loinh1106 / train_Swin_DOLG

training swin and DOLG

HCM AI CHALLENGE 2022 - Event Retrieval from Visual Data

---

To do task

• DOLG
• Hybrid swin transformer
• Mish activation
• Centralize Gradient

---

Requirements

pip install -r requirements.txt

---

Setup

If apex folder is not exist

!git clone https://github.com/NVIDIA/apex
%cd apex
!python setup.py install

Prepare CSV file for train, validation (or test)

python prepare_data_csv.py \
--data_path /content/small_dog_cat_dataset \
--output_path /content/data \
--split_test

Note

--data_path: Đường dẫn đến data images

--output_path: Đường dẫn chứa các file CSV

--split_test: Nếu cần tách thành 3 file train.csv, val.csv và test.csv (Tỷ lệ Train-Val-Test là 7-2-1)

Train

python -u -m torch.distributed.launch --nproc_per_node=1 \
          train.py \
          --model_name dolg \
          --config_name dolg_b5_step3 \
          --trainCSVPath /content/drive/MyDrive/AIC_HCM/Image_Retrieval_from_Visual_Data/data/train.csv \
          --valCSVPath /content/drive/MyDrive/AIC_HCM/Image_Retrieval_from_Visual_Data/data/val.csv \
          --loss_name cosface \
          --use_central_gradient \
          --use_mish \ 

Note

--model_name: 'dolg' cho DOLG và 'swin' cho Swin Transformer

--config_name: tên file config tương ứng với model

--trainCSVPath: Đường dẫn đến file train csv

--valCSVPath: Đường dẫn đến file val csv

--loss_name: Tên của Loss.

• arcface_dynamicmargin: ArcFaceLossAdaptiveMargin Loss
• cosface : Cosface Loss
• CE_smooth_loss : CrossEntropyLossWithLabelSmoothing Loss
• smooth_CE_loss: LabelSmoothingCrossEntropy Loss
• circleloss: CircleLoss

--use_central_gradient: Sử dụng central_gradient, nếu không dùng chỉ cần comment nó lại

--use_mish: Sử dụng Mish Function thay cho SiLU Activation nếu không dùng chỉ cần comment nó lại

Load Trained Model and Continue Training

python -u -m torch.distributed.launch --nproc_per_node=1 \
          train.py \
          --model_name dolg \
          --config_name dolg_b5_step3 \
          --trainCSVPath /content/drive/MyDrive/AIC_HCM/Image_Retrieval_from_Visual_Data/data/train.csv \
          --valCSVPath /content/drive/MyDrive/AIC_HCM/Image_Retrieval_from_Visual_Data/data/val.csv \
          --checkpoint './run/saved/dolg_efficientnet_b5_ns_step3_2.pth' \
          --loss_name cosface \
          --use_central_gradient \
          --use_mish \ 

---

Folder Structure

  Main-folder/
  │
  ├── config/ 
  │   ├── config.py - configuration
  │
  ├── data/ - default directory for storing input data
  │
  ├── data_loader/ - anything about data loading goes here
  │   └── dataset.py
  |
  ├── model/ - this folder contains any net of your project.
  │   ├── model.py
  │   ├── metric.py
  │   └── loss.py
  │
  ├── saved/
  │   ├── models/ - trained models are saved here
  │   └── log/ - default logdir for tensorboard and logging 
  │   └── submission/ -  submission file are saved here
  │
  ├── scripts/ - main function 
  │   └── pipeline.py
  │   └── OCR.py
  │   └── segment.py
  |
  ├── test/ - test functions
  │   └── run.py
  │   └── ...
  |
  ├── tools/ - open source are saved here
  │   └── detectron2 dir
  │   └── ...
  │  
  └── utils/ - small utility functions
      ├── util.py
      └── ...