pyNetBuilder is a modular pytonic interface with builtin modules for generating popular caffe networks.
A neural network is a Directed acyclic graph (DAG) of layers. The caffe layers and the network is represented using prototxt format. As we go deeper, and add more layers or build more complex DAG's using basic layers, writing the prototxt files becomes tedious. This tool aims to provide a pytonic interface to generate prototxt files.
- It provides reusable components in form of legos from popular networks like inception, resnet, squeezenet etc.
- It provides apps which can be used to generate these popular networks for classification and object detection.
- The complexity module computes the number of parameters and connections or flops of the network generated by the app, which can be used to tweak various network architectures so that you can train networks which suit your memory / runtime needs.
- We also release some models pretrained on imagenet with few tweaks to the residual network architectures (which also demonstrate how to generate networks using pynetbuilder). Refer residual network generation , SSD generation for more details.
- Caffe (version should include the layer definitions you need for your network)
- Google protobuf
pyNetBuilder builds on top of caffe's NetSpec class to stich together a network. NetSpec "provides a way to write nets directly in Python, using a natural, functional style." Here is a basic example of writing nets in caffe. pyNetBuilder is to provide generic wrappers to attach blocks of layers to NetSpec in form of Legos.
A lego is a basic neural network building block. It has:
- Required parameters - These parameters must be specified while creating a lego, and usually change for each instantiation.
- Default parameters (which can be overridden). The default parameters can be stored and read from a config file.
- A attach method which attaches a caffe layer to a netspec object.
- BaseLegoFunction: This is a classes which can attach all the core caffe layers using functional style. The default parameters for layers can be added / updated in the default.config file. For example you can attach a convolutional layer to a netspec object as follows:
from lego.base import BaseLegoFunction conv_params = dict(name='conv1', num_output=64, kernel_size=7, use_global_stats='True', pad=3, stride=2) conv = BaseLegoFunction('Convolution',params).attach(netspec, [netspec.data])
- Hybrid - These are combinations of core legos. Example - ShortcutLego (resnets), FireLego (squeezenet), InceptionLego (google). Example - you can attach a ShortcutLego (residual networks) to a netspec object as follows:
from lego.hybrid import ShortcutLego params = dict(name='resnet_block', num_output=256, shortcut='identity', main_branch='bottleneck', stride=1, use_global_stats=false,) block = ShortcutLego(params).attach(netspec, [last_layer])
To generate a network, you can pass a network specification through a series of legos and the modules will get attached to the Netspec object.
The apps folder is a collection of python scripts which uses the pynetbuilder modules to create standard caffe network prototxt files from popluar papers. Currently apps for following networks are provided (other contributions are welcome):
- Cifar 10 training:
- Resnet - see quick getting started tutorial and some results here.
- Imagenet training:
- Resnet - We also release caffemodels trained on imagenet with various variants of Residual network architectures. See Tutorial and results for more details.
- Object detection networks - using SingleShot multibox Detector - Tutorial and results
Contributing to pyNetBuilder
Contributions to pynetbuilder are welcome.
- If you want to contribute hybrid legos used in your networks please inherit the
BaseLegoclass and write your hybrid lego inside the hybrid module. If your hybrid legos are more generic (example ssd-for object detection) you can create another module inside
pynetbuilder.legos.yourlegosand add the hybrid lego's inside the module.
- If you built a different generic app using already existing legos, you can add it to the app folder and contribute the app. Send a PR and we will review and merge it.
Code licensed under the BSD 2 clause license. See LICENSE file for terms.
pynetbuilder was written by Jay Mahadeokar from the Yahoo Vision and ML team. Special thanks to Jack Culpepper, Huy Nguyen, Pierre Garrigues, Sachin Farfade, Clayton Mellina and other members of the team for inputs and review.