This repo is designed for those who want to start their experiments two days before the deadline and kill the project in the last 6 hours. new_moon_with_face Inspired by fb.torch.resnet, it provides fast experiment setup and attempts to maximize the number of projects killed within the given time. Please feel free to submit issues or pull requests if you want to contribute.
Both Python 2.7 and 3 are supported; however, it was mainly tested on Python 3.
python main.py -h to show all arguments.
Train a ResNet-56 on CIFAR-10 with data augmentation using GPU0:
CUDA_VISIBLE_DEVICES=0 python main.py --data cifar10 --data_aug --arch resnet --depth 56 --save save/cifar10 -resnet-56 --epochs 164
Train a ResNet-110 on CIFAR-100 without data augmentation using GPU0 and GPU2:
CUDA_VISIBLE_DEVICES=0,2 python main.py --data cifar100 --arch resnet --depth 110 --save save/cifar100-resnet-110 --epochs 164
See scripts/cifar10.sh and scripts/cifar100.sh for more training examples.
python main.py --resume save/resnet-56/model_best.pth.tar --evaluate test --data cifar10
Adding your custom model
You can write your own model in a .py file and put it into models folder. All you need it to provide a
createModel(arg1, arg2, **kwarg) function that returns the model which is an instance of nn.Module. Then you'll be able to use your model by setting
--arch your_model_name (assuming that your model is in a the file models/your_model_name).
Show Training & Validation Results
getbest.py save/* FOLDER_1 FOLDER_2
In short, this script reads the scores.tsv in the saving folders and display the best validation errors of them.
tensorboard --logdir save --port PORT
Experiment Setup & Logging
- Ask before overwriting existing experiments, and move the old one to /tmp instead of overwriting
- Saving training/validation loss, errors, and learning rate of each epoch to a TSV file
- Automatically copying all source code to saving directory to prevent accidental deleteion of codes. This is inspired by SGAN code.
- TensorBoard support using tensorboard_logger
- One script to show all experiment results
- Display training time
- Holding out testing set and using validation set for hyperparameter tuning experiments
- GPU support
- Adding save & data folders to .gitignore to prevent commiting the datasets and trained models
- Result table
- Python 2.7 & 3.5 support
Models (See models folder for details)
Last 5000 samples in the original training set is used for validation. Each pixel is in [0, 1]. Based on experiments results, normalizing the data to zero mean and unit standard deviation seems to be redundant.
Test Error Rate (in percentage) with validation set
The number of parameters are calculated based on CIFAR-10 model.
ResNets were training with 164 epochs (the same as the default setting in fb.resnet.torch) and DenseNets were trained 300 epochs.
Both are using batch_size=64.
|Model||Parameters||CIFAR-10||CIFAR-10 (aug)||CIFAR-100||CIFAR-100 (aug)|
|ResNet-110 with Stochastic Depth||1.73M||5.25||24.2|
Top1 Testing Error Rate (in percentage)
- main.py: main script to train or evaluate models
- train.py: training and evaluation part of the code
- config: storing configuration of datasets (and maybe other things in the future)
- utils.pypy: useful functions
- getbest.py: display the best validation error of each saving folder
- dataloader.py: defines getDataloaders function which is used to load datasets
- models: a folder storing all network models. Each script in it should contain a createModel(**kwargs) function that takes the arguments and return a model (subclass of nn.Module) for training
- scripts: a folder storing example training commands in UNIX shell scripts
This code is based on the ImageNet training script provided in PyTorch examples.
The author is not familiar with licensing. Please contact me there is there are any problems with it.