mirror of https://github.com/tracel-ai/burn.git
Doc: initial architecture documentation (#308)
This commit is contained in:
Nathaniel Simard
GitHub
parent
02abc373d3
commit
c9fda41846
|
|
@ -0,0 +1,158 @@
|
|||
# Architecture
|
||||
|
||||
This file documents most major architectural decisions with the reasoning behind them.
|
||||
|
||||
__Sections__
|
||||
|
||||
* [Module](#module)
|
||||
* [Optimization](#optimization)
|
||||
* [Serialization](#serialization)
|
||||
|
||||
|
||||
## Module
|
||||
|
||||
Modules are a way of creating neural network structures that can be easily optimized, saved, and loaded with little to no boilerplate.
|
||||
Unlike other frameworks, a module does not force the declaration of the forward pass, leaving it up to the implementer to decide how it should be defined.
|
||||
Additionally, most modules are created using a (de)serializable configuration, which defines the structure of the module and its hyper-parameters.
|
||||
Parameters and hyper-parameters are not serialized into the same file and both are normaly necessary to load a module for inference.
|
||||
|
||||
### Optimization
|
||||
|
||||
Optimization is normally done with gradient descent (or ascent for reinforcement learning), and it is important to provide an easy API for optimizing modules.
|
||||
|
||||
#### Constraints
|
||||
|
||||
1. __Users should be able to control what is optimized.__
|
||||
Modules can contain anything for maximum flexibility, but not everything needs to be optimized.
|
||||
2. __Optimizers should have a serializable state that is updated during training.__
|
||||
Many optimizers keep track of previous gradients to implement some form of momentum.
|
||||
However, the state can be anything, not just tensors, allowing for easy implementation of any kind of optimizer.
|
||||
3. __The learning rate can be updated during training.__
|
||||
Learning rate schedulers are often used during training and should be considered as a key aspect.
|
||||
|
||||
#### Solution
|
||||
|
||||
The solution to this problem comprises multiple parts.
|
||||
Firstly, the `Optimizer` trait is quite similar to the `Module` trait in terms of saving and loading the state.
|
||||
Please refer to the [serialization](#serialization) section for more details.
|
||||
|
||||
Secondly, two traits were created.
|
||||
The `Optimizer` trait is general and relatively unopinionated, with a simple `step` method that takes a learning rate, a module, and the gradients.
|
||||
The other trait, `SimpleOptimizer`, aims to provide an easier API for implementing new optimizers.
|
||||
The goal is to allow implementations to avoid handling missing gradients, loading and exporting records, navigating the module parameter structure, handling tracked and untracked tensors, and other such tasks.
|
||||
|
||||
Thirdly, each tensor that will be optimized needs to be wrapped into a `Param` struct, which gives them an ID used for (de)serialization and to associate the state of the optimizer to each parameter.
|
||||
The `Module` trait has two ways to navigate over parameters.
|
||||
The first one is the `map` function, which returns `Self` and makes it easy to implement any transformation and mutate all parameters.
|
||||
The second one is the `visit` function, which has a similar signature but does not mutate the parameter tensors.
|
||||
|
||||
__SimpleOptimizer__
|
||||
|
||||
The `SimpleOptimizer` has two major assumptions:
|
||||
|
||||
1. The state of the optimizer is linked to each parameter.
|
||||
In other words, each parameter has its own optimizer state, decoupled from the other parameters.
|
||||
2. The state of the optimizer implements `Record`, `Clone`, and has a `'static` lifetime.
|
||||
|
||||
The benefits of those assumptions materialize in simplicity with little loss in flexibility.
|
||||
The state associative type is also generic over the dimension, making it extremely easy to include tensors in the state that share the same dimensionality as its parameter.
|
||||
|
||||
To wrap a simple optimizer into the more general `Optimizer` trait, the `OptimizerAdaptor` struct is used.
|
||||
|
||||
__OptimizerAdaptor__
|
||||
|
||||
The `OptimizerAdaptor` is a simple struct composed of a `SimpleOptimizer` and a hashmap with all records associated with each parameter ID.
|
||||
When performing an optimization step, the adaptor handles the following:
|
||||
|
||||
1. Updates each parameter tensor in the given module using the `Module::map` function.
|
||||
2. Checks if a gradient for the current tensor exists.
|
||||
3. Makes sure that the gradient, the tensor, and the optimizer state associated with the current parameter are on the same device.
|
||||
The device can be different if the state is loaded from disk to restart training.
|
||||
4. Performs the simple optimizer step using the inner tensor since the operations done by the optimizer should not be tracked in the autodiff graph.
