BIGBALLON/cifar-10-cnn

Play deep learning with CIFAR datasets

Convolutional Neural Networks for CIFAR-10

This repository is about some implementations of CNN Architecture for cifar10.

I just use Keras and Tensorflow to implementate all of these CNN models.
(maybe torch/pytorch version if I have time)
A pytorch version is available at CIFAR-ZOO

Requirements

• Python (3.5)
• keras (>= 2.1.5)
• tensorflow-gpu (>= 1.4.1)

Architectures and papers

• The first CNN model: LeNet
  • LeNet-5 - Yann LeCun
• Network in Network
  • Network In Network
• Vgg19 Network
  • Very Deep Convolutional Networks for Large-Scale Image Recognition
  • The 1st places in ILSVRC 2014 localization tasks
  • The 2nd places in ILSVRC 2014 classification tasks
• Residual Network
  • Deep Residual Learning for Image Recognition
  • Identity Mappings in Deep Residual Networks
  • CVPR 2016 Best Paper Award
  • 1st places in all five main tracks:
    • ILSVRC 2015 Classification: "Ultra-deep" 152-layer nets
    • ILSVRC 2015 Detection: 16% better than 2nd
    • ILSVRC 2015 Localization: 27% better than 2nd
    • COCO Detection: 11% better than 2nd
    • COCO Segmentation: 12% better than 2nd
• Wide Residual Network
  • Wide Residual Networks
• ResNeXt
  • Aggregated Residual Transformations for Deep Neural Networks
  • Used in Mask-RCNN
• DenseNet
  • Densely Connected Convolutional Networks
  • CVPR 2017 Best Paper Award
• SENet
  • Squeeze-and-Excitation Networks
  • The 1st places in ILSVRC 2017 classification tasks

Documents & tutorials

There are also some documents and tutorials in doc & issues/3.
Get it if you need.
You can also see the articles if you can speak Chinese.

Accuracy of all my implementations

In particular：
Change the batch size according to your GPU's memory.
Modify the learning rate schedule may imporve the results of accuracy!

network	GPU	params	batch size	epoch	training time	accuracy(%)
Lecun-Network	GTX1080TI	62k	128	200	30 min	76.23
Network-in-Network	GTX1080TI	0.97M	128	200	1 h 40 min	91.63
Vgg19-Network	GTX1080TI	39M	128	200	1 h 53 min	93.53
Residual-Network20	GTX1080TI	0.27M	128	200	44 min	91.82
Residual-Network32	GTX1080TI	0.47M	128	200	1 h 7 min	92.68
Residual-Network110	GTX1080TI	1.7M	128	200	3 h 38 min	93.93
Wide-resnet 16x8	GTX1080TI	11.3M	128	200	4 h 55 min	95.13
Wide-resnet 28x10	GTX1080TI	36.5M	128	200	10 h 22 min	95.78
DenseNet-100x12	GTX1080TI	0.85M	64	250	17 h 20 min	94.91
DenseNet-100x24	GTX1080TI	3.3M	64	250	22 h 27 min	95.30
DenseNet-160x24	1080 x 2	7.5M	64	250	50 h 20 min	95.90
ResNeXt-4x64d	GTX1080TI	20M	120	250	21 h 3 min	95.19
SENet(ResNeXt-4x64d)	GTX1080TI	20M	120	250	21 h 57 min	95.60

About LeNet and CNN training tips/tricks

LeNet is the first CNN network proposed by LeCun.
I used different CNN training tricks to show you how to train your model efficiently.

LeNet_keras.py is the baseline of LeNet,
LeNet_dp_keras.py used the Data Prepossessing [DP],
LeNet_dp_da_keras.py used both DP and the Data Augmentation[DA],
LeNet_dp_da_wd_keras.py used DP, DA and Weight Decay [WD]

network	GPU	DP	DA	WD	training time	accuracy(%)
LeNet_keras	GTX1080TI	-	-	-	5 min	58.48
LeNet_dp_keras	GTX1080TI	√	-	-	5 min	60.41
LeNet_dp_da_keras	GTX1080TI	√	√	-	26 min	75.06
LeNet_dp_da_wd_keras	GTX1080TI	√	√	√	26 min	76.23

For more CNN training tricks, see Must Know Tips/Tricks in Deep Neural Networks (by Xiu-Shen Wei)

About Learning Rate schedule

Different learning rate schedule may get different training/testing accuracy!
See ./htd, and HTD for more details.

About Multiple GPUs Training

Since the latest version of Keras is already supported keras.utils.multi_gpu_model, so you can simply use the following code to train your model with multiple GPUs:

from keras.utils import multi_gpu_model
from keras.applications.resnet50 import ResNet50

model = ResNet50()

# Replicates `model` on 8 GPUs.
parallel_model = multi_gpu_model(model, gpus=8)
parallel_model.compile(loss='categorical_crossentropy',optimizer='adam')

# This `fit` call will be distributed on 8 GPUs.
# Since the batch size is 256, each GPU will process 32 samples.
parallel_model.fit(x, y, epochs=20, batch_size=256)

About ResNeXt & DenseNet

Since I don't have enough machines to train the larger networks, I only trained the smallest network described in the paper. You can see the results in liuzhuang13/DenseNet and prlz77/ResNeXt.pytorch

   

Please feel free to contact me if you have any questions!

Citation

@misc{bigballon2017cifar10cnn,
  author = {Wei Li},
  title = {cifar-10-cnn: Play deep learning with CIFAR datasets},
  howpublished = {\url{https://github.com/BIGBALLON/cifar-10-cnn}},
  year = {2017}
}