Config Parser Design

[jmsalvador2395]

main ideas

• [x] have configs in a directory structure.
• [x] build configs using built-in configparser package
• [x] design interface to prompt users for configs if none are provided

parameters to account for

I've separated these parameters into categories for convenience. this structure doesn't necessarily have to be used for building configs

general parameters

• [ ] task_name: the task to train on. this is used to identify what kind of model to train (ex. word2vec_cbow)
• [ ] device: the device to use for processing tensors (ex. cuda:0)
• [ ] seed: used to manually seed the random number generators for numpy and pytorch

optimization parameters

• [ ] lr: the optimizer learning rate
• [ ] batch_size: the amount of examples to train on at a given time
• [ ] weight_decay: the regularization term to be used for the optimizer

dataset specific parameters

• [ ] dataset_name: used to retrieve a dataset from huggingface
• [ ] data_dir: used to specify where the data is located (this is only used for custom datasets)
• [ ] shuffle: tells the program whether to shuffle the data or not
• [ ] num_epochs: the number of times to iterate over the training data
• [ ] cache_dir: where to store your dataset cache. (default would be in the project root)

list of parameters used for word2vec (maybe just put them in a category called 'Model Parameters')

• [ ] embedding_dim: the dimensionality of the embeddings to train (ex. 128)
• [ ] window_size: defines the amount of context words to use when training word2vec (ex. 9)

tools (this is for things like tokenizers and vocabularies)

• [ ] tokenizer: the desired tokenizer to use on the dataset
• [ ] min_freq: the minimum frequency a token needs to occur in the dataset before being indexed as ''

[jmsalvador2395]

link to the config parser api

[jmsalvador2395]

config search menu implemented

[jmsalvador2395]

note: implement use a class called Trainer to implement training tasks.

[jmsalvador2395]

Config parser design is done