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

Update xmisc

Browse Source
This commit is contained in:
D-X-Y 2021-06-10 23:42:00 -07:00
parent 98f981dd45
commit 248686820c
8 changed files with 72 additions and 487 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
configs/yaml.data/cifar10.test
View File

@ -17,6 +17,6 @@ kwargs:
module_path: torchvision.transforms
args: []
kwargs:
mean: (0.491, 0.482, 0.447)
std: (0.247, 0.244, 0.262)
mean: [0.491, 0.482, 0.447]
std: [0.247, 0.244, 0.262]
kwargs: {}

4
configs/yaml.data/cifar10.train
View File

@ -25,6 +25,6 @@ kwargs:
module_path: torchvision.transforms
args: []
kwargs:
mean: (0.491, 0.482, 0.447)
std: (0.247, 0.244, 0.262)
mean: [0.491, 0.482, 0.447]
std: [0.247, 0.244, 0.262]
kwargs: {}

160
exps/basic/xmain.py
View File

@ -58,6 +58,7 @@ def main(args):
pin_memory=True,
drop_last=False,
)
iters_per_epoch = len(train_data) // args.batch_size
logger.log("The training loader: {:}".format(train_loader))
logger.log("The validation loader: {:}".format(valid_loader))
@ -67,159 +68,44 @@ def main(args):
lr=args.lr,
weight_decay=args.weight_decay,
)
loss = xmisc.nested_call_by_yaml(args.loss_config)
objective = xmisc.nested_call_by_yaml(args.loss_config)
logger.log("The optimizer is:\n{:}".format(optimizer))
logger.log("The loss is {:}".format(loss))
logger.log("The objective is {:}".format(objective))
logger.log("The iters_per_epoch={:}".format(iters_per_epoch))
model, loss = torch.nn.DataParallel(model).cuda(), loss.cuda()
model, objective = torch.nn.DataParallel(model).cuda(), objective.cuda()
scheduler = xmisc.LRMultiplier(
optimizer, xmisc.get_scheduler(args.scheduler, args.lr), args.steps
)
import pdb
pdb.set_trace()
train_func, valid_func = get_procedures(args.procedure)
total_epoch = optim_config.epochs + optim_config.warmup
# Main Training and Evaluation Loop
start_time = time.time()
epoch_time = AverageMeter()
for epoch in range(start_epoch, total_epoch):
scheduler.update(epoch, 0.0)
start_time, iter_time = time.time(), xmisc.AverageMeter()
for xiter, data in enumerate(train_loader):
need_time = "Time Left: {:}".format(
convert_secs2time(epoch_time.avg * (total_epoch - epoch), True)
)
epoch_str = "epoch={:03d}/{:03d}".format(epoch, total_epoch)
LRs = scheduler.get_lr()
find_best = False
# set-up drop-out ratio
if hasattr(base_model, "update_drop_path"):
base_model.update_drop_path(
model_config.drop_path_prob * epoch / total_epoch
)
logger.log(
"\n***{:s}*** start {:s} {:s}, LR=[{:.6f} ~ {:.6f}], scheduler={:}".format(
time_string(), epoch_str, need_time, min(LRs), max(LRs), scheduler
xmisc.time_utils.convert_secs2time(
iter_time.avg * (len(train_loader) - xiter), True
)
)
iter_str = "{:6d}/{:06d}".format(xiter, len(train_loader))
# train for one epoch
train_loss, train_acc1, train_acc5 = train_func(
train_loader,
network,
criterion,
scheduler,
optimizer,
optim_config,
epoch_str,
args.print_freq,
logger,
)
# log the results
logger.log(
"***{:s}*** TRAIN [{:}] loss = {:.6f}, accuracy-1 = {:.2f}, accuracy-5 = {:.2f}".format(
time_string(), epoch_str, train_loss, train_acc1, train_acc5
)
)
inputs, targets = data
targets = targets.cuda(non_blocking=True)
model.train()
