/Classification-Set-with-Tensorflow

Lightweight model

Primary LanguagePython

Classification-tensorflow

This is a image classification implementation in tensorflow.

Include：SqueezeNet,Xception,MobileNetV1、V2、V3,ShuffleNetV1、V2.
Include：LeNet,AlexNet,VGG,ResNet,Inception,ResNext,DenseNet,SENet.

TODO:

Network structure interpretation

requirement

python
cuda
cudnn
tensorflow

data structure

dateset dir

  • train
    • class1
    • class2
    • class3
    • ...
  • val
    • class1
    • class2
    • class3
    • ...

label map

label.txt

class1:0
class2:1
class3:2
...

Usage

train

python train.py   
		--train_dir="D:/train"  
		--logs_train_dir="./model_save"  
		--N_CLASSES=3   
		--size=224  
		--BATCH_SIZE=32  
		--epochs=50  
		--inin_lr=0.01  
		--decay_steps=20  
		--model="MobileNetV3_small"

inference

python inference.py   
		--data_dir="D:/RoadMapSample/"  
		--save_dir="D:/res/"  
		--logs_train_dir="./model_save/model.ckpt-24000"  
		--N_CLASSES=3  
		--cate="['class1','class2','class3'...]"  
		--model="MobileNetV3_small"

Tips

Sometimes spying images from website according to the error images is a simple but effective method to increase to accuracy.You can spy from "Baidu" using spider.py in data_preprocessing.And then rename them to avoid Chinese character.Finally,you must check if the image is undamaged with readimg.py.