File size: 8,705 Bytes
0ca2ffc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 | import os
import os.path as osp
import numpy as np
import torch
from collections import OrderedDict
import argparse
import re
from torch_geometric.loader import DataLoader, GraphSAINTRandomWalkSampler
import sys
sys.path.append(".")
from graphmetro.transform import get_trans_func_dict
from graphmetro.model_builder import create_model
from graphmetro.datasets import get_datasets
from graphmetro.config import cfg, dump_cfg, load_cfg, set_out_dir
from graphmetro.utils.device import auto_select_device
from graphmetro.utils.comp_budet import match_baseline_cfg
from graphmetro.train.optimizer import create_optimizer, create_scheduler
from graphmetro.train.train_standard import train
from graphmetro.logger import Logger
from graphmetro.utils.seed import set_seed
from graphmetro.utils.wandb import setup_wandb
from graphmetro.transform import *
from graphmetro.models.layers import MLP, Wrapper
import warnings
warnings.simplefilter("ignore")
np.set_printoptions(precision=4)
def parse_args() -> argparse.Namespace:
r"""Parses the command line arguments."""
parser = argparse.ArgumentParser(description='graphmetro')
parser.add_argument('--cfg',
dest='cfg_file',
type=str,
default=None,
help='The configuration file path.')
parser.add_argument("--use_wandb",
type=int, default=1
)
parser.add_argument("--model_path",
type=str, default=None,
help='The model path of the trained GNN'
)
parser.add_argument("--wandb_id",
type=str, default='<your team name here>',
help='The wandb username for logging'
)
parser.add_argument("--wandb_project",
type=str, default='graphmetro',
help='The wandb project name for logging'
)
parser.add_argument("--wandb_exp_name",
type=str, default=None,
help='The wandb experiment name for logging'
)
parser.add_argument('opts',
default=None,
nargs=argparse.REMAINDER,
help='See graphgym/config.py for remaining options.')
return parser.parse_args()
auto_select_device()
args = parse_args()
load_cfg(cfg, args)
set_seed(cfg.seed)
set_out_dir(cfg)
if args.use_wandb:
setup_wandb(cfg, args)
torch.set_num_threads(cfg.num_threads)
os.makedirs(cfg.dataset.dir, exist_ok=True)
datasets = get_datasets(cfg.dataset.name, root=cfg.dataset.dir)
data = datasets['train'][0]
node_feat_dim = data.x.size(-1)
edge_feat_dim = 1 if (data.edge_attr is None or len(data.edge_attr.size()) == 1) \
else data.edge_attr.size(-1)
logger = Logger.get_logger(name='main', fname=osp.join(cfg.out_dir, 'output.log'))
logger.info(f'Training data {data} \n data.x {data.x[:5]}')
batch_size = cfg.train.batch_size
loaders = {}
loaders['train'] = DataLoader(datasets['train'],
batch_size=batch_size,
shuffle=True
)
loaders['val'] = DataLoader(datasets['val'],
batch_size=batch_size,
shuffle=False
)
loaders['test'] = DataLoader(datasets['test'],
batch_size=batch_size,
shuffle=False
)
trans_func_names = ['id'] + list(set(re.split('-|/', cfg.shift.train_types)))
trans_func_dict = get_trans_func_dict(trans_func_names)
logger.info(f'Transform functions: {list(trans_func_dict.keys())}')
if cfg.train.mode == 'standard' or \
(cfg.train.mode == 'moe_shared' and args.model_path is None):
if args.model_path:
model = torch.load(args.model_path).to(cfg.device)
logger.info(f'Resume from {args.model_path}')
else:
model = create_model(datasets['train'])
optimizer = create_optimizer(model.parameters(), lr=cfg.optim.standard_lr)
scheduler = create_scheduler(optimizer)
augment = list(trans_func_dict.keys()) if cfg.train.augment else None
best_result, last_result = train(model,
loaders,
optimizer,
scheduler,
augment
)
torch.save(model.cpu(), osp.join(cfg.out_dir, 'last_standard_model.pt'))
if cfg.train.mode == 'moe_shared':
from graphmetro.models.moe_shared import MoEModel
from graphmetro.train.train_moe_shared import train_moe
gating_model = create_model(datasets['train'], dim_out=len(trans_func_dict))
if args.model_path is None:
args.model_path = osp.join(cfg.out_dir, 'best_standard_model.pt')
model = MoEModel(trans_func_dict=trans_func_dict,
gating_model=gating_model
).to(cfg.device)
optimizer = create_optimizer(model.parameters(), lr=cfg.optim.moe_lr)
scheduler = create_scheduler(optimizer)
gnn_model = torch.load(args.model_path).to(cfg.device)
best_result, last_result = train_moe(model,
gnn_model,
loaders,
optimizer,
scheduler
)
torch.save(model.cpu(), osp.join(cfg.out_dir, 'last_moe_model.pt'))
if cfg.train.mode == 'moe':
from graphmetro.models.moe import MoEModel
from graphmetro.train.train_moe import train_moe
gating_model = create_model(datasets['train'], dim_out=len(trans_func_dict))
if args.model_path:
base_model = torch.load(args.model_path)
logger.info(f'Resume from {args.model_path}')
else:
base_model = create_model(datasets['train'], has_post_mp=True)
if cfg.train.moe_shared:
encoder = base_model.encoder
detach = False if args.model_path is None else True
if detach:
for params in encoder.parameters():
params.requires_grad = False
expert_models = torch.nn.ModuleList([base_model] +
[Wrapper(encoder, MLP(dim_out=cfg.gnn.dim_inner,
skip_connection=True,
act=torch.nn.Sigmoid()))
for _ in range(len(trans_func_dict) - 1)
])
else:
expert_models = torch.nn.ModuleList([base_model] +
[create_model(datasets['train'], has_post_mp=False) \
for _ in range(len(trans_func_dict) - 1)])
model = MoEModel(trans_func_dict=trans_func_dict,
gating_model=gating_model,
expert_models=expert_models
).to(cfg.device)
optimizer = torch.optim.Adam([
{'params': model.gating_model.parameters(), 'lr': cfg.optim.moe_lr},
{'params': model.experts.parameters(), 'lr': cfg.optim.moe_lr},
{'params': model.classifier.parameters(), 'lr': cfg.optim.standard_lr}
], weight_decay=cfg.optim.weight_decay
)
scheduler = create_scheduler(optimizer)
best_result, last_result = train_moe(model,
loaders,
optimizer,
scheduler
)
torch.save(model.cpu(), osp.join(cfg.out_dir, 'last_moe_model.pt'))
# logs
logger.info('=' * 100)
logger.info('Best result' + f'{best_result}')
logger.info('Last result' + f'{last_result}')
import csv
if (not os.path.isfile(osp.join(cfg.out_dir, '..', 'best_result.csv'))):
writeheader = True
else:
writeheader = False
with open(osp.join(cfg.out_dir, '..', 'best_result.csv'), 'a+') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=['exp_name', 'best_result'])
if writeheader: writer.writeheader()
writer.writerows([{'exp_name': cfg.out_dir.split('/')[-1], 'best_result': best_result}])
with open(osp.join(cfg.out_dir, '..', 'last_result.csv'), 'a+') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=['exp_name', 'last_result'])
if writeheader: writer.writeheader()
writer.writerows([{'exp_name': cfg.out_dir.split('/')[-1], 'last_result': best_result}]) |