# evaluate the performance
if (epoch % args.eval_frequency == 0) or (epoch + 1 == total_epoch):
logger.log("-" * 150)
valid_loss, valid_acc1, valid_acc5 = valid_func(
valid_loader,
network,
criterion,
optim_config,
epoch_str,
args.print_freq_eval,
logger,
)
valid_accuracies[epoch] = valid_acc1
logger.log(
"***{:s}*** VALID [{:}] loss = {:.6f}, accuracy@1 = {:.2f}, accuracy@5 = {:.2f} | Best-Valid-Acc@1={:.2f}, Error@1={:.2f}".format(
time_string(),
epoch_str,
valid_loss,
valid_acc1,
valid_acc5,
valid_accuracies["best"],
100 - valid_accuracies["best"],
)
)
if valid_acc1 > valid_accuracies["best"]:
valid_accuracies["best"] = valid_acc1
find_best = True
logger.log(
"Currently, the best validation accuracy found at {:03d}-epoch :: acc@1={:.2f}, acc@5={:.2f}, error@1={:.2f}, error@5={:.2f}, save into {:}.".format(
epoch,
valid_acc1,
valid_acc5,
100 - valid_acc1,
100 - valid_acc5,
model_best_path,
)
)
num_bytes = (
torch.cuda.max_memory_cached(next(network.parameters()).device) * 1.0
)
logger.log(
"[GPU-Memory-Usage on {:} is {:} bytes, {:.2f} KB, {:.2f} MB, {:.2f} GB.]".format(
next(network.parameters()).device,
int(num_bytes),
num_bytes / 1e3,
num_bytes / 1e6,
num_bytes / 1e9,
)
)
max_bytes[epoch] = num_bytes
if epoch % 10 == 0:
torch.cuda.empty_cache()
optimizer.zero_grad()
outputs = model(inputs)
loss = objective(outputs, targets)
# save checkpoint
save_path = save_checkpoint(
{
"epoch": epoch,
"args": deepcopy(args),
"max_bytes": deepcopy(max_bytes),
"FLOP": flop,
"PARAM": param,
"valid_accuracies": deepcopy(valid_accuracies),
"model-config": model_config._asdict(),
"optim-config": optim_config._asdict(),
"base-model": base_model.state_dict(),
"scheduler": scheduler.state_dict(),
"optimizer": optimizer.state_dict(),
},
model_base_path,
logger,
)
if find_best:
copy_checkpoint(model_base_path, model_best_path, logger)
last_info = save_checkpoint(
{
"epoch": epoch,
"args": deepcopy(args),
"last_checkpoint": save_path,
},
logger.path("info"),
logger,
)
loss.backward()
optimizer.step()
scheduler.step()
if xiter % iters_per_epoch == 0:
logger.log("TRAIN [{:}] loss = {:.6f}".format(iter_str, loss.item()))
# measure elapsed time
epoch_time.update(time.time() - start_time)
iter_time.update(time.time() - start_time)
start_time = time.time()
logger.log("\n" + "-" * 200)
logger.log(
"Finish training/validation in {:} with Max-GPU-Memory of {:.2f} MB, and save final checkpoint into {:}".format(
convert_secs2time(epoch_time.sum, True),
max(v for k, v in max_bytes.items()) / 1e6,
logger.path("info"),
)
)
logger.log("-" * 200 + "\n")
logger.close()
@ -249,7 +135,7 @@ if __name__ == "__main__":
parser.add_argument("--weight_decay", type=float, help="The weight decay")
parser.add_argument("--scheduler", type=str, help="The scheduler indicator.")
parser.add_argument("--steps", type=int, help="The total number of steps.")
parser.add_argument("--batch_size", type=int, default=2, help="The batch size.")
parser.add_argument("--batch_size", type=int, default=256, help="The batch size.")
parser.add_argument("--workers", type=int, default=4, help="The number of workers")
# Random Seed
parser.add_argument("--rand_seed", type=int, default=-1, help="manual seed")

