finetune-in-PyTorch

Finetune networks in pytorch

Requirements

1. imgaug
2. tqdm

Usage

1. Use gen_dataset.py to generate datasets in csv files from data folders.

   The data were located in seperate folders according to classes.

   The generated datasets are train.csv and valid.csv.

2. Use finetune.sh to train networks according to generated datasets.

   ./finetune.sh 1
   

   to train networks using GPU 1.

3. Use inference.py to test some data with the trained model.

4. Use check_train.py to check the trainning dataset and fetch out the unconsistent data.

Data and dataset

1. Use gen_dataset.py to generate datasets in csv files from data folders.

2. Use gen_and_save_images.py to test the dataset and dataloader from csv files.

3. Use gen_new_data.py to generate a csv from a folder to test the files.

4. Use fetch_data_accordingto_error.py to fetch error data from the trainning or validation data.

Models

1. ResNet

   ResNet 50, 101 and 152 from official repositories.

2. Se_ResNeXt

   senet.py is the model definition, train_se_resnext.py and train_se_resnext.sh are used to train models.

3. PNasNet

   pnasnet.py is the model definition, train_pnasnet.py and train_pnasnet.sh are used to train models.

Learning rate decay methods

1. Decay according to steps

   finetune.py and finetune.sh.

2. Reduce on Plateau

   finetune_rop.py and finetune_rop.sh.

References

Some codes were borrowed and modified from pretrained-models.pytorch and pytorch-cifar.