MetaTL/hyper_main.py

import os
os.environ['CUDA_VISIBLE_DEVICES'] = "1"

from ray.tune.schedulers import ASHAScheduler
from ray.tune import CLIReporter
from ray import tune
from functools import partial
from hyper_tunning import train_metatl
import argparse
import numpy as np
import torch
import random
from trainer import *
from utils import *
from sampler import *
import copy

def get_params():
    args = argparse.ArgumentParser()
    args.add_argument("-data", "--dataset", default="electronics", type=str)
    args.add_argument("-seed", "--seed", default=None, type=int)
    args.add_argument("-K", "--K", default=3, type=int) #NUMBER OF SHOT

    # args.add_argument("-dim", "--embed_dim", default=100, type=int)
    # args.add_argument("-bs", "--batch_size", default=1024, type=int)
    # args.add_argument("-lr", "--learning_rate", default=0.001, type=float)

    args.add_argument("-epo", "--epoch", default=100000, type=int)
    # args.add_argument("-prt_epo", "--print_epoch", default=100, type=int)
    # args.add_argument("-eval_epo", "--eval_epoch", default=1000, type=int)

    # args.add_argument("-b", "--beta", default=5, type=float)
    # args.add_argument("-m", "--margin", default=1, type=float)
    # args.add_argument("-p", "--dropout_p", default=0.5, type=float)

    # args.add_argument("-gpu", "--device", default=1, type=int)

    args = args.parse_args()
    params = {}
    for k, v in vars(args).items():
        params[k] = v

    # params['device'] = torch.device('cuda:1')
    params['device'] = 0
    return params, args



def main(num_samples, gpus_per_trial=2):
    print("===============",torch.cuda.device_count(),"=======")
    params, args = get_params()

    if params['seed'] is not None:
        SEED = params['seed']
        torch.manual_seed(SEED)
        torch.cuda.manual_seed(SEED)
        torch.backends.cudnn.deterministic = True
        np.random.seed(SEED)
        random.seed(SEED)

    user_train, usernum_train, itemnum, user_input_test, user_test, user_input_valid, user_valid = data_load(args.dataset, args.K)

    # batch_size = params['batch_size']
    # sampler = WarpSampler(user_train, usernum_train, itemnum, batch_size=batch_size, maxlen=args.K, n_workers=1)
    # sampler_test = DataLoader(user_input_test, user_test, itemnum, params)
    # sampler_valid = DataLoader(user_input_valid, user_valid, itemnum, params)

    config = {
        # "l1": tune.sample_from(lambda _: 2 ** np.random.randint(2, 9)),
        # "l2": tune.sample_from(lambda _: 2 ** np.random.randint(2, 9)),
        # "lr": tune.loguniform(1e-4, 1e-1),

        "embed_dim"     :   tune.choice([50,75,100,125,150,200,300]),
        # "batch_size"    :   tune.choice([128,256,512,1024,2048]),
        "learning_rate" :   tune.choice([0.1,0.01,0.004,0.005,0.007,0.001,0.0001]),
        "beta"          :   tune.choice([0.05,0.1,1,4,4.5,5,5.5,6,10]),
        "margin"        :   tune.choice([1,0.9,0.8,1.1,1.2]),

        "batch_size"    :   tune.choice([128,256,512,1024,2048]),
        "number_of_neg" :   tune.choice([1,3,5,7,10,20,30,50,70]),
        "loss_function" :   tune.choice(["bpr"]),
        "eval_epoch"    :   tune.choice([100,250,500,1000,1500]),
        'device'        :   params['device'],


        "itemnum":itemnum,
        "params":params,
    }

    print("===============", torch.cuda.device_count(), "=======")
    scheduler = ASHAScheduler(
        metric="MRR",
        mode="max",
        max_t=params['epoch'],
        grace_period=200,
        reduction_factor=2)
    reporter = CLIReporter(
        # parameter_columns=["l1", "l2", "lr", "batch_size"],
        metric_columns=["MRR","NDCG10","NDCG5","NDCG1","Hits10","Hits5","Hits1","training_iteration"])
    result = tune.run(
        train_metatl,
        resources_per_trial={"cpu": 4, "gpu": gpus_per_trial},
        config=config,
        num_samples=num_samples,
        scheduler=scheduler,
        progress_reporter=reporter,
        log_to_file=True,
        resume=True,
        local_dir="/media/external_10TB/10TB/maheri/metaTL_ray/ray_local_dir",
        name="bpr_rnn",
    )

    best_trial = result.get_best_trial("MRR", "max", "last")
    print("Best trial config: {}".format(best_trial.config))
    print("Best trial final validation loss: {}".format(
        best_trial.last_result["loss"]))
    print("Best trial final validation MRR: {}".format(
        best_trial.last_result["MRR"]))
    print("Best trial final validation NDCG@1: {}".format(
        best_trial.last_result["NDCG@1"]))

    #
    print("=======================================================")
    print(result.results_df)
    print("=======================================================\n")



if __name__ == "__main__":
    # You can change the number of GPUs per trial here:
    main(num_samples=150, gpus_per_trial=0.5)