|
||||
5. Updates the state for the current parameter and returns the updated tensor, making sure it's properly registered into the autodiff graph if gradients are maked as required.
|
||||
|
||||
Note that a parameter can still be updated by another process, as is the case with running metrics used in batch norm.
|
||||
These tensors are still wrapped using the `Param` struct so that they are included in the module's state and given a proper parameter ID, but they are not registered in the autodiff graph.
|
||||
|
||||
### Serialization
|
||||
|
||||
An important aspect of a deep learning framework is the ability to save and load models from disk.
|
||||
Despite appearing as a simple feature, it involves numerous constraints that require a proper solution.
|
||||
|
||||
#### Constraints
|
||||
|
||||
1. __Users should be able to declare the precision of the model to be saved, independent of the backend in use.__
|
||||
|
||||
The modules should not be duplicated in RAM in another precision to support this.
|
||||
Conversion should be done lazily during (de)serialization.
|
||||
|
||||
2. __Users should be able to add any field to a module, even fields that are not serializable.__
|
||||
|
||||
This can include constants, database connections, other module references, or any other information.
|
||||
Only parameters should be serialized since the structure of the module itself should be encapsulated with module configurations (hyper-parameters).
|
||||
|
||||
3. __Users should be able to declare the format in which the module should be saved.__
|
||||
|
||||
This can involve saving to a compressed JSON file or directly to bytes in memory for `no-std` environments.
|
||||
|
||||
4. __Users should be able to create a module with its saved parameters without having to initialize the module first.__
|
||||
|
||||
This will avoid unnecessary module initialization and tensor loading, resulting in reduced cold start when dealing with inference.
|
||||
|
||||
In addition to all of these constraints, the solution should be easy to use.
|
||||
|
||||
#### Solution
|
||||
|
||||
In order to be able to add any field to a module without requiring it to be (de)serializable, we decouple the module type from its state.
|
||||
We create a new type for each module that only contains the parameters that need to be saved.
|
||||
To generate that type automatically, the user must either declare which field is a parameter or a constant, or we assume that each field implements the module trait.
|
||||
|
||||
The second solution was chosen as it simplifies the code generation and reduces the size of the user API.
|
||||
This means that the `Module` trait should be implemented by [primitives types](./burn-core/src/module/param/primitive.rs).
|
||||
The following diagrams highlight the main types and traits used in the solution.
|
||||
|
||||
<div align="center">
|
||||
<h4>Module Serialization Types</h4>
|
||||
<img src="./assets/ModuleSerialization.png" width="700px"/>
|
||||
<div align="left">
|
||||
|
||||
The way the types interact with each other is pretty straightforward.
|
||||
First, a module can be converted into a record using `into_record()`.
|
||||
Note that tensors can be cloned, but it won't actually copy any data; it will create another reference to the same data.
|
||||
|
||||
Then, you can call `record()`, which takes a `RecordSettings` as a parameter.
|
||||
The function is automatically implemented for each record.
|
||||
It calls `into_item::<RecordSettings>()` on the record, which creates a serializable item following the given settings.
|
||||
Note that tensors implement record, and their item is just a wrapper struct that contains information about the precision in which the tensor should be saved or loaded.
|
||||
No actual copy of the tensor is made until this point.
|
||||
The tensor is converted to the `Data` struct and then converted into the specified precision only when `serialize()` or `deserialize()` are called, which makes the whole process lazy.
|
||||
|
||||
To recapitulate, the `Module` trait has an associated type that implements `Record`, which only contains the parameters of the model.
|
||||
The `Record` trait has a generic associated type (GAT) that specifies a family of types that can be (de)serialized given any `RecordSettings`.
|
||||
Records are therefore decoupled from the backend in use, and the saved items can be loaded on any backend with any precision, since the conversion is type-safe and done when `serialize()` and `deserialize()` are called.
|
||||
All of the types are generated using simple derive macros without any conditional statements or complex syntax, as `Record` and `Module` are implemented for all primitive types.
|
||||
This makes the code simple and easy to maintain.
|
||||
In addition, you can extend the current system with your own `Recorder` and `RecordSettings` to control how your modules should be saved and loaded.
|
||||
|
||||
##### Pros
|
||||
|
||||
* All constraints are respected.
|
||||
* The code is simple and easy to maintain, with very few conditional statements.
|
||||
It is just recursive data structures, where all the complexity is handled by the framework in primitive implementations.
|
||||
* The user API is simple and small, with only two derives (`Record` and `Module`) and no additional attributes.
