Melu_L2L_hyperTunning/learnToLearn.py

import os
import torch
import pickle
from options import config
from data_generation import generate
from embedding_module import EmbeddingModule
import learn2learn as l2l
import random
from learnToLearnTest import test
from fast_adapt import fast_adapt
import gc
from learn2learn.optim.transforms import KroneckerTransform
import argparse

def parse_args():
    print("==============")
    parser = argparse.ArgumentParser([], description='Fast Context Adaptation via Meta-Learning (CAVIA),'
                                                     'Clasification experiments.')
    print("==============\n")

    parser.add_argument('--seed', type=int, default=53)
    parser.add_argument('--task', type=str, default='multi', help='problem setting: sine or celeba')
    parser.add_argument('--tasks_per_metaupdate', type=int, default=32,
                        help='number of tasks in each batch per meta-update')

    parser.add_argument('--lr_inner', type=float, default=5e-6, help='inner-loop learning rate (per task)')
    parser.add_argument('--lr_meta', type=float, default=5e-5,
                        help='outer-loop learning rate (used with Adam optimiser)')
    # parser.add_argument('--lr_meta_decay', type=float, default=0.9, help='decay factor for meta learning rate')

    parser.add_argument('--inner', type=int, default=5,
                        help='number of gradient steps in inner loop (during training)')
    parser.add_argument('--inner_eval', type=int, default=5,
                        help='number of gradient updates at test time (for evaluation)')

    parser.add_argument('--first_order', action='store_true', default=False,
                        help='run first order approximation of CAVIA')
    parser.add_argument('--adapt_transform', action='store_true', default=False,
                        help='run adaptation transform')
    parser.add_argument('--transformer', type=str, default="kronoker",
                        help='transformer type')
    parser.add_argument('--meta_algo', type=str, default="gbml",
                        help='MAML/MetaSGD/GBML')
    parser.add_argument('--gpu', type=int, default=0,
                        help='number of gpu to run the code')
    parser.add_argument('--epochs', type=int, default=config['num_epoch'],
                        help='number of gpu to run the code')



    # parser.add_argument('--data_root', type=str, default="./movielens/ml-1m", help='path to data root')
    # parser.add_argument('--num_workers', type=int, default=4, help='num of workers to use')
    # parser.add_argument('--test', action='store_true', default=False, help='num of workers to use')

    # parser.add_argument('--embedding_dim', type=int, default=32, help='num of workers to use')
    # parser.add_argument('--first_fc_hidden_dim', type=int, default=64, help='num of workers to use')
    # parser.add_argument('--second_fc_hidden_dim', type=int, default=64, help='num of workers to use')
    # parser.add_argument('--num_epoch', type=int, default=30, help='num of workers to use')
    # parser.add_argument('--num_genre', type=int, default=25, help='num of workers to use')
    # parser.add_argument('--num_director', type=int, default=2186, help='num of workers to use')
    # parser.add_argument('--num_actor', type=int, default=8030, help='num of workers to use')
    # parser.add_argument('--num_rate', type=int, default=6, help='num of workers to use')
    # parser.add_argument('--num_gender', type=int, default=2, help='num of workers to use')
    # parser.add_argument('--num_age', type=int, default=7, help='num of workers to use')
    # parser.add_argument('--num_occupation', type=int, default=21, help='num of workers to use')
    # parser.add_argument('--num_zipcode', type=int, default=3402, help='num of workers to use')

    # parser.add_argument('--rerun', action='store_true', default=False,
    #                     help='Re-run experiment (will override previously saved results)')

    args = parser.parse_args()
    # use the GPU if available
    # args.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    # print('Running on device: {}'.format(args.device))
    return args

if __name__ == '__main__':
    args = parse_args()
    print(args)

    if config['use_cuda']:
        os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
        os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
    master_path= "/media/external_10TB/10TB/maheri/melu_data5"

