AntBee

Toy example of image classification

In this example, we finetune a pretrained resnet18 for classification of images with two categries: Ant and Bee. This example is a PyMIC implementation of pytorch's "transfer learning for computer vision tutorial". The orginal tutorial can be found here. In PyMIC's implementation, we only need to edit the configure file to run the code.

Data and preprocessing

1. The dataset contains about 120 training images each for ants and bees. There are 75 validation images for each class. Download the data from here and extract it to PyMIC_data. Then the path for training and validation set should be PyMIC_data/hymenoptera_data/train and PyMIC_data/hymenoptera_data/val, respectively.
2. Run python write_csv_files.py to create two csv files storing the paths and labels of training and validation images. They are train_data.csv and valid_data.csv and saved in ./config.

Finetuning all layers of resnet18

1. Here we use resnet18 for finetuning, and update all the layers. Open the configure file config/train_test_ce1.cfg. In the network section we can find details for the network. Here update_mode = all means updating all the layers.

# type of network
net_type   = resnet18
pretrain   = True
input_chns = 3
# finetune all the layers
update_mode = all 

Then start to train by running:

pymic_train config/train_test_ce1.cfg

2. During training or after training, run tensorboard --logdir model/resnet18_ce1 and you will see a link in the output, such as http://your-computer:6006. Open the link in the browser and you can observe the average loss and accuracy during the training stage, such as shown in the following images, where blue and red curves are for training set and validation set respectively. The iteration number obtained the highest accuracy on the validation set was 400, and may be different based on the hardware environment. After training, you can find the trained models in ./model/resnet18_ce1.

Testing and evaluation

1. Run the following command to obtain classification results of testing images. By default we use the best performing checkpoint based on the validation set. You can set ckpt_mode to 0 in config/train_test_ce1.cfg to use the latest checkpoint.

mkdir result
pymic_test config/train_test_ce1.cfg

2. Then run the following command to obtain quantitative evaluation results in terms of accuracy.

pymic_eval_cls -cfg config/evaluation.cfg

The obtained accuracy by default setting should be around 0.9477, and the AUC will be around 0.9745.

3. Run python show_roc.py to show the receiver operating characteristic curve.

Finetuning the last layer of resnet18

Similarly to the above example, we further try to only finetune the last layer of resnet18 for the same classification task. Use a different configure file config/train_test_ce2.cfg for training and testing, where update_mode = last in the network section means updating the last layer only:

net_type   = resnet18
pretrain   = True
input_chns = 3
# finetune the last layer only
update_mode = last

Edit config/evaluation.cfg accordinly for evaluation. The corresponding accuracy and AUC would be around 0.9477 and 0.9778, respectively.