This file is to help with writing a customized config file
All default configuration files can be seen in config/verbose/default.yaml
Note: When training models, the complete config file (i.e., containing all parameters for a specified model) will be generated and stored in the results folder
The config file can be divide into 3 main parts: data, model and training_parameters
data: describe all the data files like embeddings, features and annotations needed for training and evaluation
training_parameters: describe parametes for the training process like such as max iterations, warm-up ratio etc
model: describe the configuration of models, the model structure is illustrated in the following figure
All available configs can be found in config/config.py and config/function_config_lib.py In the config file, there are many method/par pairs. This is because different methods can have different parameters, for example, for modal_combine (for multi modal combination), there are MFH and non_linear_elmt_multiply. If only the method is changed, the coresponding default parameters will be loaded automatically
To change a config file, the user only needs to replace the correlated part, for example, to run model with image feature fine-tune, the config/keep/detectron23_finetune.yaml can be used, while the verbose config file generated is config/verbose/dectectron_finetune.yaml
To avoid creating a separate config file for each minor change, pythia also allows to overwrite config files from command-line arguments by using --config_overwrite. The overwrite a string follow a semi-json format, i.e. no quote for keys. The following are several examples
change the steps to adjust learning rate
python train.py \
--config_overwrite '{training_parameters:{lr_steps:[17000,20000,23000],max_iter:25000}}' change the dataset to use all val2014 in training
python train.py \
--config_overwrite '{data:{imdb_file_train:[\"imdb/imdb_train2014.npy\",\"imdb/imdb_val2014.npy\"]}}'