How To Optimize Single Object Image Classification Model Accuracy using Tensorflow 2

Last updated: 01/12/2021

Table : Comparative Analysis of Classification Accuracy

activation function	loss function	Optimizer	learning rate	learning rate dacay	epoch	total loss	learning time	final accuracy
Relu	MSE	Gradient descent	0.01	every 50, *=0.96	100	0.001	2h	98%

Table : Procedure

index	1	2	3
1	layer	normalization	activation function
2	normalization	activation function	convolution layer

---

Table : Data Augmentation

Method	notation	code
Random erase
Cutout
MixUp
CutMix
Style transfer GAN
Mosaic
Random Croping		tf.keras.layers.experimental.preprocessing.RandomCrop() tf.image.random_crop()

Table : Loss Functin and Optimizer Equation

activation function	activation equation	loss function	loss equation	Optimizer	Optimizer notation
Sigmoid
Relu		MSE		Gradient descent
Leacky Relu
ELU
tanh
maxout

Table : Weight Initialization

Weight Initialization	notation	code
Random Normal		tf.keras.initializers.RandomNormal()
Xavier (=Glorot)		tf.keras.glorot_uniform()
He (for Relu)		tf.keras.initializers.he_uniform()

Table : Regularization

Regularization	notation	code
Dropout	Random Node Turn off	tf.keras.layers.Dropout(rate) (0.0<rate<1.0)
GaussianDropout	sqrt(rate / (1 - rate))	tf.keras.layers.GaussianDropout(rate)
DropBlock (for CNN)	drop range of features
Spatial Dropout		tf.keras.layers.SpatialDropout2D()
L1 Regularization
L2 Regularization
Early stoppoing	stop epoch training	tf.keras.callbacks.EarlyStopping()

Table : Normalization - Internal Covariate Shift Solution

Method	notation	code
Batch Normalization(BN) [Not use Dropout]		tf.keras.layers.BatchNormalization()
Layer Normalization(LN)		tf.keras.layers.LayerNormalization()
Instance Normalization(IN)
Group Normalization(GN) [for small batch sizes]
Switchable Normalization(SN) [for small batch sizes]