iicd/xautodl
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update LFNA version 1.0

Browse Source
This commit is contained in:
D-X-Y
2021-05-13 21:33:34 +08:00
parent 3d3a04705f
commit cfabd05de8
11 changed files with 340 additions and 299 deletions
Download Patch File Download Diff File Expand all files Collapse all files

214
exps/LFNA/lfna.py
View File

@@ -1,7 +1,7 @@
#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 #
#####################################################
# python exps/LFNA/lfna.py --env_version v1 --hidden_dim 16 --layer_dim 32 --epochs 50000
# python exps/LFNA/lfna.py --env_version v1
#####################################################
import sys, time, copy, torch, random, argparse
from tqdm import tqdm
@@ -19,56 +19,82 @@ 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 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
from lfna_models_v2 import HyperNet
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 = env_info["dynamic_env"]
model = get_model(**model_kwargs)
model = model.to(args.device)
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()
logger.log("There are {:} weights.".format(model.get_w_container().numel()))
# meta_train_range = (dynamic_env.min_timestamp, (dynamic_env.min_timestamp + dynamic_env.max_timestamp) / 2)
# meta_train_interval = dynamic_env.timestamp_interval
shape_container = model.get_w_container().to_shape_container()
shape_container = base_model.get_w_container().to_shape_container()
# pre-train the hypernetwork
timestamps = list(
dynamic_env.get_timestamp(index) for index in range(len(dynamic_env) // 2)
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,
)
hypernet = HyperNet(shape_container, args.layer_dim, args.task_dim, timestamps)
hypernet = hypernet.to(args.device)
import pdb
pdb.set_trace()
# task_embed = torch.nn.Parameter(torch.Tensor(env_info["total"], args.task_dim))
total_bar = 16
task_embeds = []
for i in range(total_bar):
tensor = torch.Tensor(1, args.task_dim).to(args.device)
task_embeds.append(torch.nn.Parameter(tensor))
for task_embed in task_embeds:
trunc_normal_(task_embed, std=0.02)
model.train()
hypernet.train()
parameters = list(hypernet.parameters()) + task_embeds
# optimizer = torch.optim.Adam(parameters, lr=args.init_lr, amsgrad=True)
optimizer = torch.optim.Adam(parameters, lr=args.init_lr, weight_decay=1e-5)
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=[
@@ -77,71 +103,59 @@ def main(args):
],
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)))
# total_bar = env_info["total"] - 1
# LFNA meta-training
loss_meter = AverageMeter()
per_epoch_time, start_time = AverageMeter(), time.time()
last_success_epoch = 0
for iepoch in range(args.epochs):
need_time = "Time Left: {:}".format(
head_str = "[{:}] [{:04d}/{:04d}] ".format(
time_string(), iepoch, args.epochs
) + "Time Left: {:}".format(
convert_secs2time(per_epoch_time.avg * (args.epochs - iepoch), True)
)
head_str = (
"[{:}] [{:04d}/{:04d}] ".format(time_string(), iepoch, args.epochs)
+ need_time
loss_meter = epoch_train(
env_loader,
meta_model,
base_model,
optimizer,
criterion,
args.device,
logger,
)
losses = []
# for ibatch in range(args.meta_batch):
for cur_time in range(total_bar):
# cur_time = random.randint(0, total_bar)
cur_task_embed = task_embeds[cur_time]
cur_container = hypernet(cur_task_embed)
cur_x = env_info["{:}-x".format(cur_time)].to(args.device)
cur_y = env_info["{:}-y".format(cur_time)].to(args.device)
cur_dataset = TimeData(cur_time, cur_x, cur_y)
preds = model.forward_with_container(cur_dataset.x, cur_container)
optimizer.zero_grad()
loss = criterion(preds, cur_dataset.y)
losses.append(loss)
final_loss = torch.stack(losses).mean()
final_loss.backward()
optimizer.step()
lr_scheduler.step()
loss_meter.update(final_loss.item())
if iepoch % 100 == 0:
logger.log(
head_str
+ " meta-loss: {:.4f} ({:.4f}) :: lr={:.5f}, batch={:}".format(
loss_meter.avg,
loss_meter.val,
min(lr_scheduler.get_last_lr()),
len(losses),
)
)
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(
{
"hypernet": hypernet.state_dict(),
"task_embed": task_embed,
"meta_model": meta_model.state_dict(),
"optimizer": optimizer.state_dict(),
"lr_scheduler": lr_scheduler.state_dict(),
"iepoch": iepoch,
"args": args,
},
logger.path("model"),
logger,
)
loss_meter.reset()
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()
print(model)
print(hypernet)
w_container_per_epoch = dict()
for idx in range(0, total_bar):
future_time = env_info["{:}-timestamp".format(idx)]
@@ -183,20 +197,26 @@ if __name__ == "__main__":
parser.add_argument(
"--hidden_dim",
type=int,
required=True,
default=16,
help="The hidden dimension.",
)
parser.add_argument(
"--layer_dim",
type=int,
required=True,
help="The hidden dimension.",
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=0.1,
default=0.01,
help="The initial learning rate for the optimizer (default is Adam)",
)
parser.add_argument(
@@ -206,10 +226,23 @@ if __name__ == "__main__":
help="The batch size for the meta-model",
)
parser.add_argument(
"--epochs",
"--sampler_enlarge",
type=int,
default=2000,
help="The total number of epochs.",
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",
@@ -223,8 +256,7 @@ 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.task_dim = args.layer_dim
args.save_dir = "{:}-{:}-d{:}".format(
args.save_dir, args.env_version, args.hidden_dim
args.save_dir = "{:}-{:}-d{:}_{:}_{:}".format(
args.save_dir, args.env_version, args.hidden_dim, args.layer_dim, args.time_dim
)
main(args)

