elite-sheep / vgg16

A vgg16 implementation with tensorflow-python

VGG16

This a implementation of vgg16 with tensorflow and python, study-oriented.

Overview of vgg16

vgg16 is an important convolutional neural network posted by Karen Simonyan. The link of their essay is here below:
https://arxiv.org/abs/1409.1556

Usage

Train

To train vgg16 with this project, you only need to provide a path. The path should contain a file named "model.json" and this JSON file contains all information to train a vgg16 network with this project.

A sample model.json is shown below:

{
	"learning_rate":0.01,
	"momentum":0.9,
	"batchsize":8,
	"batches":70,
	"channel":3,
	"classes":2,
	"classnamelist":"labels.txt",
	"trainlist":"train.txt",
	"labellist":"labellist.txt"
}

The classnamelist is a file that contains all names of classes. The number of classes should be equal to the number of classes that provides in "model.json".

The trainlist file contains the absolute path to all pictures in the training set.
The labellist file contains the index of class of a picture, corresponding to the index of pictures in the trainlist.

To train a vgg16 using this project:

#!coding=utf-8

from src.vgg16 import Vgg16

def main():
		#init a new Vgg16 instance
    net = Vgg16()
		#load json file
    net.loadWithUntrainedJson(srcDir="./example")
		#train CNN, the model will be deployed in tgtDir
    net.train(tgtDir='./example')

if __name__ == '__main__':
    main()

Predict

We provide a predict API to predict from a image list. The API is shown below:

def predict(self, tgtDir, imageListFile)

Before you use the model to predict, the model should be loaded with the following API:

def loadWithTrainedJson(self, modelDir)

The modelDir contains the model trained with this project or created by the users themselves.

The imageListFile is the list of images to be predicted. Each line of the file contains an address of an image(we encourage absolute address).

At last, the result will dump to result.txt to tgtDir.

More

Further question, please contact [e-mail]