1
scripts/experimental/train-vit.sh
View File

@ -28,4 +28,5 @@ python ./exps/basic/xmain.py --save_dir ${save_dir} --rand_seed ${rseed} \
--model_config ./configs/yaml.model/vit-cifar10.s0 \
--optim_config ./configs/yaml.opt/vit.cifar \
--loss_config ./configs/yaml.loss/cross-entropy \
--batch_size 256 \
--lr 0.003 --weight_decay 0.3 --scheduler warm-cos --steps 10000

41
xautodl/xlayers/super_linear.py
View File

@ -201,7 +201,6 @@ class SuperMLPv2(SuperModule):
self._hidden_multiplier = hidden_multiplier
self._out_features = out_features
self._drop_rate = drop
self._params = nn.ParameterDict({})
self._create_linear(
"fc1", self.in_features, int(self.in_features * self.hidden_multiplier)
@ -226,26 +225,22 @@ class SuperMLPv2(SuperModule):
return spaces.get_max(self._out_features)
def _create_linear(self, name, inC, outC):
self._params["{:}_super_weight".format(name)] = torch.nn.Parameter(
torch.Tensor(outC, inC)
self.register_parameter(
"{:}_super_weight".format(name), torch.nn.Parameter(torch.Tensor(outC, inC))
)
self._params["{:}_super_bias".format(name)] = torch.nn.Parameter(
torch.Tensor(outC)
self.register_parameter(
"{:}_super_bias".format(name), torch.nn.Parameter(torch.Tensor(outC))
)
def reset_parameters(self) -> None:
nn.init.kaiming_uniform_(self._params["fc1_super_weight"], a=math.sqrt(5))
nn.init.kaiming_uniform_(self._params["fc2_super_weight"], a=math.sqrt(5))
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(
self._params["fc1_super_weight"]
)
nn.init.kaiming_uniform_(self.fc1_super_weight, a=math.sqrt(5))
nn.init.kaiming_uniform_(self.fc2_super_weight, a=math.sqrt(5))
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.fc1_super_weight)
bound = 1 / math.sqrt(fan_in)
nn.init.uniform_(self._params["fc1_super_bias"], -bound, bound)
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(
self._params["fc2_super_weight"]
)
nn.init.uniform_(self.fc1_super_bias, -bound, bound)
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.fc2_super_weight)
bound = 1 / math.sqrt(fan_in)
nn.init.uniform_(self._params["fc2_super_bias"], -bound, bound)
nn.init.uniform_(self.fc2_super_bias, -bound, bound)
@property
def abstract_search_space(self):
@ -282,8 +277,8 @@ class SuperMLPv2(SuperModule):
else:
hmul = spaces.get_determined_value(self._hidden_multiplier)
hidden_dim = int(expected_input_dim * hmul)
_fc1_weight = self._params["fc1_super_weight"][:hidden_dim, :expected_input_dim]
_fc1_bias = self._params["fc1_super_bias"][:hidden_dim]
_fc1_weight = self.fc1_super_weight[:hidden_dim, :expected_input_dim]
_fc1_bias = self.fc1_super_bias[:hidden_dim]
x = F.linear(input, _fc1_weight, _fc1_bias)
x = self.act(x)
x = self.drop(x)
@ -292,21 +287,17 @@ class SuperMLPv2(SuperModule):
out_dim = self.abstract_child["_out_features"].value
else:
out_dim = spaces.get_determined_value(self._out_features)
_fc2_weight = self._params["fc2_super_weight"][:out_dim, :hidden_dim]
_fc2_bias = self._params["fc2_super_bias"][:out_dim]
_fc2_weight = self.fc2_super_weight[:out_dim, :hidden_dim]
_fc2_bias = self.fc2_super_bias[:out_dim]
x = F.linear(x, _fc2_weight, _fc2_bias)
x = self.drop(x)
return x
def forward_raw(self, input: torch.Tensor) -> torch.Tensor:
x = F.linear(
input, self._params["fc1_super_weight"], self._params["fc1_super_bias"]
)
x = F.linear(input, self.fc1_super_weight, self.fc1_super_bias)
x = self.act(x)
x = self.drop(x)
x = F.linear(
x, self._params["fc2_super_weight"], self._params["fc2_super_bias"]
)
x = F.linear(x, self.fc2_super_weight, self.fc2_super_bias)
x = self.drop(x)
return x

