config.py

'''Configuration'''
import os
import sys

if os.path.abspath('..') not in sys.path:
    sys.path.append(os.path.abspath('..'))
import json
import logging

import click
from pprint import pformat
from utils import accuracy
import torch 

class Obj(object):
    def __init__(self, dict_):
        self.__dict__.update(dict_)

    def __repr__(self):
        # Use pprint's pformat to print the dictionary in a pretty manner
        return pformat(self.__dict__, compact=True)


def dict2obj(d):
    return json.loads(json.dumps(d), object_hook=Obj)


def update_from_dict(obj, updates):
    for key, value in updates.items():
        # set higher priority from command line as we explore some factors
        if key in ['init'] and obj.init is not None:
            continue
        setattr(obj, key, value)


# fix setting here
def setting_config(args):
    representative_r = {
        'cora': 0.5,
        'flickr': 0.01,
        'ogbn-arxiv': 0.01
    }
    if args.reduction_rate == -1:
        args.reduction_rate = representative_r[args.dataset]
    if args.dataset in ['cora', 'ogbn-arxiv']:
        args.setting = 'trans'
    if args.dataset in ['flickr']:
        args.setting = 'ind'

    args.metric = accuracy

    args.run_inter_eval = 3
    args.eval_interval = args.epochs // 10

    args.checkpoints = list(range(-1, args.epochs + 1, args.eval_interval))
    args.eval_epochs = 300
    return args


# recommend hyperparameters here
def method_config(args):
    try:
        # print(os.path.dirname(graphslim.__file__))
        conf_dt = json.load(
            open(f"{os.path.join('configs', args.method, args.dataset)}.json"))
        update_from_dict(args, conf_dt)
    except:
        print('No config file found or error in json format, please use method_config(args)')

    return args


@click.command()
@click.option('--dataset', '-D', default='cora', show_default=True)
@click.option('--gpu_id', '-G', default=0, help='gpu id start from 0, -1 means cpu', show_default=True)
@click.option('--setting', type=click.Choice(['trans', 'ind']), show_default=True,
              help='transductive or inductive setting')
@click.option('--split', default='fixed', show_default=True,
              help='only support public split now, do not change it')  # 'fixed', 'random', 'few'
@click.option('--run_reduction', default=3, show_default=True, help='repeat times of reduction')
@click.option('--run_eval', default=10, show_default=True, help='repeat times of final evaluations')
@click.option('--run_inter_eval', default=5, show_default=True, help='repeat times of intermediate evaluations')
@click.option('--eval_interval', default=100, show_default=True)
@click.option('--hidden', '-H', default=256, show_default=True)
@click.option('--eval_epochs', '--ee', default=300, show_default=True)
@click.option('--eval_model', '--em', default='GCN',
              type=click.Choice(
                  ['GCN']
              ), show_default=True)
@click.option('--final_eval_model', '--fem', default='GCN',
              type=click.Choice(
                  ['GCN']
              ), show_default=True)
@click.option('--condense_model', default='SGC',
              type=click.Choice(
                  ['GCN', 'SGC']
              ), show_default=True)
@click.option('--epochs', '-E', default=1000, show_default=True, help='number of reduction epochs')
@click.option('--lr', default=0.01, show_default=True)
@click.option('--weight_decay', '--wd', default=0.0, show_default=True)

@click.option('--pre_norm', default=True, show_default=True,
              help='pre-normalize features, forced true for arxiv, flickr and reddit')
@click.option('--outer_loop', default=10, show_default=True)
@click.option('--inner_loop', default=1, show_default=True)
@click.option('--reduction_rate', '-R', default=-1.0, show_default=True,
              help='-1 means use representative reduction rate; reduction rate of training set, defined as (number of nodes in small graph)/(number of nodes in original graph)')
@click.option('--seed', '-S', default=1, help='Random seed', show_default=True)
@click.option('--nlayers', default=2, help='number of GNN layers of condensed model', show_default=True)
@click.option('--verbose', '-V', is_flag=True, show_default=True)
@click.option('--soft_label', default=0, show_default=True)
@click.option('--init', default='random', help='features initialization methods',
              type=click.Choice(
                  ['variation_neighborhoods', 'variation_edges', 'variation_cliques', 'heavy_edge', 'algebraic_JC',
                   'affinity_GS', 'kron', 'vng', 'clustering', 'averaging',
                   'cent_d', 'cent_p', 'kcenter', 'herding', 'random']
              ), show_default=True)

