Contrastive Code Representation Learning

    npm root -g # 输出默认路径

    npm config set prefix "D:\repository\node\node_global"   #存储路径
    npm config set cache "D:\repository\node\node_cache"  #缓存路径

    npm install # 自动安装依赖

    pip install -e "." #  安装python依赖

    def pretrain(
    run_name: str,
    # Data
    train_filepath: str = DEFAULT_CSNJS_TRAIN_FILEPATH,
    spm_filepath: str = DEFAULT_SPM_UNIGRAM_FILEPATH,
    num_workers=1,
    limit_dataset_size=-1,
    max_length=1024,
    subword_regularization_alpha: float = 0,
    program_mode="contrastive",
    loss_mode="infonce",  # infonce, mlm, or hybrid
    min_alternatives=1,
    #
    # Model
    resume_path: str = "",
    encoder_type: str = "transformer",
    lstm_project_mode: str = "hidden",
    n_encoder_layers: int = 6,
    d_model: int = 512,    #rnn encode embedding size
    #
    # Optimization
    num_epochs: int = 100,    #迭代次数
    save_every: int = 1,     
    batch_size: int = 256,       # batch_size
    lr: float = 8e-4,                  #学习率
    weight_decay: float = 0,          #L2正则化参数
    adam_betas=(0.9, 0.98),      
    warmup_steps: int = 5000,
    num_steps: int = 600000,
    #
    # Distributed
    rank: int = -1,
    dist_url: str = "env://",
    dist_backend: str = "nccl",
    #
    # Computational
    use_cuda: bool = True,   #是否使用cuda
    seed: int = 0,       #种子
    ):