0.1.1beta
Pre-releaseExCore
ExCore
is a Configuration/Registry System designed for deeplearning, with some utils.
✨ ExCore
supports auto-completion, type-hinting, docstring and code navigation for config files
ExCore
is still in an early development stage.
English | 中文
Features
Config System
Config system is the core of deeplearning projects which enable us to manage and adjust hyperparameters and expriments. There are some attempts of config system because the whole community has been suffering from the plain text config files for a long while.
Config System in ExCore
is specifically designed for deeplearning training (generally refers to all similar part, e.g. testing, evaluating) procedure. The core premise is to categorize the objects to be created in the config into three classes - Primary
, Intermediate
, and Isolated
objects
Primary
objects are those which are directly used in training, e.g. model, optimizer.ExCore
will instantiate and return them.Intermediate
objects are those which are indirectly used in training, e.g. backbone of the model, parameters of model that will pass to optimizer.ExCore
will instantiate them, and pass them to targetPrimary
objects as arguments according some rules.Isolated
objects refer to python built-in objects which will be parsed when loading toml, e.g. int, string, list and dict.
ExCore
extends the syntax of toml file, introducing some special prefix characters -- !
, @
, $
and '&' to simplify the config defination.
The config system has following features.
Get rid of `type`
Model:
type: ResNet # <----- ugly type
layers: 50
num_classes: 1
In order to get rid of type
, ExCore
regards all registered names as reserved words
. The Primary
module need to be defined like [PrimaryFields.ModuleName]
. PrimaryFields
are some pre-defined fields, e.g. Model
, Optimizer
. ModuleName
are registered names.
[Model.FCN]
layers = 50
num_classes = 1
Eliminate modules nesting
TrainData:
type: Cityscapes
dataset_root: data/cityscapes
transforms:
- type: ResizeStepScale
min_scale_factor: 0.5
max_scale_factor: 2.0
scale_step_size: 0.25
- type: RandomPaddingCrop
crop_size: [1024, 512]
- type: Normalize
mode: train
ExCore
use some special prefix characters to specify certain arguments are modules as well. More prefixes will be introduced later.
[TrainData.Cityscapes]
dataset_root = "data/cityscapes"
mode = 'train'
# use `!` to show this is a module, It's formal to use a quoted key "!transforms", but whatever
!transforms = ["ResizeStepScale", "RandomPaddingCrop", "Normalize"]
# `PrimaryFields` can be omitted in defination of `Intermediate` module
[ResizeStepScale]
min_scale_factor = 0.5
max_scale_factor = 2.0
scale_step_size = 0.25
# or explicitly specify ``PrimaryFields
[Transforms.RandomPaddingCrop]
crop_size = [1024, 512]
# It can even be undefined when there are no arguments
# [Normalize]
✨Auto-complement, type-hinting, docstring and code navigation for config files
The old-style design of plain text configs has been criticized for being difficult to write (without auto-completion) and not allowing navigation to the corresponding class. However, Language Server Protocol can be leveraged to support various code editing features, such as auto-completion, type-hinting, and code navigation. By utilizing lsp and json schema, it's able to provide the ability of auto-completion, some weak type-hinting (If code is well annotated, such as standard type hint in python, it will acheive more) and docstring of corresponding class.
ExCore
dump the mappings of class name and it file location to support code navigation. Currently only support for neovim, see excore.nvim.
Config inheritance
Use `__base__` to inherit from a toml file. Only dict can be updated locally, other types are overwritten directly.__base__ = ["xxx.toml", "xxxx.toml"]
`@`Reused module
ExCore
use @
to mark the reused module, which is shared between different modules.
# FCN and SegNet will use the same ResNet object
[Model.FCN]
@backbone = "ResNet"
[Model.SegNet]
@backbone = "ResNet"
[ResNet]
layers = 50
in_channel = 3
equls to
resnet = ResNet(layers=50, in_channel=3)
FCN(backbone=resnet)
SegNet(backbone=resnet)
# If use `!`, it equls to
FCN(backbone=ResNet(layers=50, in_channel=3))
SegNet(backbone=ResNet(layers=50, in_channel=3))
`$`Refer Class and cross file
ExCore
use $
to represents class itself, which will not be instantiated.
