Write a `.yaml` configuration file

Code for this section:

core/config/config.py
config/*

Composition of the configuration file in LibFewShot

The configuration file of LibFewShot uses a yaml format file and it also supports reading the global configuration changes from the command line. We have pre-defined a default configuration core/config/default.yaml. The users can put the custom configuration into the config/ directory, and save this file in the yaml format. At parsing, the sequencing relationship of defining the configuration of the method is default.yaml->config/->console. The latter definition overrides the same value in the former definition.

Although most of the basic configurations have been set in the default.yaml, you can not directly run a program just using the default.yaml. Before running the code, the users are required to define a configuration file of one method that has been implemented in LibFewShot in the config/ directory.

Considering that FSL menthods usually have some basic parameters, such as way, shot or device id, which are often needed to be changed, LibFewShot also supports making changes to some simple configurations on the command line without modifying the yaml file. Similarly, during training and test, because many parameters are the same of different methods, we wrap these same parameters together and put them into theconfig/headers for brevity. In this way, we can write the yaml files of the custom methods succinctly by importing them.

The following is the composition of the files in the config/headers directory.

data.yaml: The relevant configuration of the data is defined in this file.
device.yaml: The relevant configuration of GPU is defined in this file.
losses.yaml: The relevant configuration of the loss used for training is defined in this file.
misc.yaml: The miscellaneous configuration is defined in this file.
model.yaml: The relevant configuration of the model is defined in this file.
optimizer.yaml: The relevant configuration of the optimizer used for training is defined in this file.

The settings of the configuration file in LibFewShot

The following details each part of the configuration file and explain how to write them. An example of how the DN4 method is configured is also presented.

The settings for data

data_root: The storage path of the dataset.
image_size: The size of the input image.
use_momery: Whether to use memory to accelerate reading.
augment: Whether to use data augmentation.
augment_times：support_set: The number of data augmentation/transformations used. Expanding the support set data for multiple times.

augment_times_query：query_set: The number of data augmentation/transformations used. Expanding the query set data for multiple times.

data_root: /data/miniImageNet--ravi
image_size: 84
use_memory: False
augment: True
augment_times: 1
augment_times_query: 1

The settings for model

backbone: The backbone information used in the method.
- name: The name of the backbone, needs to match the case of the backbone implemented in LibFewShot.
- kwargs: The parameters used in the backbone, must keep the name consistent with the name in the code.
  - is_flatten: The default is False, and if True, the feature vector after flatten is returned.
  - avg_pool: The default is False, and if True, the feature vector after global average pooling is returned.
  - is_feature: The default is False, and if True, the output of each block in backbone is returned.
```
backbone:
    name: Conv64FLeakyReLU
    kwargs:
        is_flatten: False
```
classifier: The classifier information used in the method.
- name: The name of the classifier, needs to match the case of the classifier implemented in LibFewShot.
- kwargs: The parameters used in the classifier initialization, must keep the name consistent with the name in the code.
```
classifier:
    name: DN4
    kwargs:
        n_k: 3
```

The settings for training

epoch: The number of epoch during training.
test_epoch: The number of epoch during testing.
pretrain_path: The path of the pre-training weights. At the beginning of the training, this setting will be first checked. If it is not empty, the pre-trained weights of the target path will be loaded into the backbone of the current training.
resume: If set to True, the training status is read from the default address to support continual training.
way_num: The number of way during training.
shot_num: The number of shot during training.
query_num: The number of query during training.
test_way: The number of way during testing. If not specified, the way_num is assigned to the test_way.
test_shot: The number of shot during testing. If not specified, the shot_num is assigned to the test_way.
test_query: The number of query during testing. If not specified, the query_num is assigned to the test_way。
episode_size: The number of tasks/episodes used for the network training at each time.
batch_size: The batch size used when the pre-training model is pre-trained. In some kinds of methods, this property is useless.
train_episode: The number of tasks per epoch during training.

test_episode: The number of tasks per epoch during testing.

epoch: 50
test_epoch: 5

pretrain_path: ~
resume: False

way_num: 5
shot_num: 5
query_num: 15
test_way: ~
test_shot: ~
test_query: ~
episode_size: 1
# batch_size only works in pre-train
batch_size: 128
train_episode: 10000
test_episode: 1000

The settings for optimizer

optimizer: Optimizer information used during training.
name: The name of the Optimizer, only temporarily supports all Optimizers provided by PyTorch.
kwargs: The parameters used in the optimizer, and the name needs to be the same as the parameter name required by the pytorch optimizer.
other: Currently, the framework only supports the learning rate used by each part of a separately specified method, and the name needs to be the same as the variable name used in the method.

optimizer:
    name: Adam
    kwargs:
        lr: 0.01
    other:
        emb_func: 0.01
        #For demonstration purposes, there are no additional training parameters for dn4.
        dnf_layer: 0.001

lr_scheduler: The learning rate adjustment strategy used during training, only temporarily supports all the learning rate adjustment strategies provided by PyTorch.

name: The name of the learning rate adjustment strategy.
kwargs: Other parameters used in the learning rate adjustment strategy in PyTorch.

lr_scheduler:
  name: StepLR
  kwargs:
    gamma: 0.5
    step_size: 10

The settings for Hardware

device_ids: The gpu number, which is the same as the nvidia-smi command.

n_gpu: The number of parallel gpu used during training, if 1, it can’t apply to parallel training.

deterministic: Whether to turn on torch.backend.cudnn.benchmark and torch.backend.cudnn.deterministic and whether to determine random seeds during training.

seed: Seed points used in numpy，torch，and cuda.

device_ids: 0,1,2,3,4,5,6,7
n_gpu: 4
seed: 0
deterministic: False

The settings for Miscellaneous

log_name: If empty, use the auto-generated classifier.name-data_root-backbone-way_num-shot_num file directory.

log_level: The log output level during training.

log_interval: The number of tasks for the log output interval.

result_root: The root of the result.

save_interval: The epoch interval to save weights.

save_part: The name of the variable in the method that needs to be saved. Variables with these names are saved separately when the model is saved. The parts that need to be saved are given as a list under save_part.

log_name: ~
log_level: info
log_interval: 100
result_root: ./results
save_interval: 10
save_part:
    - emb_func
    - dn4_layer

Write a .yaml configuration file