iicd/autodl-projects
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Prototype MAML

Browse Source
This commit is contained in:
D-X-Y
2021-05-10 01:02:38 +08:00
parent 6e7b1c551f
commit cbd2afb4ef
14 changed files with 1497 additions and 702 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
exps/LFNA/basic-his.py
View File

@@ -23,6 +23,9 @@ from datasets.synthetic_core import get_synthetic_env
from models.xcore import get_model
from lfna_utils import lfna_setup
def subsample(historical_x, historical_y, maxn=10000):
total = historical_x.size(0)
if total <= maxn:
@@ -33,24 +36,7 @@ def subsample(historical_x, historical_y, maxn=10000):
def main(args):
prepare_seed(args.rand_seed)
logger = prepare_logger(args)
cache_path = (
logger.path(None) / ".." / "env-{:}-info.pth".format(args.env_version)
).resolve()
if cache_path.exists():
env_info = torch.load(cache_path)
else:
env_info = dict()
dynamic_env = get_synthetic_env(version=args.env_version)
env_info["total"] = len(dynamic_env)
for idx, (timestamp, (_allx, _ally)) in enumerate(tqdm(dynamic_env)):
env_info["{:}-timestamp".format(idx)] = timestamp
env_info["{:}-x".format(idx)] = _allx
env_info["{:}-y".format(idx)] = _ally
env_info["dynamic_env"] = dynamic_env
torch.save(env_info, cache_path)
logger, env_info = lfna_setup(args)
# check indexes to be evaluated
to_evaluate_indexes = split_str2indexes(args.srange, env_info["total"], None)
@@ -60,6 +46,8 @@ def main(args):
)
)
w_container_per_epoch = dict()
per_timestamp_time, start_time = AverageMeter(), time.time()
for i, idx in enumerate(to_evaluate_indexes):
@@ -89,9 +77,6 @@ def main(args):
output_dim=1,
act_cls="leaky_relu",
norm_cls="identity",
# norm_cls="simple_norm",
# mean=mean,
# std=std,
)
model = get_model(dict(model_type="simple_mlp"), **model_kwargs)
# build optimizer
@@ -144,6 +129,7 @@ def main(args):
save_path = logger.path(None) / "{:04d}-{:04d}.pth".format(
idx, env_info["total"]
)
w_container_per_epoch[idx] = model.get_w_container().no_grad_clone()
save_checkpoint(
{
"model_state_dict": model.state_dict(),
@@ -155,10 +141,14 @@ def main(args):
logger,
)
logger.log("")
per_timestamp_time.update(time.time() - start_time)
start_time = time.time()
save_checkpoint(
{"w_container_per_epoch": w_container_per_epoch},
logger.path(None) / "final-ckp.pth",
logger,
)
logger.log("-" * 200 + "\n")
logger.close()
@@ -210,5 +200,7 @@ if __name__ == "__main__":
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 = "{:}-{:}".format(args.save_dir, args.env_version)
args.save_dir = "{:}-{:}-d{:}".format(
args.save_dir, args.env_version, args.hidden_dim
)
main(args)