128
exps/LFNA/lfna_meta_model.py Normal file
View File

@@ -0,0 +1,128 @@
#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 #
#####################################################
import copy
import torch
import torch.nn.functional as F
from xlayers import super_core
from xlayers import trunc_normal_
from models.xcore import get_model
class LFNA_Meta(super_core.SuperModule):
"""Learning to Forecast Neural Adaptation (Meta Model Design)."""
def __init__(
self,
shape_container,
layer_embeding,
time_embedding,
meta_timestamps,
mha_depth: int = 2,
dropout: float = 0.1,
):
super(LFNA_Meta, self).__init__()
self._shape_container = shape_container
self._num_layers = len(shape_container)
self._numel_per_layer = []
for ilayer in range(self._num_layers):
self._numel_per_layer.append(shape_container[ilayer].numel())
self._raw_meta_timestamps = meta_timestamps
self.register_parameter(
"_super_layer_embed",
torch.nn.Parameter(torch.Tensor(self._num_layers, layer_embeding)),
)
self.register_parameter(
"_super_meta_embed",
torch.nn.Parameter(torch.Tensor(len(meta_timestamps), time_embedding)),
)
self.register_buffer("_meta_timestamps", torch.Tensor(meta_timestamps))
# build transformer
layers = []
for ilayer in range(mha_depth):
layers.append(
super_core.SuperTransformerEncoderLayer(
time_embedding,
4,
True,
4,
dropout,
norm_affine=False,
order=super_core.LayerOrder.PostNorm,
)
)
self.meta_corrector = super_core.SuperSequential(*layers)
model_kwargs = dict(
config=dict(model_type="dual_norm_mlp"),
input_dim=layer_embeding + time_embedding,
output_dim=max(self._numel_per_layer),
hidden_dims=[(layer_embeding + time_embedding) * 2] * 3,
act_cls="gelu",
norm_cls="layer_norm_1d",
dropout=dropout,
)
self._generator = get_model(**model_kwargs)
# print("generator: {:}".format(self._generator))
# unknown token
self.register_parameter(
"_unknown_token",
torch.nn.Parameter(torch.Tensor(1, time_embedding)),
)
# initialization
trunc_normal_(
[self._super_layer_embed, self._super_meta_embed, self._unknown_token],
std=0.02,
)
def forward_raw(self, timestamps):
# timestamps is a batch of sequence of timestamps
batch, seq = timestamps.shape
timestamps = timestamps.unsqueeze(dim=-1)
meta_timestamps = self._meta_timestamps.view(1, 1, -1)
time_diffs = timestamps - meta_timestamps
time_match_v, time_match_i = torch.min(torch.abs(time_diffs), dim=-1)
# select corresponding meta-knowledge
meta_match = torch.index_select(
self._super_meta_embed, dim=0, index=time_match_i.view(-1)
)
meta_match = meta_match.view(batch, seq, -1)
# create the probability
time_probs = (1 / torch.exp(time_match_v * 10)).view(batch, seq, 1)
time_probs[:, -1, :] = 0
unknown_token = self._unknown_token.view(1, 1, -1)
raw_meta_embed = time_probs * meta_match + (1 - time_probs) * unknown_token
meta_embed = self.meta_corrector(raw_meta_embed)
# create joint embed
num_layer, _ = self._super_layer_embed.shape
meta_embed = meta_embed.view(batch, seq, 1, -1).expand(-1, -1, num_layer, -1)
layer_embed = self._super_layer_embed.view(1, 1, num_layer, -1).expand(
batch, seq, -1, -1
)
joint_embed = torch.cat((meta_embed, layer_embed), dim=-1)
batch_weights = self._generator(joint_embed)
batch_containers = []
for seq_weights in torch.split(batch_weights, 1):
seq_containers = []
for weights in torch.split(seq_weights.squeeze(0), 1):
weights = torch.split(weights.squeeze(0), 1)
seq_containers.append(self._shape_container.translate(weights))
batch_containers.append(seq_containers)
return batch_containers
def forward_candidate(self, input):
raise NotImplementedError
def extra_repr(self) -> str:
return "(_super_layer_embed): {:}, (_super_meta_embed): {:}, (_meta_timestamps): {:}".format(
list(self._super_layer_embed.shape),
list(self._super_meta_embed.shape),
list(self._meta_timestamps.shape),
)

