263 lines
8.7 KiB
Python
263 lines
8.7 KiB
Python
#####################################################
|
|
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 #
|
|
#####################################################
|
|
# python exps/LFNA/lfna.py --env_version v1
|
|
#####################################################
|
|
import sys, time, copy, torch, random, argparse
|
|
from tqdm import tqdm
|
|
from copy import deepcopy
|
|
from pathlib import Path
|
|
|
|
lib_dir = (Path(__file__).parent / ".." / ".." / "lib").resolve()
|
|
if str(lib_dir) not in sys.path:
|
|
sys.path.insert(0, str(lib_dir))
|
|
from procedures import prepare_seed, prepare_logger, save_checkpoint, copy_checkpoint
|
|
from log_utils import time_string
|
|
from log_utils import AverageMeter, convert_secs2time
|
|
|
|
from utils import split_str2indexes
|
|
|
|
from procedures.advanced_main import basic_train_fn, basic_eval_fn
|
|
from procedures.metric_utils import SaveMetric, MSEMetric, ComposeMetric
|
|
from datasets.synthetic_core import get_synthetic_env, EnvSampler
|
|
from models.xcore import get_model
|
|
from xlayers import super_core, trunc_normal_
|
|
|
|
from lfna_utils import lfna_setup, train_model, TimeData
|
|
from lfna_meta_model import LFNA_Meta
|
|
|
|
|
|
def epoch_train(loader, meta_model, base_model, optimizer, criterion, device, logger):
|
|
base_model.train()
|
|
meta_model.train()
|
|
loss_meter = AverageMeter()
|
|
for ibatch, batch_data in enumerate(loader):
|
|
timestamps, (batch_seq_inputs, batch_seq_targets) = batch_data
|
|
timestamps = timestamps.squeeze(dim=-1).to(device)
|
|
batch_seq_inputs = batch_seq_inputs.to(device)
|
|
batch_seq_targets = batch_seq_targets.to(device)
|
|
|
|
optimizer.zero_grad()
|
|
|
|
batch_seq_containers = meta_model(timestamps)
|
|
losses = []
|
|
for seq_containers, seq_inputs, seq_targets in zip(
|
|
batch_seq_containers, batch_seq_inputs, batch_seq_targets
|
|
):
|
|
for container, inputs, targets in zip(
|
|
seq_containers, seq_inputs, seq_targets
|
|
):
|
|
predictions = base_model.forward_with_container(inputs, container)
|
|
loss = criterion(predictions, targets)
|
|
losses.append(loss)
|
|
final_loss = torch.stack(losses).mean()
|
|
final_loss.backward()
|
|
optimizer.step()
|
|
loss_meter.update(final_loss.item())
|
|
return loss_meter
|
|
|
|
|
|
def main(args):
|
|
logger, env_info, model_kwargs = lfna_setup(args)
|
|
dynamic_env = get_synthetic_env(mode="train", version=args.env_version)
|
|
base_model = get_model(**model_kwargs)
|
|
base_model = base_model.to(args.device)
|
|
criterion = torch.nn.MSELoss()
|
|
|
|
shape_container = base_model.get_w_container().to_shape_container()
|
|
|
|
# pre-train the hypernetwork
|
|
timestamps = dynamic_env.get_timestamp(None)
|
|
meta_model = LFNA_Meta(shape_container, args.layer_dim, args.time_dim, timestamps)
|
|
meta_model = meta_model.to(args.device)
|
|
|
|
logger.log("The base-model has {:} weights.".format(base_model.numel()))
|
|
logger.log("The meta-model has {:} weights.".format(meta_model.numel()))
|
|
|
|
batch_sampler = EnvSampler(dynamic_env, args.meta_batch, args.sampler_enlarge)
|
|
dynamic_env.reset_max_seq_length(args.seq_length)
|
|
"""
|
|
env_loader = torch.utils.data.DataLoader(
|
|
dynamic_env,
|
|
batch_size=args.meta_batch,
|
|
shuffle=True,
|
|
num_workers=args.workers,
|
|
pin_memory=True,
|
|
)
|
|
"""
|
|
env_loader = torch.utils.data.DataLoader(
|
|
dynamic_env,
|
|
batch_sampler=batch_sampler,
|
|
num_workers=args.workers,
|
|
pin_memory=True,
|
|
)
|
|
|
|
optimizer = torch.optim.Adam(
|
|
meta_model.parameters(), lr=args.init_lr, weight_decay=1e-5, amsgrad=True
|
|
)
|
|
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
|
|
optimizer,
|
|
milestones=[
|
|
int(args.epochs * 0.8),
|
|
int(args.epochs * 0.9),
|
|
],
|
|
gamma=0.1,
|
|
)
|
|
logger.log("The base-model is\n{:}".format(base_model))
|
|
logger.log("The meta-model is\n{:}".format(meta_model))
|
|
logger.log("The optimizer is\n{:}".format(optimizer))
|
|
logger.log("Per epoch iterations = {:}".format(len(env_loader)))
|
|
|
|
# LFNA meta-training
|
|
per_epoch_time, start_time = AverageMeter(), time.time()
|
|
