jhnine
/
JCC-DeepOD
forked from JointCloud/JCC-DeepOD
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix bugs in TCN networks

This commit is contained in:
real-lhj 2023-09-07 19:09:40 +08:00
parent 82553c65a4
commit 83218af111
4 changed files with 33 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
deepod/core/base_model.py
View File

@ -324,9 +324,7 @@ class BaseDeepAD(metaclass=ABCMeta):
return self
def _training(self):
optimizer = torch.optim.Adam(self.net.parameters(),
lr=self.lr,
weight_decay=1e-5)
optimizer = torch.optim.Adam(self.net.parameters(), lr=self.lr, eps=1e-6)
self.net.train()
for i in range(self.epochs):

46
deepod/core/networks/ts_network_tcn.py
View File

@ -15,42 +15,42 @@ class TcnAE(torch.nn.Module):
n_hidden = [int(a) for a in n_hidden]
num_layers = len(n_hidden)
self.encoder_layers = []
encoder_layers = []
# encoder
for i in range(num_layers+1):
dilation_size = 2 ** i
padding_size = (kernel_size-1) * dilation_size
in_channels = n_features if i == 0 else n_hidden[i-1]
out_channels = n_emb if i == num_layers else n_hidden[i]
self.encoder_layers += [TcnResidualBlock(in_channels, out_channels, kernel_size,
stride=1, dilation=dilation_size,
padding=padding_size, dropout=dropout, bias=bias,
activation=activation)]
encoder_layers += [TcnResidualBlock(in_channels, out_channels, kernel_size,
stride=1, dilation=dilation_size,
padding=padding_size, dropout=dropout, bias=bias,
activation=activation)]
# decoder
decoder_n_hidden = n_hidden[::-1]
self.decoder_layers = []
decoder_layers = []
for i in range(num_layers+1):
# no dilation in decoder
in_channels = n_emb if i == 0 else decoder_n_hidden[i-1]
out_channels = n_features if i==num_layers else decoder_n_hidden[i]
dilation_size = 2 ** (num_layers-i)
padding_size = (kernel_size-1) * dilation_size
self.decoder_layers += [TcnResidualBlockTranspose(in_channels, out_channels, kernel_size,
stride=1, dilation=dilation_size,
padding=padding_size, dropout=dropout,
activation=activation)]
decoder_layers += [TcnResidualBlockTranspose(in_channels, out_channels, kernel_size,
stride=1, dilation=dilation_size,
padding=padding_size, dropout=dropout, bias=bias,
activation=activation)]
# to register parameters in list of layers, each layer must be an object
self.enc_layer_names = ["enc_" + str(num) for num in range(len(self.encoder_layers))]
self.dec_layer_names = ["dec_" + str(num) for num in range(len(self.decoder_layers))]
for name, layer in zip(self.enc_layer_names, self.encoder_layers):
setattr(self, name, layer)
for name, layer in zip(self.dec_layer_names, self.decoder_layers):
setattr(self, name, layer)
# # to register parameters in list of layers, each layer must be an object
# self.enc_layer_names = ["enc_" + str(num) for num in range(len(encoder_layers))]
# self.dec_layer_names = ["dec_" + str(num) for num in range(len(decoder_layers))]
# for name, layer in zip(self.enc_layer_names, self.encoder_layers):
# setattr(self, name, layer)
# for name, layer in zip(self.dec_layer_names, self.decoder_layers):
# setattr(self, name, layer)
self.encoder = torch.nn.Sequential(*self.encoder_layers)
self.decoder = torch.nn.Sequential(*self.decoder_layers)
self.encoder = torch.nn.Sequential(*encoder_layers)
self.decoder = torch.nn.Sequential(*decoder_layers)
def forward(self, x):
out = x.permute(0, 2, 1)
@ -61,7 +61,7 @@ class TcnAE(torch.nn.Module):
class TCNnet(torch.nn.Module):
"""TCN is adapted from https://github.com/locuslab/TCN"""
def __init__(self, n_features, n_hidden='500,100', n_output=20,
def __init__(self, n_features, n_hidden='8', n_output=20,
kernel_size=2, bias=False,
dropout=0.2, activation='ReLU'):
super(TCNnet, self).__init__()
@ -140,10 +140,10 @@ class TcnResidualBlock(torch.nn.Module):
class TcnResidualBlockTranspose(torch.nn.Module):
def __init__(self, n_inputs, n_outputs, kernel_size, stride, dilation, padding,
dropout=0.2, activation='ReLU'):
dropout=0.2, activation='ReLU', bias=False):
super(TcnResidualBlockTranspose, self).__init__()
self.conv1 = weight_norm(torch.nn.ConvTranspose1d(n_inputs, n_outputs, kernel_size,
stride=stride, padding=padding,
stride=stride, padding=padding, bias=bias,
dilation=dilation))
self.pad1 = Pad1d(padding)
@ -151,7 +151,7 @@ class TcnResidualBlockTranspose(torch.nn.Module):
self.dropout1 = torch.nn.Dropout(dropout)
self.conv2 = weight_norm(torch.nn.ConvTranspose1d(n_outputs, n_outputs, kernel_size,
stride=stride, padding=padding,
stride=stride, padding=padding, bias=bias,
dilation=dilation))
self.pad2 = Pad1d(padding)
self.act2 = _instantiate_class("torch.nn.modules.activation", activation)

4
deepod/models/time_series/tcned.py
View File

@ -10,8 +10,8 @@ from deepod.core.networks.ts_network_tcn import TcnAE
class TcnED(BaseDeepAD):
def __init__(self, seq_len=100, stride=1, epochs=10, batch_size=32, lr=1e-4,
rep_dim=32, hidden_dims=32, kernel_size=3, act='LeakyReLU', bias=False, dropout=0.2,
epoch_steps=-1, prt_steps=10, device='cuda',
rep_dim=32, hidden_dims=32, kernel_size=3, act='ReLU', bias=True, dropout=0.2,
epoch_steps=-1, prt_steps=1, device='cuda',
verbose=2, random_state=42):
super(TcnED, self).__init__(
model_name='TcnED', data_type='ts', epochs=epochs, batch_size=batch_size, lr=lr,

11
testbed/configs.yaml
View File

@ -36,10 +36,13 @@ COUTA:
batch_size: 64
TcnED:
epochs: 10
lr: 0.0001
rep_dim: 64
hidden_dims: 32
epochs: 100
lr: 0.00015
dropout: 0.42
kernel_size: 2
rep_dim: 8
hidden_dims: '8,8'
batch_size: 512
AnomalyTransformer:
lr: 0.0001