319
xautodl/xlayers/super_mlp.py
View File

@ -1,319 +0,0 @@
#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.03 #
#####################################################
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from typing import Optional, Callable
from xautodl import spaces
from .super_module import SuperModule
from .super_module import IntSpaceType
from .super_module import BoolSpaceType
class SuperLinear(SuperModule):
"""Applies a linear transformation to the incoming data: :math:`y = xA^T + b`"""
def __init__(
self,
in_features: IntSpaceType,
out_features: IntSpaceType,
bias: BoolSpaceType = True,
) -> None:
super(SuperLinear, self).__init__()
# the raw input args
self._in_features = in_features
self._out_features = out_features
self._bias = bias
# weights to be optimized
self.register_parameter(
"_super_weight",
torch.nn.Parameter(torch.Tensor(self.out_features, self.in_features)),
)
if self.bias:
self.register_parameter(
"_super_bias", torch.nn.Parameter(torch.Tensor(self.out_features))
)
else:
self.register_parameter("_super_bias", None)
self.reset_parameters()
@property
def in_features(self):
return spaces.get_max(self._in_features)
@property
def out_features(self):
return spaces.get_max(self._out_features)
@property
def bias(self):
return spaces.has_categorical(self._bias, True)
@property
def abstract_search_space(self):
root_node = spaces.VirtualNode(id(self))
if not spaces.is_determined(self._in_features):
root_node.append(
"_in_features", self._in_features.abstract(reuse_last=True)
)
if not spaces.is_determined(self._out_features):
root_node.append(
"_out_features", self._out_features.abstract(reuse_last=True)
)
if not spaces.is_determined(self._bias):
root_node.append("_bias", self._bias.abstract(reuse_last=True))
return root_node
def reset_parameters(self) -> None:
nn.init.kaiming_uniform_(self._super_weight, a=math.sqrt(5))
if self.bias:
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self._super_weight)
bound = 1 / math.sqrt(fan_in)
nn.init.uniform_(self._super_bias, -bound, bound)
def forward_candidate(self, input: torch.Tensor) -> torch.Tensor:
# check inputs ->
if not spaces.is_determined(self._in_features):
expected_input_dim = self.abstract_child["_in_features"].value
else:
expected_input_dim = spaces.get_determined_value(self._in_features)
if input.size(-1) != expected_input_dim:
raise ValueError(
"Expect the input dim of {:} instead of {:}".format(
expected_input_dim, input.size(-1)
)
)
# create the weight matrix
if not spaces.is_determined(self._out_features):
out_dim = self.abstract_child["_out_features"].value
else:
out_dim = spaces.get_determined_value(self._out_features)
candidate_weight = self._super_weight[:out_dim, :expected_input_dim]
# create the bias matrix
if not spaces.is_determined(self._bias):
if self.abstract_child["_bias"].value:
candidate_bias = self._super_bias[:out_dim]
else:
candidate_bias = None
else:
if spaces.get_determined_value(self._bias):
candidate_bias = self._super_bias[:out_dim]
else:
candidate_bias = None
return F.linear(input, candidate_weight, candidate_bias)
def forward_raw(self, input: torch.Tensor) -> torch.Tensor:
return F.linear(input, self._super_weight, self._super_bias)
def extra_repr(self) -> str:
return "in_features={:}, out_features={:}, bias={:}".format(
self._in_features, self._out_features, self._bias
)
def forward_with_container(self, input, container, prefix=[]):
super_weight_name = ".".join(prefix + ["_super_weight"])
super_weight = container.query(super_weight_name)
super_bias_name = ".".join(prefix + ["_super_bias"])