last_success_epoch = 0
|
|
for iepoch in range(args.epochs):
|
|
|
|
head_str = "[{:}] [{:04d}/{:04d}] ".format(
|
|
time_string(), iepoch, args.epochs
|
|
) + "Time Left: {:}".format(
|
|
convert_secs2time(per_epoch_time.avg * (args.epochs - iepoch), True)
|
|
)
|
|
|
|
loss_meter = epoch_train(
|
|
env_loader,
|
|
meta_model,
|
|
base_model,
|
|
optimizer,
|
|
criterion,
|
|
args.device,
|
|
logger,
|
|
)
|
|
lr_scheduler.step()
|
|
logger.log(
|
|
head_str
|
|
+ " meta-loss: {meter.avg:.4f} ({meter.count:.0f})".format(meter=loss_meter)
|
|
+ " :: lr={:.5f}".format(min(lr_scheduler.get_last_lr()))
|
|
)
|
|
success, best_score = meta_model.save_best(-loss_meter.avg)
|
|
if success:
|
|
logger.log("Achieve the best with best_score = {:.3f}".format(best_score))
|
|
last_success_epoch = iepoch
|
|
save_checkpoint(
|
|
{
|
|
"meta_model": meta_model.state_dict(),
|
|
"optimizer": optimizer.state_dict(),
|
|
"lr_scheduler": lr_scheduler.state_dict(),
|
|
"iepoch": iepoch,
|
|
"args": args,
|
|
},
|
|
logger.path("model"),
|
|
logger,
|
|
)
|
|
if iepoch - last_success_epoch >= args.early_stop_thresh:
|
|
logger.log("Early stop at {:}".format(iepoch))
|
|
break
|
|
|
|
per_epoch_time.update(time.time() - start_time)
|
|
start_time = time.time()
|
|
|
|
w_container_per_epoch = dict()
|
|
for idx in range(0, total_bar):
|
|
future_time = env_info["{:}-timestamp".format(idx)]
|
|
future_x = env_info["{:}-x".format(idx)]
|
|
future_y = env_info["{:}-y".format(idx)]
|
|
future_container = hypernet(task_embeds[idx])
|
|
w_container_per_epoch[idx] = future_container.no_grad_clone()
|
|
with torch.no_grad():
|
|
future_y_hat = model.forward_with_container(
|
|
future_x, w_container_per_epoch[idx]
|
|
)
|
|
future_loss = criterion(future_y_hat, future_y)
|
|
logger.log("meta-test: [{:03d}] -> loss={:.4f}".format(idx, future_loss.item()))
|
|
|
|
save_checkpoint(
|
|
{"w_container_per_epoch": w_container_per_epoch},
|
|
logger.path(None) / "final-ckp.pth",
|
|
logger,
|
|
)
|
|
|
|
logger.log("-" * 200 + "\n")
|
|
logger.close()
|
|
|
|
|
|
if __name__ == "__main__":
|
|
parser = argparse.ArgumentParser(".")
|
|
parser.add_argument(
|
|
"--save_dir",
|
|
type=str,
|
|
default="./outputs/lfna-synthetic/lfna-battle",
|
|
help="The checkpoint directory.",
|
|
)
|
|
parser.add_argument(
|
|
"--env_version",
|
|
type=str,
|
|
required=True,
|
|
help="The synthetic enviornment version.",
|
|
)
|
|
parser.add_argument(
|
|
"--hidden_dim",
|
|
type=int,
|
|
default=16,
|
|
help="The hidden dimension.",
|
|
)
|
|
parser.add_argument(
|
|
"--layer_dim",
|
|
type=int,
|
|
default=16,
|
|
help="The layer chunk dimension.",
|
|
)
|
|
parser.add_argument(
|
|
"--time_dim",
|
|
type=int,
|
|
default=16,
|
|
help="The timestamp dimension.",
|
|
)
|
|
#####
|
|
parser.add_argument(
|
|
"--init_lr",
|
|
type=float,
|
|
default=0.01,
|
|
help="The initial learning rate for the optimizer (default is Adam)",
|
|
)
|
|
parser.add_argument(
|
|
"--meta_batch",
|
|
type=int,
|
|
default=64,
|
|
help="The batch size for the meta-model",
|
|
)
|
|
parser.add_argument(
|
|
"--sampler_enlarge",
|
|
type=int,
|
|
default=5,
|
|
help="Enlarge the #iterations for an epoch",
|
|
)
|
|
parser.add_argument("--epochs", type=int, default=1000, help="The total #epochs.")
|
|
parser.add_argument(
|
|
"--early_stop_thresh",
|
|
type=int,
|
|
default=50,
|
|
help="The maximum epochs for early stop.",
|
|
)
|
|
parser.add_argument(
|
|
"--seq_length", type=int, default=5, help="The sequence length."
|
|
)
|
|
parser.add_argument(
|
|
"--workers", type=int, default=4, help="The number of workers in parallel."
|
|
)
|
|
parser.add_argument(
|
|
"--device",
|
|
type=str,
|
|
default="cpu",
|
|
help="",
|
|
)
|
|
# Random Seed
|
|
parser.add_argument("--rand_seed", type=int, default=-1, help="manual seed")
|
|
args = parser.parse_args()
|
|
if args.rand_seed is None or args.rand_seed < 0:
|
|
args.rand_seed = random.randint(1, 100000)
|
|
assert args.save_dir is not None, "The save dir argument can not be None"
|
|
args.save_dir = "{:}-{:}-d{:}_{:}_{:}".format(
|
|
args.save_dir, args.env_version, args.hidden_dim, args.layer_dim, args.time_dim
|
|
)
|
|
main(args)
|