functional-zoo

Model definitions and pretrained weights for PyTorch and Tensorflow

PyTorch, unlike lua torch, has autograd in it's core, so using modular structure of torch.nn modules is not necessary, one can easily allocate needed Variables and write a function that utilizes them, which is sometimes more convenient. This repo contains model definitions in this functional way, with pretrained weights for some models.

Weights are serialized as a dict of arrays in hdf5, so should be easily loadable in other frameworks. Thanks to @edgarriba we have cpp_parser for loading weights in C++.

More models coming! We also plan to add definitions for other frameworks in future, probably tiny-dnn first. Contributions are welcome.

See also imagenet classification with PyTorch demo.ipynb

Models

All models were validated to produce reported accuracy using imagenet-validation.py script (depends on OpenCV python bindings).

To load weights in Python first do pip install hickle, then:

import hickle as hkl
weights = hkl.load('resnet-18-export.hkl')

Unfortunately, hickle py3 support is not yet ready, so the models will be resaved in torch pickle format with torch.utils.model_zoo.load_url support, e.g.:

weights = model_zoo.load_url('https://s3.amazonaws.com/modelzoo-networks/wide-resnet-50-2-export-5ae25d50.pth')

Also, make_dot was moved to a separate package: PyTorchViz

Folded

Models below have batch_norm parameters and statistics folded into convolutional layers for speed. It is not recommended to use them for finetuning.

ImageNet

model	notebook	val error	download	size
VGG-16	vgg-16.ipynb	30.09, 10.69	url coming	528 MB
NIN	nin-export.ipynb	32.96, 12.29	url	33 MB
ResNet-18 (fb)	resnet-18-export.ipynb	30.43, 10.76	url	42 MB
ResNet-18-AT	resnet-18-at-export.ipynb	29.44, 10.12	url	44.1 MB
ResNet-34 (fb)	resnet-34-export.ipynb	26.72, 8.74	url	78.3 MB
WRN-50-2	wide-resnet-50-2-export.ipynb	22.0, 6.05	url	246 MB

Fast Neural Style

Notebook: fast-neural-style.ipynb

Models:

model	download	size
candy.hkl	url	7.1 MB
feathers.hkl	url	7.1 MB
wave.hkl	url	7.1 MB

Models with batch normalization

Coming