Page MenuHomec4science
Files F60148732

VAE_Main.py
No OneTemporary
Actions

File Metadata

Created
Sat, Apr 27, 21:06

VAE_Main.py
View Options

from vrae import VRAE
from utils import *
from Material_processing import *
import torch
import numpy as np
import pandas as pd
from torch.nn import functional as F
import matplotlib.pyplot as plt
import seaborn as sns
from torch import optim, cuda
#%%
dload = './model_VAE_dir' #download directory
torch.cuda.empty_cache()
#%%
hidden_size = 27
hidden_layer_depth = 1
latent_length = 27
batch_size = 100
learning_rate = 0.001 #0.005
n_epochs = 300
dropout_rate = 0.2
optimizer = 'Adam' # options: ADAM, SGD
cuda = True # options: True, False
print_every=5
clip = True # options: True, False
max_grad_norm=5
loss = 'MSELoss' # options: SmoothL1Loss, MSELoss
#%%
windowsize=5000
Material1 = "Inconel"
classname="Conduction mode"
class_1 = 'Balling'
class_2 = 'LoF pores'
class_3 = 'Conduction mode'
class_4 = 'Keyhole pores'
#%%
# Whole dataset preparation
df=MaterialTsne(Material1, windowsize)
df.target.value_counts()
class_names = [class_1,class_2,class_3,class_4]
#%%
# Display weightage of different classes
sns.set(font_scale = 1)
sns.set_style("whitegrid", {'axes.grid' : False})
graphname=str(Material1)+'_weightage'+'.png'
fig, ax = plt.subplots(figsize=(7,5), dpi=100)
ax = sns.countplot(df.target,palette=["#fbab17", "#0515bf", "#10a310", "#e9150d"])
ax.set_xticklabels(class_names);
ax.xaxis.label.set_size(10)
plt.savefig(graphname,bbox_inches='tight',pad_inches=0.1,dpi=800)
plt.show()
plt.clf()
#%%
sns.set(font_scale = 1.5)
sns.set_style("whitegrid", {'axes.grid' : False})
#%%
#Plot moving average visualisation
colour = ["#fbab17", "#0515bf", "#10a310", "#e9150d"]
graphname=str(Material1)+' moving average visualisation'+'.png'
classes = df.target.unique()
classes=np.sort(classes)
fig, axs = plt.subplots(
nrows=2,
ncols=2,
sharey=False,
figsize=(15, 7),
dpi=800
)
for i, cls in enumerate(classes):
ax = axs.flat[i]
data = df[df.target == cls] .drop(labels='target', axis=1) .mean(axis=0) .to_numpy()
plot_time_series(data, class_names[i], ax,colour[i],i)
fig.tight_layout();
plt.savefig(graphname,bbox_inches='tight',pad_inches=0.1)
plt.show()
plt.clf()
#%%
Material_1 = Normal_Regime(Material1, classname,windowsize) #Select the normal regimes
Material_1_database,Material_1_labels=dataprocessing (Material_1) #Data and labels split
Material_1_train, sequence_length, number_of_features = create_dataset(Material_1_database) #Tensor
#%% Model training
vrae = VRAE(sequence_length=sequence_length,
number_of_features = number_of_features,
hidden_size = hidden_size,
latent_length = latent_length,
batch_size = batch_size,
learning_rate = learning_rate,
n_epochs = n_epochs,
dropout_rate = dropout_rate,
optimizer = optimizer,
cuda = cuda,
print_every=print_every,
clip=clip,
max_grad_norm=max_grad_norm,
loss = loss,
dload = dload)
train_loss, _=vrae.fit(Material_1_train)
PATH = 'vrae_VAE.pth'
vrae.save(PATH)
#%%
fig, ax = plt.subplots(figsize=(7,5), dpi=100)
ax.plot(train_loss,color='red')
plt.ylabel('Loss')
plt.xlabel('Epochs')
plt.legend(['VAE training loss'])
plt.title('Training loss')
plt.savefig('Training loss.png',dpi=800)
plt.show();
plt.clf()
#%% Threshold and Reconstruction Error distribution
dataset=MaterialTemplate(Material1,windowsize)
losses_three=Resconstruction_distribution(dataset,class_3,vrae)
scores_normal = np.asarray(losses_three)
normal_avg, normal_std = np.average(scores_normal), np.std(scores_normal)
Threshold = normal_avg + (normal_std * 3)
print('Threshold:',Threshold)
losses_three=classabnormaliy(dataset,class_3,vrae,Threshold,color="#10a310")
losses_one=classabnormaliy(dataset,class_1,vrae,Threshold,color="#fbab17")
losses_two=classabnormaliy(dataset,class_2,vrae,Threshold,color="#0515bf")
losses_four=classabnormaliy(dataset,class_4,vrae,Threshold,color="#e9150d")
#%%
Normal = Normal_Regime(Material1, classname,windowsize)
Normal,Normal_labels=dataprocessing (Normal)
Normal, sequence_length, number_of_features = create_dataset(Normal)
Anomaly=MaterialTemplate(Material1,windowsize)
Anomaly = Anomaly.sample(frac=1.0)
class_name = str(Material1)+'_'+classname
Anomaly = Anomaly[Anomaly.target != str(class_name)]
Anomaly, Anomaly_labels=dataprocessing (Anomaly)
Anomaly, sequence_length, number_of_features = create_dataset(Anomaly)
fig, axs = plt.subplots(
nrows=2,
ncols=5,
sharey=True,
sharex=True,
figsize=(18, 8)
)
for i, data in enumerate(Normal[:5]):
plot_prediction(data, vrae, title='Normal', ax=axs[0, i])
for i, data in enumerate(Anomaly[:5]):
plot_prediction(data, vrae, title='Anomaly', ax=axs[1, i])
fig.tight_layout();
plt.savefig('Normal and Anomaly.png',dpi=800)
plt.show()
plt.clf()
#%%
df3,one=boxplotsupport(losses_one,'Balling')
df4,two=boxplotsupport(losses_two,'LoF')
df5,three=boxplotsupport(losses_three,'Nopores')
df6,four=boxplotsupport(losses_four,'Keyhole')
categories=np.concatenate((df3,df4,df5,df6), axis=0)
classvalue=np.concatenate((one,two,three,four), axis=0)
categories = pd.DataFrame(categories)
classvalue = pd.DataFrame(classvalue)
boxplot_data = [classvalue, categories]
boxplot_data = pd.concat(boxplot_data,axis=1)
boxplot_data.columns = ['losses', 'classes']
plt.figure()
sns.boxplot(x = 'classes', y = 'losses', data = boxplot_data,linewidth=0.5,palette=["#fbab17", "#0515bf", "#10a310", "#e9150d"])
plt.savefig('losses_boxplot.png',bbox_inches='tight',pad_inches=0.1,dpi=200)
plt.show()
#%%
count_parameters(vrae)
#786736

Event Timeline

c4science · Help