72
exps/LFNA/lfna_models_v2.py
View File

@@ -1,72 +0,0 @@
#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 #
#####################################################
import copy
import torch
import torch.nn.functional as F
from xlayers import super_core
from xlayers import trunc_normal_
from models.xcore import get_model
class HyperNet(super_core.SuperModule):
"""The hyper-network."""
def __init__(
self,
shape_container,
layer_embeding,
task_embedding,
meta_timestamps,
return_container: bool = True,
):
super(HyperNet, self).__init__()
self._shape_container = shape_container
self._num_layers = len(shape_container)
self._numel_per_layer = []
for ilayer in range(self._num_layers):
self._numel_per_layer.append(shape_container[ilayer].numel())
self.register_parameter(
"_super_layer_embed",
torch.nn.Parameter(torch.Tensor(self._num_layers, layer_embeding)),
)
trunc_normal_(self._super_layer_embed, std=0.02)
model_kwargs = dict(
config=dict(model_type="dual_norm_mlp"),
input_dim=layer_embeding + task_embedding,
output_dim=max(self._numel_per_layer),
hidden_dims=[(layer_embeding + task_embedding) * 2] * 3,
act_cls="gelu",
norm_cls="layer_norm_1d",
dropout=0.2,
)
import pdb
pdb.set_trace()
self._generator = get_model(**model_kwargs)
self._return_container = return_container
print("generator: {:}".format(self._generator))
def forward_raw(self, task_embed):
# task_embed = F.normalize(task_embed, dim=-1, p=2)
# layer_embed = F.normalize(self._super_layer_embed, dim=-1, p=2)
layer_embed = self._super_layer_embed
task_embed = task_embed.view(1, -1).expand(self._num_layers, -1)
joint_embed = torch.cat((task_embed, layer_embed), dim=-1)
weights = self._generator(joint_embed)
if self._return_container:
weights = torch.split(weights, 1)
return self._shape_container.translate(weights)
else:
return weights
def forward_candidate(self, input):
raise NotImplementedError
def extra_repr(self) -> str:
return "(_super_layer_embed): {:}".format(list(self._super_layer_embed.shape))

4
exps/LFNA/vis-synthetic.py
View File

@@ -225,8 +225,8 @@ def visualize_env(save_dir, version):
def compare_algs(save_dir, version, alg_dir="./outputs/lfna-synthetic"):
save_dir = Path(str(save_dir))
for substr in ("pdf", "png"):
sub_save_dir = save_dir / substr
sub_save_dir.mkdir(parents=True, exist_ok=True)
sub_save_dir = save_dir / substr
sub_save_dir.mkdir(parents=True, exist_ok=True)
dpi, width, height = 30, 3200, 2000
figsize = width / float(dpi), height / float(dpi)