myTiny-imageNet-classification

This repository is a minimum implementation to train a image classification network on tiny ImageNet dataset. It is recommended for beginners in learning deep learning computer vision. It has detailed instruction to train on your custom dataset.

Requirements

• Ubuntu 16.04 LTS
• Python 3.5
• PyTorch 0.4.1+
• Torchvision
• TensorboardX
• OpenCV
• Numpy

Installation

In your command terminal, type these command below

• Installing git (skip this step if git has been installed)
  

sudo apt-get install git

• Clone this repository and go to the repository directory
  

git clone https://github.com/albertchristianto/myTiny-imageNet-classification
cd myTiny-imageNet-classification

• Install these python libraries
  

sudo pip3 install opencv-python numpy tensorboardx
sudo pip3 install pytorch torchvision

• You are ready to run the code

Train a Image Classification Network

To train a image classification network using this repository, you must:

• Download tiny-ImageNet dataset here and extract it
• Generate train.txt, validation.txt, and test.txt inside datalist folder by running the following commands

python3 utils/createDataList.py --tinyImagenetPath [TINY_IMAGENET_FOLDER_PATH]

• And it is ready to train the network

python3 train.py

• You can also resume your training using this following command. The checkpoint file is in the checkpoint folder.

python3 train.py --resume --checkpoint_dir [YOUR_CHECKPOINT_PATH]

• To monitor the training loss, you can run this following command on another command window and open your browser to access the provided address

tensorboard --logdir [YOUR_CHECKPOINT_PATH]

Test a Image Classification Network

To test a image classification, you can run this following command

python3 test.py --val_txtPath [PATH_TO_VALIDATION_TXT_FILE] --weight_path [PATH_TO_YOUR_PTH_FILE]

Inference Code

To see the classification result of a image, run the following command

python3 demo.py ----img_Path [PATH_TO_A_IMAGE_FILE] --weight_path [PATH_TO_YOUR_PTH_FILE]

Training on Custom Dataset

To train a image classification on your custom dataset, you can follow these steps:

• Set your dataset format look like this

${DATA_ROOT}
|-- YOUR_DATASET
    |-- train
        |-- your_class_0
            |-- 000000000009.jpg
            |-- ... 
        |-- your_class_1
            |-- 000000000139.jpg
            |-- ...
        ...
    |-- validation
        |-- your_class_0
            |-- 000000000009.jpg
            |-- ... 
        |-- your_class_1
            |-- 000000000139.jpg
            |-- ...
        ...

• Generate your train.txt, val.txt, and test.txt for your dataset using this format. For reference, you can see here.

{DATA_ROOT}/YOUR_DATASET/train/your_class_0/000000000009.jpg 0
{DATA_ROOT}/YOUR_DATASET/train/your_class_1/000000000139.jpg 1
{IMG_PATH} {LABEL}
...

• Run this following command

python3 train.py --train_txtPath [PATH_TO_TRAIN_TXT_FILE] --val_txtPath [PATH_TO_VALIDATION_TXT_FILE]

• For advanced training, run this command to see the training option

python3 train.py -h