In this repository, some important features of a deep classification model are written by the author manually instead of calling pytorch library directly e.g. the softmax cross entropy loss in training process and a number of connected dense blocks which are use to define the base architecture of netowrks.

• It's pure Python code. That means a beginner can easily understand the workflow of it and there will not be any confused and terrible C++ code for you to get confused!!!
• Hyperparameters are flexible to change. We took the method used by rbgirshick/py-faster-rcnn. In the form of easydict, all the parameters needed were set in config file. You can set some parameters freely to debug before the program runs.
• Tensorboard Support To visualize the training process, we introduce tensorboardX to our porject for a better visualization experience.

• Multiple GPUs training. Because the memory usage of general classification tasks are considerable due to the large minibatch, we will add some codes to support multiple GPUs training mode to accelerate the trainingn process and further enlarge the size of a minibatch as much as possible.
• Customizing image preprocessing process. In our now version of code, we simply use the Compose function of Torchvision to generate the method of image preprocessing process. In the future vision, we will replace them with our lambda process function

Firstly, clone the entire project

git  clone  https://github.com/heekhero/pytorch-classification-simple

• Python 3.6
• Pytorch 1.1.0 and torchvision 0.3.0 Installation instructions
• CUDA 9.0 or higher

Then use pip to prepare other packages:

pip install -r requirements.txt

create data folders:

cd classification
mkdir data
cd data
mkdir train
mkdir test

Prepare your own dataset in form of folder + images and I highly recommended you to create soft link with respect to these prepared datasets in train and test folder you have created before.

The tree architecture of the folders should look like this:

data
    train
        class_1
            xxx.jpg
            yyy.jpg
            zzz.jpg
            ...
        class_2
            xxx.jpg
            yyy.jpg
            zzz.jpg
            ...
        ...
    test
        class_1
            xxx.jpg
            yyy.jpg
            zzz.jpg
            ...
        class_2
            xxx.jpg
            yyy.jpg
            zzz.jpg
            ...
        ...

PS: class_X is the true class of your images e.g. person, train, bike... and the file name of images is not important but need to follow the correct image suffix.

For simplification, run the follow instrction to prepare a simple datasets in terms of Apple and Orange classification.

bash ./scripts/prepare_datasets.sh

You can set all the parameters used in training and testing period as you wish in lib/utils/config.py!!! Without any boring terminal parameters!!! Anything is changeable!!!

Simply run with:

CUDA_VISIBLE_DEVICES=$GPU_ID python train.py

If you want to use tensorboard and you just need to set __C.TRAIN.USE_TFBOARD=True and run with

tensorboard --logdir='logs' --port=6006

Then you can view the loss curve convergence during your training through localhost:6006

If you want to evaluate the performance of your trained model, simply run

CUDA_BISIBLE_DEVICES=$GPU_ID python test.py

Please site this repository if you utilize this project.

@article{heekhero2019clas,
    Author = {heekhero},
    Title = {A very simple Pytorch project for image classification},
    Journal = {https://github.com/heekhero/pytorch-classification-simple},
    Year = {2019}
}