JosieHong / CNNs_for_classifing_CIFAR10

An implement of LeNet, AlexNet, VGG16, GoogLeNet and ResNet with Tensorflow.

Home Page:https://josiehong.github.io/2018/08/13/CNN%20for%20classification%20CIFAR10/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

CNNs_for_classifing_CIFAR10

An implement of CNNs for classifing on CIFAR10 with tensorflow. In order to adapt to the size of CIFAR10, I adjusted some parameters in the network. And it's easy to fit it taining on other dataset.

  • LeNet
  • AlexNet
  • VGG16
  • GoogLeNet
  • ResNet50

Requirements

  • python 3.6.3
  • tensorflow 1.13.1
  • numpy 1.16.3
  • CIFAR10 can be download here. The path to ‘cifar-10-batches-py’ can be specified with the optional parameter ‘--dataset_dir’, which by default is placed in the root directory.

Train and Test

Here I only iterate 20 epoches (10000 steps), you can increase the number of iterations by using the last trained model to achieve higher accuracy. Besides, you can also change learning rate and steps in ‘main.py’.

# Train and test by default.
$ python main.py

# Train with optional patameters and test.
$ python main.py	--model_type	[LeNet/AlexNet/VGG16/GoogLeNet/ResNet50] 
			--dataset_dir	[Path to cifar-10-batches-py] 
			--model_dir	[A .ckpt file of pretrained model or A folder for saving model] 

Use GPU

  • CUDA 8.0.61
  • CUDNN 5.1
  • tensorflow_gpu 1.2.0
# Chose GPU to use
$ CUDA_VISIBLE_DEVICES=0 python main.py (optional patameters...)

Logs

$ tensorboard --logdir=/logs

About

An implement of LeNet, AlexNet, VGG16, GoogLeNet and ResNet with Tensorflow.

https://josiehong.github.io/2018/08/13/CNN%20for%20classification%20CIFAR10/

License:MIT License


Languages

Language:Python 100.0%