@click.option('--eval_wd', '--ewd', default=0.0, show_default=True)
@click.option('--eval_loss', '--eloss', default='CE',
                type=click.Choice(
                  ['CE', 'KLD','MSE']
              ), show_default=True)
@click.option('--method', '-M', default='hdgc', show_default=True)
@click.option('--activation', default='relu', help='activation function when do NAS',
              type=click.Choice(
                  ['sigmoid', 'tanh', 'relu', 'linear', 'softplus', 'leakyrelu', 'relu6', 'elu']
              ), show_default=True)
@click.option('--attack', '-A', default=None, help='corruption method',
              type=click.Choice(
                  ['random_adj', 'metattack', 'random_feat']
              ), show_default=True)
@click.option('--coarsen_strategy', '--cs', default='greedy', help='for edge contraction method',
              type=click.Choice(
                  ['optimal', 'greedy']
              ), show_default=True)
@click.option('--ptb_r', '-P', default=0.25, show_default=True, help='perturbation rate for corruptions')
@click.option('--agg', is_flag=True, show_default=True, help='use aggregation for coreset methods')
@click.option('--multi_label', is_flag=True, show_default=True, help='use aggregation for coreset methods')
@click.option('--dis_metric', default='ours', show_default=True,
              help='distance metric for all condensation methods,ours means metric used in GCond paper')
@click.option('--lr_adj', default=1e-4, show_default=True)
@click.option('--lr_feat', default=1e-4, show_default=True)
@click.option('--optim', default="Adam", show_default=True)
@click.option('--threshold', default=0.0, show_default=True, help='sparsificaiton threshold before evaluation')
@click.option('--dropout', default=0.0, show_default=True)
@click.option('--ntrans', default=1, show_default=True, help='number of transformations in SGC and APPNP')
@click.option('--with_bn', is_flag=True, show_default=True)
@click.option('--no_buff', is_flag=True, show_default=True,
              help='skip the buffer generation and use existing in geom,sfgc')
@click.option('--batch_adj', default=1, show_default=True, help='batch size for msgc')
# model specific args
@click.option('--alpha', default=0.1, help='for appnp', show_default=True)
@click.option('--mx_size', default=100, help='for gcsntk methods, avoid SVD error', show_default=True)
@click.option('--ts', default=4, help='for tspanner', show_default=True)
@click.option('--save_path', '--sp', default='.', show_default=True, help='save path for synthetic graph')
@click.option('--load_path', '--lp', default='data', show_default=True, help='save path for synthetic graph')
@click.option('--eval_whole', '-W', is_flag=True, show_default=True, help='if run on whole graph')
@click.option('--with_structure', default=1, show_default=True, help='if synthesizing structure')
# ======================================simgc====================================== #
@click.option('--feat_alpha', default=10, show_default=True, help='feature loss weight')
@click.option('--smoothness_alpha', default=0.1, show_default=True, help='smoothness loss weight')
#==================gdem=====#
@click.option('--eigen_k', default=60, show_default=True, help='number of eigenvalues')
@click.option('--ratio', default=0.8, show_default=True, help='eigenvalue loss weight')
@click.option('--lr_eigenvec', default=0.01, show_default=True, help='eigenvalue loss weight')
@click.option('--gamma', default=0.5, show_default=True, help='eigenvalue loss weight')

@click.option('--min_noise_scale', default=0.1, show_default=True, help='min noise scale')
@click.option('--target_delta', default=1e-5, show_default=True, help='target delta')
@click.option('--chi_square_rho', default=0.5, show_default=True, help='chi square rho')
@click.option('--dro_alpha', default=0.1, show_default=True, help='dro alpha')
@click.option('--max_grad_norm', default=1.0, show_default=True, help='max grad norm')
@click.option('--min_noise_multiplier', default=0.01, show_default=True, help='min noise multiplier')
@click.option('--apply_dp', is_flag=False, show_default=False, help='apply dp')
@click.option('--apply_dro', is_flag=False, show_default=False, help='apply dro')
@click.option('--target_epsilon', '-te', default=0.0, show_default=True, help='target epsilon')
@click.option('--down_epsilon', '--de', default=0.0, show_default=True, help='down epsilon')

@click.pass_context
def cli(ctx, **kwargs):
    args = dict2obj(kwargs)
    if args.gpu_id >= 0 and torch.cuda.is_available():
        os.environ["CUDA_VISIBLE_DEVICES"] = f"{args.gpu_id}"
        args.device = f'cuda:0'
    else:
        args.device = 'cpu'
    path = args.save_path
    # for benchmark, we need unified settings and reduce flexibility of args
    args = method_config(args)
    # setting_config has higher priority than methods_config
    args = setting_config(args)
    for key, value in ctx.params.items():
        if ctx.get_parameter_source(key) == click.core.ParameterSource.COMMANDLINE:
            setattr(args, key, value)
    if not os.path.exists(f'{path}/logs/{args.method}'):
        try:
            os.makedirs(f'{path}/logs/{args.method}')
        except:
            print(f'{path}/logs/{args.method} exists!')
    logging.basicConfig(filename=f'{path}/logs/{args.method}/{args.dataset}_{args.reduction_rate}.log',
                        level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
    args.logger = logging.getLogger(__name__)
    args.logger.addHandler(logging.StreamHandler())
    args.logger.info(args)
    return args


def get_args():
    return cli(standalone_mode=False)


if __name__ == '__main__':
    cli()