convblock

Artificial neural networks architectures implemented via pytorch. This library's main purpose is to simplify creation, design and experiments with convolutional neural networks. Library is build on top of pytorch framework.

Installation

git clone https://github.com/kirillemilio/convblock.git && cd convblock && pip install .

ConvBlock module

This module provides simple, easy to understand and unified way to create typical convolutional layout. For example, simple sequential module can be described in a following way:

from convblock import ConvBlock

ConvBlock(input_shape=(32, 32, 32), layout='cna cna p',
          c=dict(kernel_size=3, filters=16), 
          p=dict(kernel_size=2, stride=2, mode='max'))

Various ResNet blocks are really simple to create:

from convblock import ConvBlock

ConvBlock(input_shape=(32, 32, 32),
          layout='+ cna cna cn + a',
          c=dict(kernel_size=[1, 3, 1], filters=[8, 8, 32], stride=[1, 2, 1]),
          shortcut=dict(layout='c', kernel_size=3))

Simple DenseNet block can be created in a few lines of code:

ConvBlock(input_shape=(32, 32, 32),
          layout='. cna cna .' * 4,
          c=dict(kernel_size=[1, 3] * 4,
                 filters=[8, 32] * 4))

If you want to reuse the same block structure but in 3D space(using 3D convolutions and batch normalization) you need just to change input_shape all other modules inside ConvBlock will be adapted to spatial dimension of input. So, the following code will generate the same DenseBlock that can be used on 3D image(for example, CT scans)

from convblock import ConvBlock

ConvBlock(input_shape=(32, 32, 32, 32),
          layout='. cna cna .' * 4,
          c=dict(kernel_size=[1, 3] * 4,
                 filters=[8, 32] * 4))

More complex structures with bunch of branches can also be created and designed using just a few lines of code. For instance, here is implementation of InceptionB module using ConvBranches module that allows to describe branches of ConvBlock on par with mixing operation(., *, +):

from convblock import ConvBranches

ConvBranches(
    input_shape=(32, 32, 32), mode='.',
    branch1x1={
        'layout': 'cna',
        'c': {'filters': 192, 'kernel_size': 1}
    },
    branch_pool={
        'layout': 'p cna',
        'c': {'filters': 192, 'kernel_size': 1},
        'p': {'mode': 'avg', 'kernel_size': 3, 'stride': 1}
    },
    branch7x7={
        'layout': 'cna cna cna',
        'c': {'kernel_size': [(1, 1), (1, 7), (7, 1)],
        'filters': [128, 128, 192]}
    },
    branch7x7dbl={'layout': 'cna cna cna cna cna',
                  'c': {'kernel_size': [(1, 1), (1, 7), (7, 1), (1, 7), (7, 1)],
                        'filters': (128, 128, 128, 128, 192)}}
)

Library also contains flexibly parametrized implementation of several popular CNN architectures for image classificaton/segmentation and detection. Here is non-exhaustive list of architectures:

• VGG
• ResNet
• DenseNet
• MobileNetV1, MobileNetV2, MobileNetV3
• ShuffleNetV1, ShuffleNetV2
• SqueezeNet
• InceptionV3, InceptionV4
• InceptionResNet
• Xception
• UNet
• VNet
• SSD

It's easy to change parameters of creation of various models. For example, here is a code for changing default ResNet to SEResNext50 model

from convblock import ResNet
from convblock import Config

config = Config({
    'input_shape': (3, 256, 256),
    'num_blocks': (3, 4, 6, 3),
    'use_bottleneck': True,
    'body/cardinality': 32,
    'body/use_se': True,
    'body/shortcut/layout': 'p cna',
    'body/shortcut/pool_mode': 'avg'
})

se_resnext50 = ResNet(config)