[Model.ResNet]
$block = "BasicBlock"
layers = 50
in_channel = 3
equls to
from xxx import ResNet, BasicBlock
ResNet(block=BasicBlock, layers=50, in_channel=3)
In order to refer module accross files, $
can be used before PrimaryFields
. For example:
File A:
[Block.BasicBlock]
File B:
[Block.BottleneckBlock]
File C:
[Model.ResNet]
!block="$Block"
So we can combine file A and C or file B and C with a toml file
__base__ = ["A.toml", "C.toml"]
# or
__base__ = ["B.toml", "C.toml"]
`&`Variable reference
ExCore
use &
to refer a variable from the top-level of config.
Note: The value may be overwritten when inheriting, so the call it variable.
size = 224
[TrainData.ImageNet]
&train_size = "size"
!transforms = ['RandomResize', 'Pad']
data_path = 'xxx'
[Transform.Pad]
&pad_size = "size"
[TestData.ImageNet]
!transforms = ['Normalize']
&test_size = "size"
data_path = 'xxx'
✨Using python module in config file
The Registry
in ExCore
is able to register a module:
from excore import Registry
import torch
MODULE = Registry("module")
MODULE.register_module(torch)
Then you can use torch in config file:
[Model.ResNet]
$activation = "torch.nn.ReLU"
# or
!activation = "torch.nn.ReLU"
import torch
from xxx import ResNet
ResNet(torch.nn.ReLU)
# or
ResNet(torch.nn.ReLU())
Note: You shouldn't define arguments of a module.
✨Argument-level hook
ExCore
provide a simple way to call argument-level hooks without arguments.
[Optimizer.AdamW]
@params = "$Model.parameters()"
weight_decay = 0.01
If you want to call a class or static method.
[Model.XXX]
$backbone = "A.from_pretained()"
Attributes can also be used.
[Model.XXX]
!channel = "$Block.out_channel"
It also can be chained invoke.
[Model.XXX]
!channel = "$Block.last_conv.out_channels"
This way requsts you to define such methods or attributes in target class and can not pass arguments. So ExCore
provides ConfigArgumentHook
.
class ConfigArgumentHook(node, enabled)
You need to implements your own class inherited from ConfigArgumentHook
. For example:
from excore.engine.hook import ConfigArgumentHook
from . import HOOKS
@HOOKS.register()
class BnWeightDecayHook(ConfigArgumentHook):
def __init__(self, node, enabled: bool, bn_weight_decay: bool, weight_decay: float):
super().__init__(node, enabled)
self.bn_weight_decay = bn_weight_decay
self.weight_decay = weight_decay
def hook(self):
model = self.node()
if self.bn_weight_decay:
optim_params = model.parameters()
else:
p_bn = [p for n, p in model.named_parameters() if "bn" in n]
p_non_bn = [p for n, p in model.named_parameters() if "bn" not in n]
optim_params = [
{"params": p_bn, "weight_decay": 0},
{"params": p_non_bn, "weight_decay": self.weight_decay},
]
return optim_params
[Optimizer.SGD]
@params = "$Model@BnWeightDecayHook"
lr = 0.05
momentum = 0.9
weight_decay = 0.0001
[ConfigHook.BnWeightDecayHook]
weight_decay = 0.0001
bn_weight_decay = false
enabled = true
Use @
to call user defined hooks.
Instance-level hook
If the logic of module building are too complicated, instance-level hook may be helpful.
TODO
✨Lazy Config with simple API
The core conception of LazyConfig is 'Lazy', which represents a status of delay. Before instantiating, all the parameters will be stored in a special dict which additionally contains what the target class is. So It's easy to alter any parameters of the module and control which module should be instantiated and which module should not.It's also used to address the defects of plain text configs through python lsp which is able to provide code navigation, auto-completion and more.
ExCore
implements some nodes - MoudleNode
, InternNode
, ReusedNode
, ClassNode
, ConfigHookNode
, ChainedInvocationWrapper
and VariableReference
and a LazyConfig
to manage all nodes.
ExCore
provides only 2 simple API to build moduels -- 'load' and build_all
.
