Spaces:
Sleeping
Sleeping
File size: 8,130 Bytes
d03866e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 |
"""
This function is adapted from [usad] by [manigalati]
Original source: [https://github.com/manigalati/usad]
"""
from __future__ import division
from __future__ import print_function
import numpy as np
import math
import torch
import torch.nn.functional as F
from sklearn.utils import check_array
from sklearn.utils.validation import check_is_fitted
from torch import nn
from torch.utils.data import DataLoader
from sklearn.preprocessing import MinMaxScaler
import tqdm
from .base import BaseDetector
from ..utils.dataset import ReconstructDataset
from ..utils.torch_utility import EarlyStoppingTorch, get_gpu
class USADModel(nn.Module):
def __init__(self, feats, n_window=5):
super(USADModel, self).__init__()
self.name = 'USAD'
self.lr = 0.0001
self.n_feats = feats
self.n_hidden = 16
self.n_latent = 5
self.n_window = n_window # USAD w_size = 5
self.n = self.n_feats * self.n_window
self.encoder = nn.Sequential(
nn.Flatten(),
nn.Linear(self.n, self.n_hidden), nn.ReLU(True),
nn.Linear(self.n_hidden, self.n_hidden), nn.ReLU(True),
nn.Linear(self.n_hidden, self.n_latent), nn.ReLU(True),
)
self.decoder1 = nn.Sequential(
nn.Linear(self.n_latent,self.n_hidden), nn.ReLU(True),
nn.Linear(self.n_hidden, self.n_hidden), nn.ReLU(True),
nn.Linear(self.n_hidden, self.n), nn.Sigmoid(),
)
self.decoder2 = nn.Sequential(
nn.Linear(self.n_latent,self.n_hidden), nn.ReLU(True),
nn.Linear(self.n_hidden, self.n_hidden), nn.ReLU(True),
nn.Linear(self.n_hidden, self.n), nn.Sigmoid(),
)
def forward(self, g):
bs = g.shape[0]
## Encode
# z = self.encoder(g.view(1,-1))
z = self.encoder(g.view(bs, self.n))
## Decoders (Phase 1)
ae1 = self.decoder1(z)
ae2 = self.decoder2(z)
## Encode-Decode (Phase 2)
ae2ae1 = self.decoder2(self.encoder(ae1))
# return ae1.view(-1), ae2.view(-1), ae2ae1.view(-1)
return ae1.view(bs, self.n), ae2.view(bs, self.n), ae2ae1.view(bs, self.n)
class USAD(BaseDetector):
def __init__(self,
win_size = 5,
feats = 1,
batch_size = 128,
epochs = 10,
patience = 3,
lr = 1e-4,
validation_size=0.2
):
super().__init__()
self.__anomaly_score = None
self.cuda = True
self.device = get_gpu(self.cuda)
self.win_size = win_size
self.batch_size = batch_size
self.epochs = epochs
self.feats = feats
self.validation_size = validation_size
self.model = USADModel(feats=self.feats, n_window=self.win_size).to(self.device)
self.optimizer = torch.optim.AdamW(
self.model.parameters(), lr=lr, weight_decay=1e-5
)
self.scheduler = torch.optim.lr_scheduler.StepLR(self.optimizer, 5, 0.9)
self.criterion = nn.MSELoss(reduction = 'none')
self.early_stopping = EarlyStoppingTorch(None, patience=patience)
def fit(self, data):
tsTrain = data[:int((1-self.validation_size)*len(data))]
tsValid = data[int((1-self.validation_size)*len(data)):]
train_loader = DataLoader(
dataset=ReconstructDataset(tsTrain, window_size=self.win_size),
batch_size=self.batch_size,
shuffle=True
)
valid_loader = DataLoader(
dataset=ReconstructDataset(tsValid, window_size=self.win_size),
