project-01-dlwxmon-go

project-01-dlwxmon-go created by GitHub Classroom

1.Packages Requirements

We use several packages:
pickle, numpy, PIL, skimage, torch, sklearn, math, matplotlib

import pickle
import numpy as np
import matplotlib.pyplot as plt
import math

from PIL import Image
from skimage.color import rgb2gray
import skimage.filters as filt

import skimage
from skimage import filters

from skimage.morphology import skeletonize, thin
import torch.optim as optim
import torch.nn.functional as F

from sklearn.model_selection import train_test_split

import torch 
import torch.nn as nn
from torch.autograd import Variable
import torch.utils.data as Data
from torch.utils.data import Dataset
from torch.optim.lr_scheduler import MultiStepLR

from sklearn.ensemble import IsolationForest

from sklearn.svm import OneClassSVM

3. Files Included

0: Please put all files in same file or you have to add path.

1 checkpoing.pt : the parameters are saved in this file

using code below to load file

  model.load_state_dict(torch.load('checkpoint.pt')) 

2 iso_train.pickle : this file stores the model for IsolationForests to classify unknow class as -1 laebl.

with open('iso_train.pickle', 'rb') as f:
  clf = pickle.load(f)

3 svm_train.pickle

with open('svm_train.pickle', 'wb') as f:
  clf1= pickle.load(clf1,f)

4. train.py

5. test.py

2. Input Format

1 Input format the same as train_data from last assignment.

2 ATTENTION： If you want to have a prediction on specific image,let us say,

train_data[1]

please input as format:

train_data[1:2] 

Because we defalutly treat input data as a collection of images.!!

3 Load data：

train_data = load_pkl('train_data.pkl')
**PLEASE USING　VARIABLE test_data**
test_data = load_pkl('PLEASE FILL FILE') 

3.Parameters

1. modification split size

train_data_, validation_data_, train_labels_, validation_labels_ = train_test_split(train_data,train_labels, test_size=0.2,shuffle=True)

2. SGD with momentum=0.5

optimizer = optim.SGD(model.parameters(), lr=learningRate, momentum=momentum)

3. Lr,epochs,loss function

#learningRate = 0.02 with step dicreasing by 0.7 at epochs [[6,15,20,30]]：
scheduler = MultiStepLR(optimizer, milestones=[6,15,20,30], gamma=0.7)

epoches = 50  #fixed
batch_size = 80 # fixed

#Weighed Cross Entropy Loss Function with weights =[1.4,1.4,0.8,0.8,1,0.9,0.9,1]  
lossCriterion = nn.CrossEntropyLoss(weight=torch.FloatTensor([1.4,1.4,0.8,0.8,1,0.9,0.9,1])) 

# train data to get parameters
train_validate(0.02,patience=4)

4. CNN structure

(channels,size)
Input :
(1,32,32)

after conv1:
(16,32,32)

after Max Pooling :
(16,16,16)

after conv2:
(32,16,16)

after Max Pooling:
(32,8,8)

full_connected layer1:
in:32*8*8,
out:64

full_connected layer_2:
in:64
out:8

And there are BN layers before every Relu Activation Function.

5. Isolation Forests

    clf=IsolationForest(n_estimators=140,behaviour='new',max_samples='auto',contamination=0.001,max_features=5)

6. One Class SVM

    clf1 = OneClassSVM(kernel='rbf',tol=0.01,nu=0.001，gamma='auto')