if container.has(super_bias_name):
super_bias = container.query(super_bias_name)
else:
super_bias = None
return F.linear(input, super_weight, super_bias)
class SuperMLPv1(SuperModule):
"""An MLP layer: FC -> Activation -> Drop -> FC -> Drop."""
def __init__(
self,
in_features: IntSpaceType,
hidden_features: IntSpaceType,
out_features: IntSpaceType,
act_layer: Callable[[], nn.Module] = nn.GELU,
drop: Optional[float] = None,
):
super(SuperMLPv1, self).__init__()
self._in_features = in_features
self._hidden_features = hidden_features
self._out_features = out_features
self._drop_rate = drop
self.fc1 = SuperLinear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = SuperLinear(hidden_features, out_features)
self.drop = nn.Dropout(drop or 0.0)
@property
def abstract_search_space(self):
root_node = spaces.VirtualNode(id(self))
space_fc1 = self.fc1.abstract_search_space
space_fc2 = self.fc2.abstract_search_space
if not spaces.is_determined(space_fc1):
root_node.append("fc1", space_fc1)
if not spaces.is_determined(space_fc2):
root_node.append("fc2", space_fc2)
return root_node
def apply_candidate(self, abstract_child: spaces.VirtualNode):
super(SuperMLPv1, self).apply_candidate(abstract_child)
if "fc1" in abstract_child:
self.fc1.apply_candidate(abstract_child["fc1"])
if "fc2" in abstract_child:
self.fc2.apply_candidate(abstract_child["fc2"])
def forward_candidate(self, input: torch.Tensor) -> torch.Tensor:
return self.forward_raw(input)
def forward_raw(self, input: torch.Tensor) -> torch.Tensor:
x = self.fc1(input)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
def extra_repr(self) -> str:
return "in_features={:}, hidden_features={:}, out_features={:}, drop={:}, fc1 -> act -> drop -> fc2 -> drop,".format(
self._in_features,
self._hidden_features,
self._out_features,
self._drop_rate,
)
class SuperMLPv2(SuperModule):
"""An MLP layer: FC -> Activation -> Drop -> FC -> Drop."""
def __init__(
self,
in_features: IntSpaceType,
hidden_multiplier: IntSpaceType,
out_features: IntSpaceType,
act_layer: Callable[[], nn.Module] = nn.GELU,
drop: Optional[float] = None,
):
super(SuperMLPv2, self).__init__()
self._in_features = in_features
self._hidden_multiplier = hidden_multiplier
self._out_features = out_features
self._drop_rate = drop
self._params = nn.ParameterDict({})
self._create_linear(
"fc1", self.in_features, int(self.in_features * self.hidden_multiplier)
)
self._create_linear(
"fc2", int(self.in_features * self.hidden_multiplier), self.out_features
)
self.act = act_layer()
self.drop = nn.Dropout(drop or 0.0)
self.reset_parameters()
@property
def in_features(self):
return spaces.get_max(self._in_features)
@property
def hidden_multiplier(self):
return spaces.get_max(self._hidden_multiplier)
@property
def out_features(self):
return spaces.get_max(self._out_features)
def _create_linear(self, name, inC, outC):
self._params["{:}_super_weight".format(name)] = torch.nn.Parameter(
torch.Tensor(outC, inC)
)
self._params["{:}_super_bias".format(name)] = torch.nn.Parameter(
torch.Tensor(outC)
)
def reset_parameters(self) -> None:
nn.init.kaiming_uniform_(self._params["fc1_super_weight"], a=math.sqrt(5))
nn.init.kaiming_uniform_(self._params["fc2_super_weight"], a=math.sqrt(5))
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(
self._params["fc1_super_weight"]
)
bound = 1 / math.sqrt(fan_in)
nn.init.uniform_(self._params["fc1_super_bias"], -bound, bound)
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(
self._params["fc2_super_weight"]
)
bound = 1 / math.sqrt(fan_in)