Typically, we follow the following procedure.
from excore import config
layz_cfg = config.load('xxx.toml')
module_dict, run_info = config.build_all(layz_cfg)
The results of build_all
are respectively Primary
modules and Isolated
objects.
If you only want to use a certain module.
from excore import config
layz_cfg = config.load('xxx.toml')
model = lazy_cfg.Model() # Model is one of `PrimaryFields`
# or
model = layz_cfg['Model']()
If you want to follow other logic to build modules, you can still use LazyConfig
to adjust the arguments of node
s and more things.
from excore import config
layz_cfg = config.load('xxx.toml')
lazy_cfg.Model.add_params(pre_trained='./')
module_dict, run_info = config.build_all(layz_cfg)
Config print
from excore import config
cfg = config.load_config('xx.toml')
print(cfg)
Result:
╒══════════════════════════╤══════════════════════════════════════════════════════════════════════╕
│ size │ 1024 │
├──────────────────────────┼──────────────────────────────────────────────────────────────────────┤
│ TrainData.CityScapes │ ╒═════════════╤════════════════════════════════════════════════════╕ │
│ │ │ &train_size │ size │ │
│ │ ├─────────────┼────────────────────────────────────────────────────┤ │
│ │ │ !transforms │ ['RandomResize', 'RandomFlip', 'Normalize', 'Pad'] │ │
│ │ ├─────────────┼────────────────────────────────────────────────────┤ │
│ │ │ data_path │ xxx │ │
│ │ ╘═════════════╧════════════════════════════════════════════════════╛ │
├──────────────────────────┼──────────────────────────────────────────────────────────────────────┤
│ Transform.RandomFlip │ ╒══════╤═════╕ │
│ │ │ prob │ 0.5 │ │
│ │ ├──────┼─────┤ │
│ │ │ axis │ 0 │ │
│ │ ╘══════╧═════╛ │
├──────────────────────────┼──────────────────────────────────────────────────────────────────────┤
│ Transform.Pad │ ╒═══════════╤══════╕ │
│ │ │ &pad_size │ size │ │
│ │ ╘═══════════╧══════╛ │
├──────────────────────────┼──────────────────────────────────────────────────────────────────────┤
│ Normalize.std │ [0.5, 0.5, 0.5] │
├──────────────────────────┼──────────────────────────────────────────────────────────────────────┤
│ Normalize.mean │ [0.5, 0.5, 0.5] │
├──────────────────────────┼──────────────────────────────────────────────────────────────────────┤
│ TestData.CityScapes │ ╒═════════════╤═══════════════╕ │
│ │ │ !transforms │ ['Normalize'] │ │
│ │ ├─────────────┼───────────────┤ │
│ │ │ &test_size │ size │ │
│ │ ├─────────────┼───────────────┤ │
│ │ │ data_path │ xxx │ │
│ │ ╘═════════════╧═══════════════╛ │
├──────────────────────────┼──────────────────────────────────────────────────────────────────────┤
│ Model.FCN │ ╒═══════════╤════════════╕ │
│ │ │ @backbone │ ResNet │ │
│ │ ├───────────┼────────────┤ │
│ │ │ @head │ SimpleHead │ │
│ │ ╘═══════════╧════════════╛ │
...
