Tensorflow Model Zoo for Torch7 and PyTorch (OBSOLETE)

13/07/2017: Please use the new repo pretrained-models.pytorch which includes inceptionv4 and inceptionresnetv2 with a nicer API.

This is a porting of tensorflow pretrained models made by Remi Cadene and Micael Carvalho. Special thanks to Moustapha Cissé. All models have been tested on Imagenet.

This work was inspired by inception-v3.torch.

Using pretrained models

Torch7

Requirements

Please install torchnet-vision.

luarocks install --server=http://luarocks.org/dev torchnet-vision

Models available:

• inceptionv3
• inceptionv4
• inceptionresnetv2
• resnet{18, 34, 50, 101, 152, 200}
• overfeat
• vggm
• vgg16

Simple example

require 'image'
tnt = require 'torchnet'
vision = require 'torchnet-vision'
model = vision.models.inceptionresnetv2
net = model.load()

augmentation = tnt.transform.compose{
   vision.image.transformimage.randomScale{
   	minSize = 299, maxSize = 350
   },
   vision.image.transformimage.randomCrop(299),
   vision.image.transformimage.colorNormalize{
      mean = model.mean, std  = model.std
   },
   function(img) return img:float() end
}

net:evaluate()
output = net:forward(augmentation(image.lena()))

PyTorch

Currently available in this repo only On pytorch/vision maybe!

Models available:

• inceptionv4
• inceptionresnetv2

Simple example

import torch
from inceptionv4.pytorch_load import inceptionv4
net = inceptionv4()
input = torch.autograd.Variable(torch.ones(1,3,299,299))
output = net.forward(input)

Reproducing the porting

Requirements

• Tensorflow
• Torch7
• PyTorch
• hdf5 for python3
• hdf5 for lua

Example of commands

In Tensorflow: Download tensorflow parameters and extract them in ./dump directory.

python3 inceptionv4/tensorflow_dump.py

In Torch7 or PyTorch: Create the network, load the parameters, launch few tests and save the network in ./save directory.

th inceptionv4/torch_load.lua
python3 inceptionv4/pytorch_load.py