220
exps/LFNA/basic-maml.py Normal file
View File

@@ -0,0 +1,220 @@
#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 #
#####################################################
# python exps/LFNA/basic-maml.py --env_version v1 #
# python exps/LFNA/basic-maml.py --env_version v2 #
#####################################################
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
from models.xcore import get_model
from xlayers import super_core
from lfna_utils import lfna_setup, TimeData
class MAML:
"""A LFNA meta-model that uses the MLP as delta-net."""
def __init__(self, container, criterion, meta_lr, inner_lr=0.01, inner_step=1):
self.criterion = criterion
self.container = container
self.meta_optimizer = torch.optim.Adam(
self.container.parameters(), lr=meta_lr, amsgrad=True
)
self.inner_lr = inner_lr
self.inner_step = inner_step
def adapt(self, model, dataset):
# create a container for the future timestamp
y_hat = model.forward_with_container(dataset.x, self.container)
loss = self.criterion(y_hat, dataset.y)
grads = torch.autograd.grad(loss, self.container.parameters())
fast_container = self.container.additive(
[-self.inner_lr * grad for grad in grads]
)
import pdb
pdb.set_trace()
w_container.requires_grad_(True)
containers = [w_container]
for idx, dataset in enumerate(seq_datasets):
x, y = dataset.x, dataset.y
y_hat = model.forward_with_container(x, containers[-1])
loss = criterion(y_hat, y)
gradients = torch.autograd.grad(loss, containers[-1].tensors)
with torch.no_grad():
flatten_w = containers[-1].flatten().view(-1, 1)
flatten_g = containers[-1].flatten(gradients).view(-1, 1)
input_statistics = torch.tensor([x.mean(), x.std()]).view(1, 2)
input_statistics = input_statistics.expand(flatten_w.numel(), -1)
delta_inputs = torch.cat((flatten_w, flatten_g, input_statistics), dim=-1)
delta = self.delta_net(delta_inputs).view(-1)
delta = torch.clamp(delta, -0.5, 0.5)
unflatten_delta = containers[-1].unflatten(delta)
future_container = containers[-1].no_grad_clone().additive(unflatten_delta)
# future_container = containers[-1].additive(unflatten_delta)
containers.append(future_container)
# containers = containers[1:]
meta_loss = []
temp_containers = []
for idx, dataset in enumerate(seq_datasets):
if idx == 0:
continue
current_container = containers[idx]
y_hat = model.forward_with_container(dataset.x, current_container)
loss = criterion(y_hat, dataset.y)
meta_loss.append(loss)
temp_containers.append((dataset.timestamp, current_container, -loss.item()))
meta_loss = sum(meta_loss)
w_container.requires_grad_(False)
# meta_loss.backward()
# self.meta_optimizer.step()
return meta_loss, temp_containers
def step(self):
torch.nn.utils.clip_grad_norm_(self.delta_net.parameters(), 1.0)
self.meta_optimizer.step()
def zero_grad(self):
self.meta_optimizer.zero_grad()
def main(args):
logger, env_info = lfna_setup(args)
total_time = env_info["total"]
for i in range(total_time):
for xkey in ("timestamp", "x", "y"):
nkey = "{:}-{:}".format(i, xkey)
assert nkey in env_info, "{:} no in {:}".format(nkey, list(env_info.keys()))
train_time_bar = total_time // 2
base_model = get_model(
dict(model_type="simple_mlp"),
act_cls="leaky_relu",
norm_cls="identity",
input_dim=1,
output_dim=1,
)
w_container = base_model.get_w_container()
criterion = torch.nn.MSELoss()
print("There are {:} weights.".format(w_container.numel()))
maml = MAML(w_container, criterion, args.meta_lr, args.inner_lr, args.inner_step)
# meta-training
per_epoch_time, start_time = AverageMeter(), time.time()
for iepoch in range(args.epochs):
need_time = "Time Left: {:}".format(
convert_secs2time(per_epoch_time.avg * (args.epochs - iepoch), True)
)
logger.log(
"[{:}] [{:04d}/{:04d}] ".format(time_string(), iepoch, args.epochs)
+ need_time
)
maml.zero_grad()
all_meta_losses = []
for ibatch in range(args.meta_batch):
sampled_timestamp = random.randint(0, train_time_bar)
past_dataset = TimeData(
sampled_timestamp,
env_info["{:}-x".format(sampled_timestamp)],
env_info["{:}-y".format(sampled_timestamp)],
)
future_dataset = TimeData(
sampled_timestamp + 1,
env_info["{:}-x".format(sampled_timestamp + 1)],
env_info["{:}-y".format(sampled_timestamp + 1)],
)
maml.adapt(base_model, past_dataset)
import pdb
pdb.set_trace()
meta_loss = torch.stack(all_meta_losses).mean()
meta_loss.backward()
adaptor.step()
debug_str = pool.debug_info(debug_timestamp)
logger.log("meta-loss: {:.4f}".format(meta_loss.item()))
per_epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log("-" * 200 + "\n")
logger.close()
if __name__ == "__main__":
parser = argparse.ArgumentParser("Use the data in the past.")
parser.add_argument(
"--save_dir",
type=str,
default="./outputs/lfna-synthetic/maml",
help="The checkpoint directory.",
)
parser.add_argument(
"--env_version",
type=str,
required=True,
help="The synthetic enviornment version.",
)
parser.add_argument(
"--meta_lr",
type=float,
default=0.01,
help="The learning rate for the MAML optimizer (default is Adam)",
)
parser.add_argument(
"--inner_lr",
type=float,
default=0.01,
help="The learning rate for the inner optimization",
)
parser.add_argument(
"--inner_step", type=int, default=1, help="The inner loop steps for MAML."
)
parser.add_argument(
"--meta_batch",
type=int,
default=5,
help="The batch size for the meta-model",
)
parser.add_argument(
"--epochs",
type=int,
default=1000,
help="The total number of epochs.",
)
parser.add_argument(
"--workers",
type=int,
default=4,
help="The number of data loading workers (default: 4)",
)
# 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"
main(args)