Registry
✨LazyRegistry
To reduce the unnecessary imports, `ExCore` provides `LazyRegistry`, which store the mappings of class/function name to its `qualname` and dump the mappings to local. When config parsing, the necessary modules will be imported.Extra information
from excore import Registry
Models = Registry("Model", extra_field="is_backbone")
@Models.register(is_backbone=True)
class ResNet:
pass
Modules classification and fuzzy search
from excore import Registry
Models = Registry("Model", extra_field="is_backbone")
@Models.register(is_backbone=True)
class ResNet:
pass
@Models.register(is_backbone=True)
class ResNet50:
pass
@Models.register(is_backbone=True)
class ResNet101:
pass
@Models.register(is_backbone=False)
class head:
pass
print(Models.module_table(select_info='is_backbone'))
print(Models.module_table(filter='**Res**'))
results:
╒═══════════╤═══════════════╕
│ Model │ is_backbone │
╞═══════════╪═══════════════╡
│ ResNet │ True │
├───────────┼───────────────┤
│ ResNet101 │ True │
├───────────┼───────────────┤
│ ResNet50 │ True │
├───────────┼───────────────┤
│ head │ False │
╘═══════════╧═══════════════╛
╒═══════════╕
│ Model │
╞═══════════╡
│ ResNet │
├───────────┤
│ ResNet101 │
├───────────┤
│ ResNet50 │
╘═══════════╛
Register all
from torch import optim
from excore import Registry
OPTIM = Registry("Optimizer")
def _get_modules(name: str, module) -> bool:
if name[0].isupper():
return True
return False
OPTIM.match(optim, _get_modules)
print(OPTIM)
results:
╒════════════╤════════════════════════════════════╕
│ NAEM │ DIR │
╞════════════╪════════════════════════════════════╡
│ Adadelta │ torch.optim.adadelta.Adadelta │
├────────────┼────────────────────────────────────┤
│ Adagrad │ torch.optim.adagrad.Adagrad │
├────────────┼────────────────────────────────────┤
│ Adam │ torch.optim.adam.Adam │
├────────────┼────────────────────────────────────┤
│ AdamW │ torch.optim.adamw.AdamW │
├────────────┼────────────────────────────────────┤
│ SparseAdam │ torch.optim.sparse_adam.SparseAdam │
├────────────┼────────────────────────────────────┤
│ Adamax │ torch.optim.adamax.Adamax │
├────────────┼────────────────────────────────────┤
│ ASGD │ torch.optim.asgd.ASGD │
├────────────┼────────────────────────────────────┤
│ SGD │ torch.optim.sgd.SGD │
├────────────┼────────────────────────────────────┤
│ RAdam │ torch.optim.radam.RAdam │
├────────────┼────────────────────────────────────┤
│ Rprop │ torch.optim.rprop.Rprop │
├────────────┼────────────────────────────────────┤
│ RMSprop │ torch.optim.rmsprop.RMSprop │
├────────────┼────────────────────────────────────┤
│ Optimizer │ torch.optim.optimizer.Optimizer │
├────────────┼────────────────────────────────────┤
│ NAdam │ torch.optim.nadam.NAdam │
├────────────┼────────────────────────────────────┤
│ LBFGS │ torch.optim.lbfgs.LBFGS │
╘════════════╧════════════════════════════════════╛
All in one
Through Registry to find all registries. Make registries into a global one.
from excore import Registry
MODEL = Registry.get_registry("Model")
G = Registry.make_global()
✨Register module
Registry
is able to not only register class or function, but also a python module, for example:
from excore import Registry
import torch
MODULE = Registry("module")
MODULE.register_module(torch)
Then you can use torch in config file:
[Model.ResNet]
$activation = "torch.nn.ReLU"
# or
!activation = "torch.nn.ReLU"
equls to
import torch
from xxx import ResNet
ResNet(torch.nn.ReLU)
# or
ResNet(torch.nn.ReLU())
RoadMap
For more features you may refer to Roadmap of ExCore
What's Changed
- ✨ Support LazyRegistry by @Asthestarsfalll in #9
- ✨ Support completion for config file by @Asthestarsfalll in #11
- 👷 Add CI by @Asthestarsfalll in #15
- ✨ add workspace config and cli by @Asthestarsfalll in #16
- ✨ Support non-instance value: class and post-hook by @Asthestarsfalll in #18
- 🐛 fix some config/json_schema bugs by @Asthestarsfalll in #19
- ✨ Support Variable Reference by @Asthestarsfalll in #22
- 🎨 Turn to ruff by @Asthestarsfalll in #23
- 🎨 Use black as formatter by @Asthestarsfalll in #24
- ♻️✅ Refactor config parsing, add tests for co… by @Asthestarsfalll in #27
- ✅ Add more tests by @Asthestarsfalll in #28
- ✨ Add simple code navigation by @Asthestarsfalll in #29
- ✅ Add more tests of config by @Asthestarsfalll in #30
- ✨ Support type-hints for modules by @Asthestarsfalll in #32
Full Changelog: https://github.com/Asthestarsfalll/ExCore/commits/0.1.1beta