🔎A Simple Deep Learning Trainer

🚀 Usage

• This is a PyTorch-based deep learning trainer designed for beginners to get started quickly.

1. Installation

• Clone the repo:

  git clone https://github.com/xigua7105/DLTrainer.git
  cd DLTrainer
  

• Environment setup: DLTrainer works with Python 3.8+ and PyTorch 2.0+.

  conda create -n DLTrainer python=3.8
  conda activate DLTrainer
  pip install -r requirements.txt
  

2. QuickStart

• For example, training ResNet-50 on the CIFAR-100 dataset:

   torchrun --nproc_per_node=1 --nnodes=1 --standalone train.py --c configs/resnet50-cifar100.yaml
  

📝 Configs Structure

• The configs mainly consist of six major parts: model, data, optim, loss, trainer, and logger. The following is the explanation of the usage of different parameters for each part.

1. Model

• name (optional): Your custom naming for the network, such as convnet-lite, convnet-base, convnet-large, etc.
• task (optional): A brief introduction to the task.
• struct (required): Parameters required for defining the model architecture.

2. Data

• dir (required): The path of the dataset.
• dataset_type (required): The type of the dataset, such as IRDataset (image restoration dataset), ImageFolderDefault (image classification dataset), etc. You can define the dataset in dataset and register it in _register.
• is_multi_loader (required): Whether there are multiple training sets or test sets.
• train_transforms:
  • name (required): The names of the defined transforms function. You can define and register it in transforms.
  • kwargs (optional):
• test_transforms (required):
• train_target_transforms (required):
• test_target_transforms (required):

3. Optim

• optimizer:
  • name (required): Such as Adam, SGD, etc. You can register them in _register.
  • lr (required): Learning rate.
• scheduler:
  • name (required): Such as Cosine, MultiStepLR, etc.

4. Loss

• loss_terms:
  • name (required): Such as CrossEntropyLoss, IRLoss, etc. You can define and register it in loss.

5. Trainer

• name (required): The type of the Trainer, such as CLSTrainer, IRTrainer, etc. You can define it in trainer and register it in _register.
• ckpt_dir (required): The folder for saving the checkpoint.
• batch_size (required): Global train batch size.
• batch_size_test (required): Global test batch size.
• num_workers_per_gpu (required): Dataloaders parameters.
• drop_last (required): Dataloaders parameters.
• pin_memory (required): Dataloaders parameters.
• save_freq (required): The frequency of saving the checkpoint.
• amp (required): Whether to enable mixed precision to accelerate the training of the model.

6. Logger

• dir: The folder for saving the log.
• log_freq: The frequency of saving the log.

License

This project is released under the MIT license. Please see the LICENSE file for more information.