49
exps/LFNA/basic-same.py
View File

@@ -1,7 +1,8 @@
#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 #
#####################################################
# python exps/LFNA/basic-same.py --srange 1-999
# python exps/LFNA/basic-same.py --srange 1-999 --env_version v1 --hidden_dim 16
# python exps/LFNA/basic-same.py --srange 1-999 --env_version v2 --hidden_dim
#####################################################
import sys, time, copy, torch, random, argparse
from tqdm import tqdm
@@ -22,6 +23,8 @@ from procedures.metric_utils import SaveMetric, MSEMetric, ComposeMetric
from datasets.synthetic_core import get_synthetic_env
from models.xcore import get_model
from lfna_utils import lfna_setup
def subsample(historical_x, historical_y, maxn=10000):
total = historical_x.size(0)
@@ -33,22 +36,7 @@ def subsample(historical_x, historical_y, maxn=10000):
def main(args):
prepare_seed(args.rand_seed)
logger = prepare_logger(args)
cache_path = (logger.path(None) / ".." / "env-info.pth").resolve()
if cache_path.exists():
env_info = torch.load(cache_path)
else:
env_info = dict()
dynamic_env = get_synthetic_env()
env_info["total"] = len(dynamic_env)
for idx, (timestamp, (_allx, _ally)) in enumerate(tqdm(dynamic_env)):
env_info["{:}-timestamp".format(idx)] = timestamp
env_info["{:}-x".format(idx)] = _allx
env_info["{:}-y".format(idx)] = _ally
env_info["dynamic_env"] = dynamic_env
torch.save(env_info, cache_path)
logger, env_info, model_kwargs = lfna_setup(args)
# check indexes to be evaluated
to_evaluate_indexes = split_str2indexes(args.srange, env_info["total"], None)
@@ -78,16 +66,6 @@ def main(args):
historical_x = env_info["{:}-x".format(idx)]
historical_y = env_info["{:}-y".format(idx)]
# build model
mean, std = historical_x.mean().item(), historical_x.std().item()
model_kwargs = dict(
input_dim=1,
output_dim=1,
act_cls="leaky_relu",
norm_cls="identity",
# norm_cls="simple_norm",
# mean=mean,
# std=std,
)
model = get_model(dict(model_type="simple_mlp"), **model_kwargs)
# build optimizer
optimizer = torch.optim.Adam(model.parameters(), lr=args.init_lr, amsgrad=True)
@@ -151,9 +129,9 @@ def main(args):
logger,
)
logger.log("")
per_timestamp_time.update(time.time() - start_time)
start_time = time.time()
save_checkpoint(
{"w_container_per_epoch": w_container_per_epoch},
logger.path(None) / "final-ckp.pth",
@@ -172,6 +150,18 @@ if __name__ == "__main__":
default="./outputs/lfna-synthetic/use-same-timestamp",
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,
required=True,
help="The hidden dimension.",
)
parser.add_argument(
"--init_lr",
type=float,
@@ -205,4 +195,7 @@ if __name__ == "__main__":
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
)
main(args)

