Cifar10 Tensorflow Project
Get Started
- environment:
tensorflow-gpu1.8+cude9.0 - datasets from kaggle : CIFAR-10 - Object Recognition in Images, first you can download the train and test dataset.
- then use the
utils/get_data_list.pyandutils/get_dataset_mean.pyscripts to generatetrain.txtandval.txt.
How to Learn this project
- one step: you can modify trian parameters in
config/cifar10_config.json. - two step: you can learn how load datasets before training from
src/datasets/cifar10_dataloader.py. - three step: you can learn how to write network from
src/models/layersandsrc/models/simple_model.py, you can easily create you own model. - four step: you should finish trian scripts
tools/train_cifar10.py, in this process you will finish loss function and metric funtion：src/loss/cross_entropy.pyandsrc/metrics/acc_metric.py; in this scriptstools/train_cifar10.py, we will first create graph and then run session. at the same time, we will record train models and use tensorboard to visual loss and accuracy inexperiments/chekpointandexperiment/summaryfolder. - five step: you can run train scripts:
tools/train_cifar10.py. - six step: when you get train model, you can predict image and get class name in
demo/prdict.py.
- seven step: you can also get some extra information from
demo/visual.py, such as weights or visual feature map.
- other: you can fimilar how to use some tool function in
tools/utils.py.
The optimization process
- The detailed information you can get from there.
- run scripts
tools/trian_cifar10.pyinclude adjust lr , add data augmentation ,add dropout ,weight decay training tricks stack 3*3 conv. you can learn how train model acc from 70%+ to 91+%.
- run scripts
tools/trian_cifar10_v2.pyinclude add batch_norm ...
