ImageNet

This is a ImageNet Example for classification task.

Installation

I suggest you use conda to create a environment and use pip to install packges.(Because pip is faster than conda especially use this mirror https://pypi.tuna.tsinghua.edu.cn/simple/)

git clone https://github.com/LOOKCC/ImageNet.git
pip install torch torchvision -i https://pypi.tuna.tsinghua.edu.cn/simple/
pip install tensorboard
pip install PyYAML

Config

config is at config.yaml, They are commonly used configurations, such as lr or batch_szie. There are some important config need to be explanded.

• fix_fc Because imagenet has 1000 classes, when your task has 2 classes, you need to fix last fc layer. But if the fc is 2, you can't load imagenet pretrain models. So you need to fix fc layer after load imagenet checkpoint.
• max_epoch and max_iteration Your training will stop when epoch reach max_epoch or iteration reach max_iteration. So you should make both of them a lager number to avoid stopping early.
• lr just follow this formula, and lr_lowest represents the lowest lr.

lr = args.base_lr * (args.lr_decay_factor ** (epoch // args.lr_decay_epoch))

• validation_freq controls how many iterations to test.

Dataset

In config, there are two text file represented train_set and test_set. In text file, One line represents one sample. They format as follows:

path_to_image/example.txt label

The tpye of label is int, for example, if it's a Binary classification problem, then label is 0 or 1.

Usage

usage: main.py [-h] [-c CONFIG] [--resume PATH] [-e] [--save SAVE]

PyTorch ImageNet Training

optional arguments:
  -h, --help            show this help message and exit
  -c CONFIG, --config CONFIG
                        path to config file.
  --resume PATH         path to latest checkpoint (default: none)
  -e, --evaluate        evaluate model on validation set
  --save SAVE           save_dir

tips: you may need to use CUDA_VISIBLE_DEVICES. Because the default device is cuda:0
sample example:

CUDA_VISIBLE_DEVICE=2,3 python main.py -c config.yaml --save save_test

Tensorboard

If you want to use Tensorboard:

tesorboard --logdir=. --port=8090 --blind_all

Add Models

If you want to add your Net, such as Resnet. You need to do followings:

1. Add your_net.py in ./net
2. In your_net.py, us __all__ to set interface. if you don't konw about this, just do this link do: https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py
3. modify ./net/__init__.py, add from xxx import xxx
4. modify confi: arch to the net name

Result

If your --save is save_test, there will be a file called report.txt be created.

iter train_loss train_acc test_loss test_acc lr
1000 0.5406224970574622 87.31268310546875 0.4301627216339111 86.00000762939453 0.1
2000 0.48305593270888514 87.4562759399414 0.41288787841796876 86.00000762939453 0.1

This informaion will help you to choose checkpoints.

Finally

I will very appreciate if this repository helps your training. GL&HF.