272
exps/LFNA/lfna-v0.py Normal file
View File

@@ -0,0 +1,272 @@
#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 #
#####################################################
# python exps/LFNA/lfna-v0.py
#####################################################
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
from models.xcore import get_model
from xlayers import super_core
class LFNAmlp:
"""A LFNA meta-model that uses the MLP as delta-net."""
def __init__(self, obs_dim, hidden_sizes, act_name):
self.delta_net = super_core.SuperSequential(
super_core.SuperLinear(obs_dim, hidden_sizes[0]),
super_core.super_name2activation[act_name](),
super_core.SuperLinear(hidden_sizes[0], hidden_sizes[1]),
super_core.super_name2activation[act_name](),
super_core.SuperLinear(hidden_sizes[1], 1),
)
self.meta_optimizer = torch.optim.Adam(
self.delta_net.parameters(), lr=0.01, amsgrad=True
)
def adapt(self, model, criterion, w_container, seq_datasets):
w_container.requires_grad_(True)
containers = [w_container]
for idx, dataset in enumerate(seq_datasets):
x, y = dataset.x, dataset.y
y_hat = model.forward_with_container(x, containers[-1])
loss = criterion(y_hat, y)
gradients = torch.autograd.grad(loss, containers[-1].tensors)
with torch.no_grad():
flatten_w = containers[-1].flatten().view(-1, 1)
flatten_g = containers[-1].flatten(gradients).view(-1, 1)
input_statistics = torch.tensor([x.mean(), x.std()]).view(1, 2)
input_statistics = input_statistics.expand(flatten_w.numel(), -1)
delta_inputs = torch.cat((flatten_w, flatten_g, input_statistics), dim=-1)
delta = self.delta_net(delta_inputs).view(-1)
delta = torch.clamp(delta, -0.5, 0.5)
unflatten_delta = containers[-1].unflatten(delta)
future_container = containers[-1].no_grad_clone().additive(unflatten_delta)
# future_container = containers[-1].additive(unflatten_delta)
containers.append(future_container)
# containers = containers[1:]
meta_loss = []
temp_containers = []
for idx, dataset in enumerate(seq_datasets):
if idx == 0:
continue
current_container = containers[idx]
y_hat = model.forward_with_container(dataset.x, current_container)
loss = criterion(y_hat, dataset.y)
meta_loss.append(loss)
temp_containers.append((dataset.timestamp, current_container, -loss.item()))
meta_loss = sum(meta_loss)
w_container.requires_grad_(False)
# meta_loss.backward()
# self.meta_optimizer.step()
return meta_loss, temp_containers
def step(self):
torch.nn.utils.clip_grad_norm_(self.delta_net.parameters(), 1.0)
self.meta_optimizer.step()
def zero_grad(self):
self.meta_optimizer.zero_grad()
self.delta_net.zero_grad()
class TimeData:
def __init__(self, timestamp, xs, ys):
self._timestamp = timestamp
self._xs = xs
self._ys = ys
@property
def x(self):
return self._xs
@property
def y(self):
return self._ys
@property
def timestamp(self):
return self._timestamp
class Population:
"""A population used to maintain models at different timestamps."""
def __init__(self):
self._time2model = dict()
self._time2score = dict() # higher is better
def append(self, timestamp, model, score):
if timestamp in self._time2model:
if self._time2score[timestamp] > score:
return
self._time2model[timestamp] = model.no_grad_clone()
self._time2score[timestamp] = score
def query(self, timestamp):
closet_timestamp = None
for xtime, model in self._time2model.items():
if closet_timestamp is None or (
xtime < timestamp and timestamp - closet_timestamp >= timestamp - xtime
):
closet_timestamp = xtime
return self._time2model[closet_timestamp], closet_timestamp
def debug_info(self, timestamps):
xstrs = []
for timestamp in timestamps:
if timestamp in self._time2score:
xstrs.append(
"{:04d}: {:.4f}".format(timestamp, self._time2score[timestamp])
)
return ", ".join(xstrs)
def main(args):
prepare_seed(args.rand_seed)
logger = prepare_logger(args)
cache_path = (logger.path(None) / ".." / "env-info.pth").resolve()
if cache_path.exists():
env_info = torch.load(cache_path)
else:
env_info = dict()
dynamic_env = get_synthetic_env()
env_info["total"] = len(dynamic_env)
for idx, (timestamp, (_allx, _ally)) in enumerate(tqdm(dynamic_env)):
env_info["{:}-timestamp".format(idx)] = timestamp
env_info["{:}-x".format(idx)] = _allx
env_info["{:}-y".format(idx)] = _ally
env_info["dynamic_env"] = dynamic_env
torch.save(env_info, cache_path)
total_time = env_info["total"]
for i in range(total_time):
for xkey in ("timestamp", "x", "y"):
nkey = "{:}-{:}".format(i, xkey)
assert nkey in env_info, "{:} no in {:}".format(nkey, list(env_info.keys()))
train_time_bar = total_time // 2
base_model = get_model(
dict(model_type="simple_mlp"),
act_cls="leaky_relu",
norm_cls="identity",
input_dim=1,
output_dim=1,
)
w_container = base_model.get_w_container()
criterion = torch.nn.MSELoss()
print("There are {:} weights.".format(w_container.numel()))
adaptor = LFNAmlp(4, (50, 20), "leaky_relu")
pool = Population()
pool.append(0, w_container, -100)
# LFNA meta-training
per_epoch_time, start_time = AverageMeter(), time.time()
for iepoch in range(args.epochs):
need_time = "Time Left: {:}".format(
convert_secs2time(per_epoch_time.avg * (args.epochs - iepoch), True)
)
logger.log(
"[{:}] [{:04d}/{:04d}] ".format(time_string(), iepoch, args.epochs)
+ need_time
)
adaptor.zero_grad()
debug_timestamp = set()
all_meta_losses = []
for ibatch in range(args.meta_batch):
sampled_timestamp = random.randint(0, train_time_bar)
query_w_container, query_timestamp = pool.query(sampled_timestamp)
# def adapt(self, model, w_container, xs, ys):
seq_datasets = []
# xs, ys = [], []
for it in range(sampled_timestamp, sampled_timestamp + args.max_seq):
xs = env_info["{:}-x".format(it)]
ys = env_info["{:}-y".format(it)]
seq_datasets.append(TimeData(it, xs, ys))
temp_meta_loss, temp_containers = adaptor.adapt(
base_model, criterion, query_w_container, seq_datasets
)
all_meta_losses.append(temp_meta_loss)
for temp_time, temp_container, temp_score in temp_containers:
pool.append(temp_time, temp_container, temp_score)
debug_timestamp.add(temp_time)
meta_loss = torch.stack(all_meta_losses).mean()
meta_loss.backward()
adaptor.step()
debug_str = pool.debug_info(debug_timestamp)
logger.log("meta-loss: {:.4f}".format(meta_loss.item()))
per_epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log("-" * 200 + "\n")
logger.close()
if __name__ == "__main__":
parser = argparse.ArgumentParser("Use the data in the past.")
parser.add_argument(
"--save_dir",
type=str,
default="./outputs/lfna-synthetic/lfna-v1",
help="The checkpoint directory.",
)
parser.add_argument(
"--init_lr",
type=float,
default=0.1,
help="The initial learning rate for the optimizer (default is Adam)",
)
parser.add_argument(
"--meta_batch",
type=int,
default=5,
help="The batch size for the meta-model",
)
parser.add_argument(
"--epochs",
type=int,
default=1000,
help="The total number of epochs.",
)
parser.add_argument(
"--max_seq",
type=int,
default=5,
help="The maximum length of the sequence.",
)
parser.add_argument(
"--workers",
type=int,
default=4,
help="The number of data loading workers (default: 4)",
)
# 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"
main(args)

