Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add Autoencoder pretraining module in FeaWAD #61

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Add Autoencoder pretraining module in FeaWAD #61

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 15 additions & 4 deletions deepod/models/tabular/feawad.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,7 +64,8 @@ class FeaWAD(BaseDeepAD):
random_state： int, optional (default=42)
the seed used by the random
"""
def __init__(self, epochs=100, batch_size=64, lr=1e-3,

def __init__(self, epochs=100, pretrain_epochs=50, batch_size=64, lr=1e-3,
rep_dim=128, hidden_dims='100,50', act='ReLU', bias=False,
margin=5.,
epoch_steps=-1, prt_steps=10, device='cuda',
Expand All @@ -76,13 +77,15 @@ def __init__(self, epochs=100, batch_size=64, lr=1e-3,
verbose=verbose, random_state=random_state
)

self.pretrain_epochs = pretrain_epochs
self.margin = margin

self.rep_dim = rep_dim
self.hidden_dims = hidden_dims
self.act = act
self.bias = bias

self.cur_epoch = None
return

def training_prepare(self, X, y):
Expand All @@ -107,6 +110,8 @@ def training_prepare(self, X, y):
}
net = FeaWadNet(**network_params).to(self.device)
criterion = FeaWADLoss(margin=self.margin)
self.cur_epoch = 0

if self.verbose >= 2:
print(net)

Expand All @@ -123,6 +128,8 @@ def training_forward(self, batch_x, net, criterion):
batch_x = batch_x.float().to(self.device)
batch_y = batch_y.to(self.device)
pred, sub_result = net(batch_x)
if self.cur_epoch <= self.pretrain_epochs:
return torch.nn.functional.mse_loss(batch_x, net.AEmodel(batch_x)[0])
loss = criterion(batch_y, pred, sub_result)
return loss

Expand All @@ -133,6 +140,9 @@ def inference_forward(self, batch_x, net, criterion):
batch_z = batch_x
return batch_z, s

def epoch_update(self):
self.cur_epoch += 1


class FeaWadNet(torch.nn.Module):
def __init__(self, n_features, network, n_hidden='500,100', n_hidden2='256,32', n_emb=20,
Expand All @@ -143,7 +153,7 @@ def __init__(self, n_features, network, n_hidden='500,100', n_hidden2='256,32',
FWmodel = get_network('MLP')
self.AEmodel = AEmodel_class(n_features, n_hidden=n_hidden, n_emb=n_emb,
activation=activation, bias=bias)
self.LinearModel = FWmodel(n_features+n_emb, n_hidden=n_hidden2, n_output=1,
self.LinearModel = FWmodel(n_features + n_emb, n_hidden=n_hidden2, n_output=1,
activation=activation, bias=bias)

def forward(self, x):
Expand Down Expand Up @@ -181,6 +191,7 @@ class FeaWADLoss(torch.nn.Module):
- If ``'sum'``: the output will be summed

"""

def __init__(self, margin=5., reduction='mean'):
super(FeaWADLoss, self).__init__()
self.margin = margin
Expand All @@ -192,8 +203,8 @@ def forward(self, y_true, y_pred, sub_result):
inlier_loss = torch.abs(dev)
outlier_loss = torch.abs(torch.maximum(self.margin - dev, torch.tensor(0.)))

sub_nor = torch.norm(sub_result, p=2, dim=1 if len(sub_result.shape)==2 else [1,2])
outlier_sub_loss = torch.abs(torch.maximum(self.margin-sub_nor, torch.tensor(0.)))
sub_nor = torch.norm(sub_result, p=2, dim=1 if len(sub_result.shape) == 2 else [1, 2])
outlier_sub_loss = torch.abs(torch.maximum(self.margin - sub_nor, torch.tensor(0.)))
loss = (1 - y_true) * (inlier_loss + sub_nor) + y_true * (outlier_loss + outlier_sub_loss)

if self.reduction == 'mean':
Expand Down