config_Clinc_LSTM.json

{
    "name": "LTA",
    "n_gpu": 1,

    "step1": false,
    "step2": true,
    "encoder_type": "lstm",
    "arch_step1": {
        "type": "Encoder",
        "args": {
            "encoder_param": {
                "hidden_size": 64,
                "num_layers": 1,
                "bi": true,
                "freeze_emb": false,
                "tau": 10.0
            }
        },
        "lr": 1e-3
    },
    "arch_step2": {
        "type": "LTA",
        "args": {
            "encoder_param": {
                "hidden_size": 64,
                "num_layers": 1,
                "dropout": 0.5,
                "bi": true,
                "freeze_emb": false
            },
            "tau": 10.0,
            "alpha": 10.0,
            "d_r": 64
        },
        "lr": {
            "1e-2": ["tau"],
            "1e-3": ["proto_adapt_net", "extractor", "seen_class_protos"],
            "1e-4": ["lstm", "others"]
        },

        "ablation": {
            "init": true,
            "proto_adapt": true,
            "feat_adapt": true
        }
    },

    "dataset": {
        "name": "Clinc",
        "load_path": "data/ver1/Clinc/Clinc_Goog.pkl"
    },
    "data_sampler": {
        "type": "MNFBatchSampler",
        "args":{
            "iterations": 20,
            "seen_N": 10,
            "seen_K": 5,
            "unseen_N": 10,
            "unseen_K": 5
        }
    },
    "optimizer": {
        "type": "Adam",
        "args":{
            "weight_decay": 0,
            "amsgrad": true
        }
    },
    "lr_scheduler": {
        "type": "StepLR",
        "args": {
            "step_size": 500,
            "gamma": 0.1
        }
    },
    "loss": "CosT",
    "save_dir": "saved/",
    "trainer_step1": {
        "type": "StepOneTrainer",
        "args": {
            "epochs": 10,
            "len_epoch": 20,
            "save_per_epochs": 500,

            "verbosity": 2,

            "train_batch_size": 64,
            "valid_batch_size": 256,
            "monitor": "max val_GZSL_acc_hm",
            "early_stop": 3,

            "tensorboard": false,
            "histogram": false
        }
    },
    "trainer_step2": {
        "type": "StepTwoTrainer",
        "args": {
            "epochs": 300,

            "save_per_epochs": 500,

            "verbosity": 2,

            "valid_batch_size": 256,
            "monitor": "max val_GZSL_acc_hm",
            "early_stop": 30,

            "tensorboard": true,
            "histogram": false
        }
    },
    "test": {
        "batch_size": 1024
    }
}