61
exps/LFNA/lfna_utils.py Normal file
View File

@@ -0,0 +1,61 @@
#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 #
#####################################################
import torch
from tqdm import tqdm
from procedures import prepare_seed, prepare_logger
from datasets.synthetic_core import get_synthetic_env
def lfna_setup(args):
prepare_seed(args.rand_seed)
logger = prepare_logger(args)
cache_path = (
logger.path(None) / ".." / "env-{:}-info.pth".format(args.env_version)
).resolve()
if cache_path.exists():
env_info = torch.load(cache_path)
else:
env_info = dict()
dynamic_env = get_synthetic_env(version=args.env_version)
env_info["total"] = len(dynamic_env)
for idx, (timestamp, (_allx, _ally)) in enumerate(tqdm(dynamic_env)):
env_info["{:}-timestamp".format(idx)] = timestamp
env_info["{:}-x".format(idx)] = _allx
env_info["{:}-y".format(idx)] = _ally
env_info["dynamic_env"] = dynamic_env
torch.save(env_info, cache_path)
model_kwargs = dict(
input_dim=1,
output_dim=1,
hidden_dim=args.hidden_dim,
act_cls="leaky_relu",
norm_cls="identity",
)
return logger, env_info, model_kwargs
class TimeData:
def __init__(self, timestamp, xs, ys):
self._timestamp = timestamp
self._xs = xs
self._ys = ys
@property
def x(self):
return self._xs
@property
def y(self):
return self._ys
@property
def timestamp(self):
return self._timestamp
def __repr__(self):
return "{name}(timestamp={:}, with {num} samples)".format(
name=self.__class__.__name__, timestamp=self._timestamp, num=len(self._xs)
)