Getting started

This section shows an example of a process of using LibFewShot.

Prepare the dataset (use miniImageNet as an example)

1. download and extract miniimagent–ravi.

2. check the structure of the dataset：

   The dataset must be in the following structure:

   dataset_folder/
   ├── images/
   │   ├── images_1.jpg
   │   ├── ...
   │   └── images_n.jpg
   ├── train.csv *
   ├── test.csv *
   └── val.csv *
   

Modify the config file

Use ProtoNet as an example：

1. create a new yaml file getting_started.yaml in config/

2. write the following commands into the created file:

   includes:
     - headers/data.yaml
     - headers/device.yaml
     - headers/losses.yaml
     - headers/misc.yaml
     - headers/model.yaml
     - headers/optimizer.yaml
     - classifiers/Proto.yaml
     - backbones/Conv64F.yaml
   

More details can be referred to write a config yaml.

Run

1. set the config as follows in run_trainer.py:

   config = Config("./config/getting_started.yaml").get_config_dict()
   

2. train with the console command:

   python run_trainer.py
   

3. wait for the end of training.

View the log files

After running the program, you can find a symlink of results/ProtoNet-miniImageNet-Conv64F-5-1 and a directory of results/ProtoNet-miniImageNet-Conv64F-5-1-$TS, where TS means the timestamp. The directory contains two folders: checkpoint/ and log_files/, and a configuration file: config.yaml. Note that the symlink will always link to the directory created at the last time, when you train the model with the same few-shot learning configuration for multiple times.

config.yaml contains all the settings used in the training phase.

log_files/ contains tensorboard files, training log files and test log files.

checkpoints/ contains model checkpoints saved at $save_insterval intervals, the last model checkpoint (used to resume) and the best model checkpoint (used to test). The checkpoint files are generally divided into emb_func.pth, classifier.pth, and model.pth (a combination of the first two), respectively.