batch_size=self.batch_size,
shuffle=False
)
l1s, l2s = [], []
for epoch in range(1, self.epochs + 1):
self.model.train(mode=True)
n = epoch + 1
avg_loss = 0
loop = tqdm.tqdm(
enumerate(train_loader), total=len(train_loader), leave=True
)
for idx, (d, _) in loop:
d = d.to(self.device) # (bs, win, feat)
# print('d: ', d.shape)
ae1s, ae2s, ae2ae1s = self.model(d)
# print('ae2ae1s: ', ae2ae1s.shape)
d = d.view(ae2ae1s.shape[0], self.feats*self.win_size)
l1 = (1 / n) * self.criterion(ae1s, d) + (1 - 1/n) * self.criterion(ae2ae1s, d)
l2 = (1 / n) * self.criterion(ae2s, d) - (1 - 1/n) * self.criterion(ae2ae1s, d)
# print('l1: ', l1.shape)
l1s.append(torch.mean(l1).item())
l2s.append(torch.mean(l2).item())
loss = torch.mean(l1 + l2)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
avg_loss += loss.cpu().item()
loop.set_description(f"Training Epoch [{epoch}/{self.epochs}]")
loop.set_postfix(loss=loss.item(), avg_loss=avg_loss / (idx + 1))
if len(valid_loader) > 0:
self.model.eval()
avg_loss_val = 0
loop = tqdm.tqdm(
enumerate(valid_loader), total=len(valid_loader), leave=True
)
with torch.no_grad():
for idx, (d, _) in loop:
d = d.to(self.device)
ae1s, ae2s, ae2ae1s = self.model(d)
d = d.view(ae2ae1s.shape[0], self.feats*self.win_size)
l1 = (1 / n) * self.criterion(ae1s, d) + (1 - 1/n) * self.criterion(ae2ae1s, d)
l2 = (1 / n) * self.criterion(ae2s, d) - (1 - 1/n) * self.criterion(ae2ae1s, d)
l1s.append(torch.mean(l1).item())
l2s.append(torch.mean(l2).item())
loss = torch.mean(l1 + l2)
avg_loss_val += loss.cpu().item()
loop.set_description(
f"Validation Epoch [{epoch}/{self.epochs}]"
)
loop.set_postfix(loss=loss.item(), avg_loss_val=avg_loss_val / (idx + 1))
self.scheduler.step()
if len(valid_loader) > 0:
avg_loss = avg_loss_val / len(valid_loader)
else:
avg_loss = avg_loss / len(train_loader)
self.early_stopping(avg_loss, self.model)
if self.early_stopping.early_stop:
print(" Early stopping<<<")
break
def decision_function(self, data):
test_loader = DataLoader(
dataset=ReconstructDataset(data, window_size=self.win_size),
batch_size=self.batch_size,
shuffle=False
)
self.model.eval()
scores = []
loop = tqdm.tqdm(enumerate(test_loader), total=len(test_loader), leave=True)
with torch.no_grad():
for idx, (d, _) in loop:
d = d.to(self.device)
# print('d: ', d.shape)
ae1, ae2, ae2ae1 = self.model(d)
d = d.view(ae2ae1.shape[0], self.feats*self.win_size)
# print('ae2ae1: ', ae2ae1.shape)
# print('d: ', d.shape)
loss = 0.1 * self.criterion(ae1, d) + 0.9 * self.criterion(ae2ae1, d)
# print('loss: ', loss.shape)
loss = torch.mean(loss, axis=-1)
scores.append(loss.cpu())
scores = torch.cat(scores, dim=0)
scores = scores.numpy()
self.__anomaly_score = scores
if self.__anomaly_score.shape[0] < len(data):
self.__anomaly_score = np.array([self.__anomaly_score[0]]*math.ceil((self.win_size-1)/2) +
list(self.__anomaly_score) + [self.__anomaly_score[-1]]*((self.win_size-1)//2))
return self.__anomaly_score
def anomaly_score(self) -> np.ndarray:
return self.__anomaly_score
def param_statistic(self, save_file):
pass
|