[DDesigner API] Deep-learning Designer API

1. About

1.1. DDesignerAPI?

It is a API for deep-learning learning and inference, and an API for application development using multi-platform

1.2. Functions

1.2.1. Layers and Blocks

• Accelerator enabled layers and the ability to define special layers that are not defined in Keras and others
• A function that defines a combination of layers as a block and easily composes a block (ex. CONV + BN + ACT + DROPOUT= ConvBlock)

1.2.2. Optimization for Accelerator Usage (XWN)

• Optimized function to use accelerator
  
  
  

2. Support

2.1. Platforms

• Tensorflow 2.6.0
• PyTorch 1.13.1

2.2. Components of Network

2.2.1. Layers

• Accelerator enabled layers and Custom layers that perform specific functions

2.2.1.1. Summary

2.2.1.2. Detail

• Convolution : 1D, 2D with XWN optimization
• TransposeConvolution : 1D, 2D with XWN optimization
• CascadeConvolution : A Layer that decomposes a layer with large kernel into multiple layers with smaller kernels to lighten the model / 1D, 2D with XWN optimization
  
  

2.2.2. Blocks

• A set of defined layers for user convenience

2.2.2.1. Summary

2.2.2.2. Detail

• ConvBlock : Convolution N-D Block (CONV + BN + ACT + DROPOUT), support Conv1DBlock, Conv2DBlock
• TConvBlock : Transpose Convolution 2D Block (TCONV + BN + ACT + DROPOUT), support TConv2DBlock
• CascadeConvBlock : Cascade Convolution N-D Block (CONV + BN + ACT + DROPOUT), support CascadeConv1DBlock, CascadeConv2DBlock
  
  

2.3. XWN (Applies only to convolution operations)

2.3.1. Transform Configuration (data type / default value / description)

• transform : bool / False / Choose whether to use
• bit : int / 4 / Quantization range (bit-1 ** 2)
• max_scale : float / 4.0 / Max value

2.3.2. Pruning Configuration

• pruning : bool / False / Choose whether to use
• prun_weight : float / 0.5 / Weights for puning edge generation

2.3.3. Summary

3. Command Usage

3.1. XWN

3.1.1. Single Convolution

3.1.1.1. Tensorflow

    >>> from ddesigner_api.tensorflow.xwn import tf_nn as nn
    >>> nn.conv2d(
            x,
            kernel,
            ...
            use_transform=True,
            bit=4,
            max_scale=4.0,
            use_pruning=False
        )

3.1.1.2. Keras

    >>> from ddesigner_api.tensorflow.xwn import keras_layers as klayers
    >>> klayers.Conv2D(
            2, 3, 
            ...
            use_transform=True,
            bit=4,
            max_scale=4.0
            use_pruning=True,
            prun_weight=0.5
        )

3.1.1.3. PyTorch

    >>> from ddesigner_api.pytorch.xwn import torch_nn as nn
    >>> nn.Conv2d(
            in_channels=1,
            out_channels=2,
            ...
            use_transform=True,
            bit=4,
            max_scale=4.0,
            use_pruning=False
        )

3.1.2. Custum Layer and Block (CascadeConv, ...)

3.1.2.1. Keras

    >>> from ddesigner_api.tensorflow import dpi_layers as dlayers
    >>> dlayers.CascadeConv2d(
            2, 3, 
            ...
            transform=4,
            max_scale=4.0,
            pruning=None,
        )

3.1.2.2. PyTorch

    >>> from ddesigner_api.pytorch import dpi_nn as dnn
    >>> dnn.CascadeConv2d(
            16, # in_channels 
            32, # out_channels
            7, # kernel_size
            stride=(1,1), 
            bias=False,
            ...
            transform=4,
            max_scale=4.0,
            pruning=None,
        )

3.2. Blocks

3.2.1. Keras

3.2.1.1. Conv1DBlock

    >>> from ddesigner_api.tensorflow import dpi_blocks as db
    >>> dtype='mixed_float16'
    >>> db.Conv1DBlock(
            64, 3, strides=1, padding='SAME', use_bias=False,
            activation=tf.keras.layers.ReLU(dtype=dtype), 
            batchnormalization=tf.keras.layers.BatchNormalization(dtype=dtype), 
            dtype=dtype,
            transform=4, max_scale=4.0,
            pruning=0.5
        )

3.2.1.2. Conv2DBlock

    >>> from ddesigner_api.tensorflow import dpi_blocks as db
    >>> dtype='mixed_float16'
    >>> db.Conv2DBlock(
            64, (3,3), strides=(1,1), padding='SAME', use_bias=False,
            activation=tf.keras.layers.ReLU(dtype=dtype), 
            batchnormalization=tf.keras.layers.BatchNormalization(dtype=dtype), 
            dtype=dtype,
            transform=4, max_scale=4.0,
            pruning=0.5
        )

3.2.1.3. TConv2DBlock

    >>> from ddesigner_api.tensorflow import dpi_blocks as db
    >>> dtype='mixed_float16'
    >>> db.TConv2DBlock(
            64, (3,3), strides=(2,2), padding='SAME', use_bias=False,
            activation=tf.keras.layers.ReLU(dtype=dtype), 
            batchnormalization=tf.keras.layers.BatchNormalization(dtype=dtype), 
            dtype=dtype,
            transform=4, max_scale=4.0,
            pruning=0.5
        )

3.3. Examples

• An example of comparing and printing results before optimization(XWN) and after XWN for the same input on a supported platform.

3.3.1. Tensorflow

    >>> import ddesigner_api.tensorflow.examples.examples_tensorflow as ex
    >>> ex.main()
    >>> ====== TENSORFLOW Examples======
    >>> 1: Fixed  Float32 Input Conv2D
    >>> q: Quit
    >>> Select Case: ...

3.3.2. Keras

    >>> import ddesigner_api.tensorflow.examples.examples_keras as ex
    >>> ex.main()
    >>> ====== KERAS Examples======
    >>> 1: Fixed  Float32 Input Conv2D
    >>> 2: Random Float32 Input Conv2D
    >>> 3: Random Float32 Input Conv2DTranspose
    >>> 4: Random Float16 Input Conv2D
    >>> q: Quit
    >>> Select Case: ...

3.3.3. PyTorch

    >>> import ddesigner_api.pytorch.examples.examples_pytorch as ex
    >>> ex.main()
    >>> ====== PYTORCH Examples======
    >>> 1: Fixed  Float32 Input Conv2D
    >>> 2: Random Float32 Input Conv2D
    >>> 3: Fixed  Float32 Input Conv1D
    >>> 4: Fixed  Float32 Input Conv1DTranspose
    >>> 5: Random Float32 Input CascadeConv2D
    >>> 6: Random Float32 Input CascadeConv1D
    >>> q: Quit
    >>> Select Case: ...

3.3.4. Numpy

    >>> import ddesigner_api.numpy.examples.examples_numpy as ex
    >>> ex.main()
    >>> ====== NUMPY Examples======
    >>> 1: XWN Transform
    >>> 2: XWN Transform and Pruning
    >>> q: Quit
    >>> Select Case: ...