
- model_type: Which model to use. "bnn" for a BNN, "ridge" for Ridge and "ard" for ARD.
+ model_type: Which model to use. "bnn" for a BNN, "bnn_rvm" for a BNN with RVM, "ridge" for Ridge and "ard" for ARD.
"""
def __init__(self,
@@ -586,11 +586,11 @@ class Model(TransformerMixin, BaseEstimator):
tof_start: Optional[int]=None,
delta_tof: Optional[int]=300,
validation_size: float=0.05,
- model_type: Literal["bnn", "ridge", "ard"]="ard",
+ model_type: Literal["bnn", "bnn_rvm", "ridge", "ard"]="ard",
):
self.high_res_sigma = high_res_sigma
# models
- self.x_select = SelectRelevantLowResolution(channels, tof_start, delta_tof, poly=(model_type not in ["bnn"]))
+ self.x_select = SelectRelevantLowResolution(channels, tof_start, delta_tof, poly=(model_type not in ["bnn", "bnn_rvm"]))
x_model_steps = list()
x_model_steps += [
('pca', PCA(n_pca_lr, whiten=True)),
@@ -606,6 +606,8 @@ class Model(TransformerMixin, BaseEstimator):
for ch in channels+['full']}
if model_type == "bnn":
self.fit_model = BNNModel()
+ elif model_type == "bnn_rvm":
+ self.fit_model = BNNModel(rvm=True)
elif model_type == "ridge":
self.fit_model = MultiOutputRidgeWithStd(BayesianRidge(n_iter=300, tol=1e-8, verbose=True), n_jobs=8)
elif model_type == "ard":
@@ -627,7 +629,7 @@ class Model(TransformerMixin, BaseEstimator):
def n_pars(self) -> float:
"""Get number of parameters."""
- if self.model_type == "bnn":
+ if self.model_type in ("bnn", "bnn_rvm"):
return sum(p.numel() for p in self.fit_model.model.parameters())
return sum(len(estimator.coef_) + 1 for estimator in self.fit_model.estimators_)
@@ -1015,7 +1017,7 @@ class Model(TransformerMixin, BaseEstimator):
joblib.dump([self.x_select,
self.x_model,
self.y_model,
- self.fit_model.state_dict() if self.model_type == "bnn" else self.fit_model,
+ self.fit_model.state_dict() if self.model_type in ("bnn", "bnn_rvm") else self.fit_model,
self.channel_pca,
#self.channel_fit_model
DataHolder(dict(
@@ -1073,7 +1075,9 @@ class Model(TransformerMixin, BaseEstimator):
obj.x_model = x_model
obj.y_model = y_model
if obj.model_type == "bnn":
- obj.fit_model = BNNModel(state_dict=fit_model)
+ obj.fit_model = BNNModel(state_dict=fit_model, rvm=False)
+ elif obj.model_type == "bnn_rvm":
+ obj.fit_model = BNNModel(state_dict=fit_model, rvm=True)
else:
obj.fit_model = fit_model
obj.channel_pca = channel_pca
diff --git a/pes_to_spec/test/offline_analysis.py b/pes_to_spec/test/offline_analysis.py
index 6c530092c3464c3dd3272d1b789dca55e64b24f7..a199e8b663e5a0d9c813c9c967afbdd5729da4c6 100755
--- a/pes_to_spec/test/offline_analysis.py
+++ b/pes_to_spec/test/offline_analysis.py
@@ -131,7 +131,7 @@ def main():
parser.add_argument('-X', '--xgm', type=str, metavar='NAME', default="SA3_XTD10_XGM/XGM/DOOCS:output", help='XGM name')
parser.add_argument('-o', '--offset', type=int, metavar='INT', default=0, help='Train ID offset')
parser.add_argument('-c', '--xgm_cut', type=float, metavar='INTENSITY', default=0, help='XGM intensity threshold in uJ.')
- parser.add_argument('-T', '--model-type', type=str, metavar='TYPE', default="ard", choices=["bnn", "ridge", "ard"], help='Which model type to use.')
+ parser.add_argument('-T', '--model-type', type=str, metavar='TYPE', default="ard", choices=["bnn", "bnn_rvm", "ridge", "ard"], help='Which model type to use.')
parser.add_argument('-w', '--weight', action="store_true", default=False, help='Whether to reweight data as a function of the pulse energy to make it invariant to that.')
args = parser.parse_args()
diff --git a/pes_to_spec/test/prepare_plots.py b/pes_to_spec/test/prepare_plots.py
index 7dc0888e72f74e18164d0c4a0ac37de81b858693..12ca07cd4d98a448665ee6a1d876d0adc2d294ed 100755
--- a/pes_to_spec/test/prepare_plots.py
+++ b/pes_to_spec/test/prepare_plots.py
@@ -12,8 +12,8 @@ from matplotlib.gridspec import GridSpec
import seaborn as sns
SMALL_SIZE = 12
-MEDIUM_SIZE = 18
-BIGGER_SIZE = 24
+MEDIUM_SIZE = 22
+BIGGER_SIZE = 26
plt.rc('font', size=BIGGER_SIZE) # controls default text sizes
plt.rc('axes', titlesize=BIGGER_SIZE) # fontsize of the axes title
diff --git a/pyproject.toml b/pyproject.toml
index 86ea8d577381b8e30672742abe3d63e0d5ff09b6..9ea807b4f8cce6209cbb409d43aec2cc6e905656 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -29,7 +29,6 @@ dependencies = [
"scipy>=1.6",
"scikit-learn>=1.2.0",
"torch",
- "torchbnn",
]
[project.optional-dependencies]