    # DATA GENERATION
    print("DATA GENERATION PHASE")
    if not os.path.exists("{}/".format(master_path)):
        os.mkdir("{}/".format(master_path))
        # preparing dataset. It needs about 22GB of your hard disk space.
        generate(master_path)

    # TRAINING
    print("TRAINING PHASE")
    embedding_dim = config['embedding_dim']
    fc1_in_dim = config['embedding_dim'] * 8
    fc2_in_dim = config['first_fc_hidden_dim']
    fc2_out_dim = config['second_fc_hidden_dim']
    use_cuda = config['use_cuda']

    fc1 = torch.nn.Linear(fc1_in_dim, fc2_in_dim)
    fc2 = torch.nn.Linear(fc2_in_dim, fc2_out_dim)
    linear_out = torch.nn.Linear(fc2_out_dim, 1)
    head = torch.nn.Sequential(fc1,fc2,linear_out)

    if use_cuda:
        emb = EmbeddingModule(config).cuda()
    else:
        emb = EmbeddingModule(config)

    # META LEARNING
    print("META LEARNING PHASE")

    # define transformer
    transform = None
    if args.transformer == "kronoker":
        transform = KroneckerTransform(l2l.nn.KroneckerLinear)
    elif args.transformer == "linear":
        transform = l2l.optim.ModuleTransform(torch.nn.Linear)

    # define meta algorithm
    if args.meta_algo == "maml":
        head = l2l.algorithms.MAML(head, lr=args.lr_inner,first_order=args.first_order)
    elif args.meta_algo == 'metasgd':
        head = l2l.algorithms.MetaSGD(head, lr=args.lr_inner,first_order=args.first_order)
    elif args.meta_algo == 'gbml':
        head = l2l.algorithms.GBML(head, transform=transform, lr=args.lr_inner,adapt_transform=args.adapt_transform, first_order=args.first_order)

    if use_cuda:
        head.cuda()

    # Setup optimization
    print("SETUP OPTIMIZATION PHASE")
    all_parameters =  list(emb.parameters()) + list(head.parameters())
    optimizer = torch.optim.Adam(all_parameters, lr=args.lr_meta)
    # loss = torch.nn.MSELoss(reduction='mean')

    # Load training dataset.
    print("LOAD DATASET PHASE")
    training_set_size = int(len(os.listdir("{}/warm_state".format(master_path))) / 4)
    supp_xs_s = []
    supp_ys_s = []
    query_xs_s = []
    query_ys_s = []
    for idx in range(training_set_size):
        supp_xs_s.append(pickle.load(open("{}/warm_state/supp_x_{}.pkl".format(master_path, idx), "rb")))
        supp_ys_s.append(pickle.load(open("{}/warm_state/supp_y_{}.pkl".format(master_path, idx), "rb")))
        query_xs_s.append(pickle.load(open("{}/warm_state/query_x_{}.pkl".format(master_path, idx), "rb")))
        query_ys_s.append(pickle.load(open("{}/warm_state/query_y_{}.pkl".format(master_path, idx), "rb")))
    total_dataset = list(zip(supp_xs_s, supp_ys_s, query_xs_s, query_ys_s))
    del(supp_xs_s, supp_ys_s, query_xs_s, query_ys_s)
    training_set_size = len(total_dataset)
    batch_size = config['batch_size']
    # torch.cuda.empty_cache()

    random.shuffle(total_dataset)
    num_batch = int(training_set_size / batch_size)
    a, b, c, d = zip(*total_dataset)

    print("\n\n\n")
    for iteration in range(args.epochs):
        for i in range(num_batch):
            optimizer.zero_grad()
            meta_train_error = 0.0
            meta_train_accuracy = 0.0
            meta_valid_error = 0.0
            meta_valid_accuracy = 0.0
            meta_test_error = 0.0
            meta_test_accuracy = 0.0

            print("EPOCH: ", iteration, " BATCH: ", i)
            supp_xs = list(a[batch_size * i:batch_size * (i + 1)])
            supp_ys = list(b[batch_size * i:batch_size * (i + 1)])
            query_xs = list(c[batch_size * i:batch_size * (i + 1)])
            query_ys = list(d[batch_size * i:batch_size * (i + 1)])
            batch_sz = len(supp_xs)

