Tensor Tools

What is Tensor Tools ?

Tensor Tools is a library which contain ready tensorflow layers. In Tensor Tools you can find many methods to build popular models such as ResNet. Tensor Tools is developped for speed up the coding process and provide flexibility when implement popular models.

Dependencies

• Tensorflow

Which methods that include ?

Tensor Tools contain:

• MobileNets Bottleneck Blocks
  • MobileNetV1 Block
  • MobileNetv2 Block
  • Relu 6 ( Number can be changed )
• SqueezeNet Blocks
  • Fire Module
• ResNet Blocks
  • Residual Block
• Math
  • Trainable Math Multiplication Layer
  • Element Wise Multiplication Layer

What is advantage of using tensorTools library ?

• Building models quickly
• Easy To Use
• Flexibility ( You can change some parameters in ready methods such as kernel size )

Documentation

tensorTools.mobileNetV1Block (x , depthWiseKernel = (3,3) , filters = None , useStride = False , strides = (2,2))

Customizable MobileNet V1 Block

x( Mandatory )

• Must be a tensor with at least 4 dimension Example: (None , 32 , 32 , 3)

depthWiseKernel ( Arbitary )

• Default value is same with original article

filters ( Arbitary )

• Must be an integer , default value equal to double of x's channel

useStride ( Arbitary )

• If True , stride will apply to Depthwise Convolution layer

strides ( Arbitary )

• If useStride is False , this is unnecessary. If useStride is True, stride value can be changeable

---

tensorTools.mobileNetV2Block (x , outputChannel , depthWiseKernel = (3,3) , t = 6 , useStride = True , strides = (2,2))

Customizable MobileNet V2 Bottleneck Block

x ( Mandatory )

• Must be a tensor with at least 4 dimension Example: (None , 32 , 32 , 3)

outputChannel ( Mandatory )

• Must be an integer , specify the output channel size

depthWiseKernel ( Arbitary )

• Default value is same with original article

filters ( Arbitary )

• Must be an integer , default value equal to double of x's channel

t ( Arbitary )

• T is used for calculate first expand convolution layer (expand = x's channel * t) , default value equal to 6

useStride ( Arbitary )

• If True , stride will apply to Depthwise Convolution layer

strides ( Arbitary )

• If useStride is False , this is unnecessary. If useStride is True, stride value can be changeable

---

tensorTools.ReluN ( n )

ReluN is made for mobileNet Relu6 Activation Layer , but you can change the number with this method

n ( Mandatory )

• If you want to use default Relu6 Activation, send 6 to parameter n , else you can specify n value

---

tensorTools.fireModule ( x , squeeze = 16 , expand = 64 , kernels = ( (1,1) , (3,3) ) , bypass = False , complexBypass = False )

Customizable SqueezeNet Fire Module

x ( Mandatory )

• Must be a tensor with at least 4 dimension Example: (None , 32 , 32 , 3)

squeeze ( Arbitary )

• Must be an integer , specify the filter amount of squeeze layer

expand ( Arbitary )

• Must be an integer , specify the filter amount of expand layer

kernels ( Arbitary )

• Must be an list/tuple/array with 2 length , specify the kernels of squeeze and expand layer, default value is same with the article

bypass ( Arbitary )

• Default is False , apply fire module with bypass method

complexBypass ( Arbitary )

• Default is False , apply fire module with complex bypass method

---

tensorTools.residualBlock ( x , blockType = 'type2' , filters = None , filKerSelection = 'auto' , kernels = None , useStrideFirst = False , strides = (2,2))

Customizable ResNet Residual Block

x ( Mandatory )

• Must be a tensor with at least 4 dimension Example: (None , 32 , 32 , 3)

blockType ( Arbitary )

• Options: type1/type2/custom, 'type1' is residual block with 2 convolutional layer , 'type2' is residual block with 3 convolutional layer , if blockType is 'custom' , both kernels and filters must be given (length of kernels and length of filters must be equal)

filters ( Arbitary )

• Must be a list/tuple/array, if blockType is 'type1' , length of filters must be 2 , if blockType is 'type2' , length of filters must be 3.

NOTE: If filKerSelection is 'auto' you can't change the filters you must change filKerSelection to 'manual'

kernels ( Arbitary )

• Must be a list/tuple/array, if blockType is 'type1' , length of kernels must be 2 , if blockType is 'type2' , length of kernels must be 3.

NOTE: If filKerSelection is 'auto' you can't change the filters you must change filKerSelection to 'manual'

filKerSelection ( Arbitary )

• Default is 'auto' , options: auto/manual

useStrideFirst ( Arbitary )

• Default is False , if True , strides will apply to first convolutional layer

strides ( Arbitary )

• Default is (2,2) , if useStrideFirst is False , it's unnecessary