TensorFlow_VGG_train_test

Implementation of training and testing process of VGG16 in TensorFlow

• Pre-requisites
• Easy Run Testing
• Training on CIFAR-10 dataset
• Files Overview

Pre-requisites

• TensorFlow
• Python 2.7.x
• Pre-trained VGG16 model parameters VGG16_modelParams.tensorflow
• Training dataset CIFAR-10 dataset

Easy Run Testing

The testing code testing.py test on a image of weasel using the pre-traind VGG16 model parameters provided by Davi Frossard. We saved the model parameters as a tensorflow readable file VGG16_modelParams.tensorflow. Download the parameter file to the same folder and run

$ python testing.py

Testing result

Restoring VGG16 model parameters ...

Classification Result:
	Category Name: weasel 
	Probability: 88.05%

	Category Name: polecat, fitch, foulmart, foumart, Mustela putorius 
	Probability: 6.82%

	Category Name: mink 
	Probability: 4.79%

Training on CIFAR-10 dataset (in debugging)

The training code training.py shows a demo of the training process in TensorFlow. Download the CIFAR-10 dataset to the same folder and run

$ python training.py

The print information during traing will be like the follow. Of course, the results may vary with the initalization. The epoch number is set to 10 to get a quick result. For the same reason, we trained on a small model (3 layers) rather than on VGG16 model. Although the accuracy is not good enough, it roughly shows the trend of lower loss and higher accuracy with more epoch.

Loading dataset ...
Color center of images: [ 0.59276742  0.57485735  0.44807526]
Training ...
...
(GPU Info)
...
Epoch  1/10:
	Train Loss = 2.33	 Accuracy = 18.00%
	Valid Loss = 2.13	 Accuracy = 25.00%
Epoch  2/10:
	Train Loss = 2.08	 Accuracy = 20.00%
	Valid Loss = 1.95	 Accuracy = 33.30%
Epoch  3/10:
	Train Loss = 1.99	 Accuracy = 26.00%
	Valid Loss = 1.85	 Accuracy = 34.70%
Epoch  4/10:
	Train Loss = 1.80	 Accuracy = 40.00%
	Valid Loss = 1.79	 Accuracy = 37.20%
Epoch  5/10:
	Train Loss = 1.47	 Accuracy = 44.00%
	Valid Loss = 1.66	 Accuracy = 41.10%
Epoch  6/10:
	Train Loss = 1.25	 Accuracy = 54.00%
	Valid Loss = 1.61	 Accuracy = 43.60%
Epoch  7/10:
	Train Loss = 1.24	 Accuracy = 60.00%
	Valid Loss = 1.68	 Accuracy = 43.20%
Epoch  8/10:
	Train Loss = 1.06	 Accuracy = 64.00%
	Valid Loss = 1.62	 Accuracy = 46.00%
Epoch  9/10:
	Train Loss = 0.98	 Accuracy = 72.00%
	Valid Loss = 1.55	 Accuracy = 47.00%
Epoch 10/10:
	Train Loss = 0.90	 Accuracy = 76.00%
	Valid Loss = 1.60	 Accuracy = 45.40%

Files Overview

• layerConstructor.py provides higher level functions to build basic convolution, pooling, and fully connection layers.
• VGG16_model.py build the VGG16 model, respectively, using layerConstructor.py
• imagenet_classes.py stores the class names of the ImageNet dataset, cited from Davi Frossard
• weasel.png is an example image for testing
• training.py and testing.py are demos of traing and testing