            # if use_cuda:
            #     for j in range(batch_size):
            #         supp_xs[j] = supp_xs[j].cuda()
            #         supp_ys[j] = supp_ys[j].cuda()
            #         query_xs[j] = query_xs[j].cuda()
            #         query_ys[j] = query_ys[j].cuda()

            for task in range(batch_sz):
                # print("EPOCH: ", iteration," BATCH: ",i, "TASK: ",task)
                # Compute meta-training loss
                # if use_cuda:
                sxs = supp_xs[task].cuda()
                qxs = query_xs[task].cuda()
                sys = supp_ys[task].cuda()
                qys = query_ys[task].cuda()

                learner = head.clone()
                temp_sxs = emb(sxs)
                temp_qxs = emb(qxs)

                evaluation_error = fast_adapt(learner,
                                               temp_sxs,
                                               temp_qxs,
                                               sys,
                                               qys,
                                               args.inner)
                                               # config['inner'])

                evaluation_error.backward()
                meta_train_error += evaluation_error.item()

                del(sxs,qxs,sys,qys)
                supp_xs[task].cpu()
                query_xs[task].cpu()
                supp_ys[task].cpu()
                query_ys[task].cpu()


            # Print some metrics
            print('Iteration', iteration)
            print('Meta Train Error', meta_train_error / batch_sz)
            # print('Meta Train Accuracy', meta_train_accuracy / batch_sz)
            # print('Meta Valid Error', meta_valid_error / batch_sz)
            # print('Meta Valid Accuracy', meta_valid_accuracy / batch_sz)
            # print('Meta Test Error', meta_test_error / batch_sz)
            # print('Meta Test Accuracy', meta_test_accuracy / batch_sz)

            # Average the accumulated gradients and optimize
            for p in all_parameters:
                p.grad.data.mul_(1.0 / batch_sz)
            optimizer.step()

            # torch.cuda.empty_cache()
            del(supp_xs,supp_ys,query_xs,query_ys)
            gc.collect()
            print("===============================================\n")


    # save model
    final_model = torch.nn.Sequential(emb,head)
    torch.save(final_model.state_dict(), master_path + "/models_gbml.pkl")

    # testing
    print("start of test phase")
    for test_state in ['warm_state', 'user_cold_state', 'item_cold_state', 'user_and_item_cold_state']:
        test_dataset = None
        test_set_size = int(len(os.listdir("{}/{}".format(master_path, test_state))) / 4)
        supp_xs_s = []
        supp_ys_s = []
        query_xs_s = []
        query_ys_s = []
        for idx in range(test_set_size):
            supp_xs_s.append(pickle.load(open("{}/{}/supp_x_{}.pkl".format(master_path, test_state, idx), "rb")))
            supp_ys_s.append(pickle.load(open("{}/{}/supp_y_{}.pkl".format(master_path, test_state, idx), "rb")))
            query_xs_s.append(pickle.load(open("{}/{}/query_x_{}.pkl".format(master_path, test_state, idx), "rb")))
            query_ys_s.append(pickle.load(open("{}/{}/query_y_{}.pkl".format(master_path, test_state, idx), "rb")))
        test_dataset = list(zip(supp_xs_s, supp_ys_s, query_xs_s, query_ys_s))
        del (supp_xs_s, supp_ys_s, query_xs_s, query_ys_s)

        print("===================== " + test_state + " =====================")
        test(emb,head, test_dataset, batch_size=config['batch_size'], num_epoch=args.epochs,adaptation_step=args.inner_eval)
        print("===================================================\n\n\n")
    print(args)