nn.init.uniform_(self._params["fc2_super_bias"], -bound, bound)
@property
def abstract_search_space(self):
root_node = spaces.VirtualNode(id(self))
if not spaces.is_determined(self._in_features):
root_node.append(
"_in_features", self._in_features.abstract(reuse_last=True)
)
if not spaces.is_determined(self._hidden_multiplier):
root_node.append(
"_hidden_multiplier", self._hidden_multiplier.abstract(reuse_last=True)
)
if not spaces.is_determined(self._out_features):
root_node.append(
"_out_features", self._out_features.abstract(reuse_last=True)
)
return root_node
def forward_candidate(self, input: torch.Tensor) -> torch.Tensor:
# check inputs ->
if not spaces.is_determined(self._in_features):
expected_input_dim = self.abstract_child["_in_features"].value
else:
expected_input_dim = spaces.get_determined_value(self._in_features)
if input.size(-1) != expected_input_dim:
raise ValueError(
"Expect the input dim of {:} instead of {:}".format(
expected_input_dim, input.size(-1)
)
)
# create the weight and bias matrix for fc1
if not spaces.is_determined(self._hidden_multiplier):
hmul = self.abstract_child["_hidden_multiplier"].value * expected_input_dim
else:
hmul = spaces.get_determined_value(self._hidden_multiplier)
hidden_dim = int(expected_input_dim * hmul)
_fc1_weight = self._params["fc1_super_weight"][:hidden_dim, :expected_input_dim]
_fc1_bias = self._params["fc1_super_bias"][:hidden_dim]
x = F.linear(input, _fc1_weight, _fc1_bias)
x = self.act(x)
x = self.drop(x)
# create the weight and bias matrix for fc2
if not spaces.is_determined(self._out_features):
out_dim = self.abstract_child["_out_features"].value
else:
out_dim = spaces.get_determined_value(self._out_features)
_fc2_weight = self._params["fc2_super_weight"][:out_dim, :hidden_dim]
_fc2_bias = self._params["fc2_super_bias"][:out_dim]
x = F.linear(x, _fc2_weight, _fc2_bias)
x = self.drop(x)
return x
def forward_raw(self, input: torch.Tensor) -> torch.Tensor:
x = F.linear(
input, self._params["fc1_super_weight"], self._params["fc1_super_bias"]
)
x = self.act(x)
x = self.drop(x)
x = F.linear(
x, self._params["fc2_super_weight"], self._params["fc2_super_bias"]
)
x = self.drop(x)
return x
def extra_repr(self) -> str:
return "in_features={:}, hidden_multiplier={:}, out_features={:}, drop={:}, fc1 -> act -> drop -> fc2 -> drop,".format(
self._in_features,
self._hidden_multiplier,
self._out_features,
self._drop_rate,
)

8
xautodl/xmisc/__init__.py
View File

@ -1,6 +1,7 @@
#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.06 #
#####################################################
"""The module and yaml related functions."""
from .module_utils import call_by_dict
from .module_utils import call_by_yaml
from .module_utils import nested_call_by_dict
@ -11,10 +12,13 @@ from .torch_utils import count_parameters
from .logger_utils import Logger
# sampler
"""The data sampler related classes."""
from .sampler_utils import BatchSampler
# scheduler related
"""The meter related classes."""
from .meter_utils import AverageMeter
"""The scheduler related classes."""
from .scheduler_utils import CosineParamScheduler, WarmupParamScheduler, LRMultiplier

22
xautodl/xmisc/meter_utils.py Normal file
View File

@ -0,0 +1,22 @@
class AverageMeter:
"""Computes and stores the average and current value"""
def __init__(self):
self.reset()
def reset(self):
self.val = 0.0
self.avg = 0.0
self.sum = 0.0
self.count = 0.0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
def __repr__(self):
return "{name}(val={val}, avg={avg}, count={count})".format(
name=self.__class__.__name__, **self.__dict__
)