tensorflow-cifar100

Tensorflow implementation on cifar100.

All models have achieved high accuracy (> 0.7).

Usage

Requirements:

1. tensorflow-gpu=1.11.1
2. tensorlayer=1.11.0

download dataset:

Download Website

download repo:

$ git clone https://github.com/Ecohnoch/tensorflow-cifar100

train:

$ python3 -u train.py train --batch_size 64 --epoch 200 --network resnet50 --opt momentum --train_path /data/ChuyuanXiong/up/cifar-100-python/train --test_path /data/ChuyuanXiong/up/cifar-100-python/test

params:

• batch_size: 64 default
• epoch: 200 is best
• network: resnet18/resnet50/resnet110/resnet152/seresnet50/seresnet110/seresnet152/densenet121/densenet169/densenet161/densenet201/resnext50/resnext110/resnext152/seresnext50/seresnext110/seresnext152/densenet100bc/densenet190bc
• opt: adam/momentum/nesterov
• train_path: your train path
• test_path: your test path

Have Done

ResNet18
ResNet34
ResNet50
ResNet110
ResNet152
ResNeXt50
ResNeXt110
ResNeXt152
SENet50
SENet110
SENet152
SE-ResNext50
SE-ResNext110
SE-ResNext152
DenseNet121
DenseNet169
DenseNet201
DenseNet100BC
DenseNet190BC

# TODO
preresnet
mobilenet

test:

python3 -u train.py test --network resnet18 --test_path '/data/ChuyuanXiong/up/cifar-100-python/test' --ckpt 'params/resnet18/Speaker_vox_iter_58000.ckpt'

params:

• network: resnet18/resnet50
• test_path: your test path
• ckpt: your pre-trained model. You can try the [$THIS_REPO/params/resnet18/Speaker_vox_iter_58000.ckpt]

Also, If you have the pre-trained model, you can try the interface to quickly access to the test results. Just try this:

$ cd [to the root directory of this repo]
$ python
>>> from pretrained.cifar100 import cifar100
>>> model = cifar100(model='resnet18')
>>> model.test()

Or you can cd to the dir and then edit and run example.py.

Results

dataset	network	top1 acc	epoch (lr=0.1)	epoch (lr=0.02)	batch_size	initializer	warmup	weight decay
cifar100	resnet18	0.740	60	> 60	128	msra	0	0
cifar100	densenet169	0.743	60	> 60	64	orth	1	5e-4
cifar100	densenet100bc	0.728	60	> 60	32	orth	1	5e-4
cifar100	densenet201	0.753	60	> 60	64	orth	1	5e-4
cifar100	seresnet110	0.725	60	> 60	64	orth	1	5e-4

// TODO

• preresnet18
• ...

Pre-trained model download

Continuous update!

1. ResNet18,Accuracy=0.740
2. DenseNet169,Accuracy=0.743,Password=7qj2
3. DenseNet100-BC,Accuracy=0.728,Password=fwi4
4. Se-ResNet110,Accuracy=0.725,Password=we64

References

1. pytorch-cifar100

Author

Ecohnoch (Chuyuan Xiong)

If this project is very helpful for you, please star it!