Experiment to improve the faulty steel classification

Experiments with Pytorch-based Model that classified Steel plate faults into 7 types using 27 features

Data

Kaggle Faulty STeel Plates Data / .csv format

Preview head 5 features in 27 features

Y_Perimeter	Sum_of_Luminosity	Minimum_of_Luminosity	Maximum_of_Luminosity	Length_of_Conveyer
44	24220	76	108	1687

Experiments Category

• Activation Function : Sigmoid, Tanh / ReLU, Swish
• Weight Initialization : Random, Xavier(+Sigmoid), He(+ReLU)
• Optimization : SGD, Momentum, AdaGrad, RMSProp, Adam(Momentum + RMSProp)
• Overfitting Prevention : Batch Normalization, Drotout, L2-Regularization(Weight Decay)

Experimental Results

Hyphter Parameter Setting

Epoch = 1200, Learning Rate = lr, Batch Size = 128, Hidden Layer Neuron = 20, Input Size = 27, Output Size = 7

Train Data : Test Data = 1600 : 300, Totla Data = 1900

Hidden Layer의 크기를 정함에 있어서 가장 흔하게 신뢰되어 지는 것은, 'Hidden Layer의 최적의 크기는 보통 Input과 Output Layer의 크기 사이이다'

Network	Activation Function	Weight Initialzation	Optimization	Overfitting Prevention	Train Acc Test Acc	Code
2 Layer	Sigmoid	Random / std=0.01	SGD / lr=0.01	.	0.3906 0.3958	.
2 Layer	Sigmoid	Random / std=0.01	SGD / lr=0.01	Batch Normalization	0.5112 0.4692	.
2 Layer	Sigmoid	He	SGD / lr=0.01	Batch Normalization	0.51 0.5190	.
2 Layer	Sigmoid	Xavier	SGD / lr=0.01	Batch Normalization	0.5031 0.5395	.
2 Layer	Sigmoid	Xavier	Adam / lr=0.001	Batch Normalization	0.5731 0.3049	.
2 Layer	Sigmoid	Xavier	Adam / lr=0.001	Batch Normalization Drouput ratio=0.15	0.56 0.5571	.
2 Layer	Sigmoid	Xavier	AdaGrad / lr=0.01	Batch Normalization	0.5512 0.5747	.
3 Layer	Sigmoid	Xavier	SGD / lr=0.01	Batch Normalization	0.5056 0.4750	.
2 Layer	ReLU	He	SGD / lr=0.01	Batch Normalization	0.5143 0.4985	.
2 Layer	ReLU	He	Adam / lr=0.001	Batch Normalization	0.5681 0.5131	.
2 Layer	ReLU	Xavier	Adam / lr=0.001	Batch Normalization	0.585 0.5601	.
3 Layer	ReLU	He	SGD / lr=0.01	Batch Normalization	0.5131 0.5425	.