|
||||
* Users can create their own `Module` and `Record` primitive types, which gives them the flexibility to control how their data is serialized without having to fork the framework.
|
||||
|
||||
##### Cons
|
||||
|
||||
* There are more types, but most of them are automatically generated and single-purpose, so users don't need to interact with them for common use cases.
|
||||
However, they can do so if necessary.
|
||||
* When instantiating a new record manually, each field must be set to something, even if the type itself is `()`, which represents no value.
|
||||
Since the code generation step uses associative types, it doesn't know that a field type is actually nothing.
|
||||
Creating a record manually without using the generated function `into_record` or loading it from a file is only useful to load a set of parameters into a module from an arbitrary source.
|
||||
Using the record may not be the optimal solution to this problem, and another API could be created in the future.
|
||||
|
||||
##### Compatibility
|
||||
|
||||
Record may become incompatible with previous versions of Burn, depending on the chosen format.
|
||||
The more compact format (bincode) store minimal information about the type, making it significantly smaller but less resilient to type changes such adding an optional field.
|
||||
At some point, it might be necessary to provide a translation script that can translate a more resilient format from a previous version to a more compact one.
|
||||
Binary file not shown.
|
After Width: | Height: | Size: 109 KiB |
|
|
@ -0,0 +1,134 @@
|
|||
<mxfile host="app.diagrams.net" modified="2023-04-23T15:16:56.674Z" agent="Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/112.0.0.0 Safari/537.36" etag="yRaqMiYc5aCAWcy34Vtg" version="21.1.9" type="device">
|
||||
<diagram name="Page-1" id="p9OtIezBOMlZGQ46mofZ">
|
||||
<mxGraphModel dx="2181" dy="3049" grid="1" gridSize="10" guides="1" tooltips="1" connect="1" arrows="1" fold="1" page="1" pageScale="1" pageWidth="827" pageHeight="1169" math="0" shadow="0">
|
||||
<root>
|
||||
<mxCell id="0" />
|
||||
<mxCell id="1" parent="0" />
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-37" value="" style="rounded=1;whiteSpace=wrap;html=1;strokeWidth=3;" vertex="1" parent="1">
|
||||
<mxGeometry x="950" y="-670" width="260" height="270" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-9" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;entryX=0.5;entryY=0;entryDx=0;entryDy=0;strokeWidth=3;endArrow=open;endFill=0;endSize=14;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-1" target="0A-74KZi9Q9iTp0nixR_-3">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-1" value="Module" style="rounded=1;whiteSpace=wrap;html=1;fontSize=18;fontStyle=1;strokeWidth=3;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
|
||||
<mxGeometry x="30" y="-1140" width="170" height="70" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-7" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;entryX=0;entryY=0.5;entryDx=0;entryDy=0;strokeWidth=3;endArrow=diamondThin;endFill=0;fontStyle=1;endSize=20;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-3" target="0A-74KZi9Q9iTp0nixR_-6">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-3" value="<span style="font-size: 18px;">Module<br style="font-size: 18px;">Trait<br style="font-size: 18px;"></span>" style="rhombus;whiteSpace=wrap;html=1;strokeWidth=3;fontSize=18;fontStyle=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
|
||||
<mxGeometry x="40" y="-990" width="150" height="150" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-30" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;strokeWidth=3;endArrow=diamondThin;endFill=0;endSize=20;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-6" target="0A-74KZi9Q9iTp0nixR_-29">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-6" value="<span style="font-size: 18px;">Record<br style="font-size: 18px;">Trait<br style="font-size: 18px;"></span>" style="rhombus;whiteSpace=wrap;html=1;strokeWidth=3;fontSize=18;fontStyle=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
|
||||
<mxGeometry x="290" y="-990" width="150" height="150" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-24" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;entryX=0.5;entryY=0;entryDx=0;entryDy=0;strokeWidth=3;endArrow=open;endFill=0;endSize=14;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-10" target="0A-74KZi9Q9iTp0nixR_-6">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-10" value="Module<br style="font-size: 18px;">Record" style="rounded=1;whiteSpace=wrap;html=1;fontSize=18;fontStyle=1;strokeWidth=3;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
|
||||
<mxGeometry x="280" y="-1140" width="170" height="70" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-12" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;entryX=0;entryY=0.5;entryDx=0;entryDy=0;strokeWidth=3;endArrow=diamondThin;endFill=0;fontStyle=1;endSize=20;" edge="1" parent="1" target="0A-74KZi9Q9iTp0nixR_-13">
|
||||
<mxGeometry relative="1" as="geometry">
|
||||
<mxPoint x="973" y="-585" as="sourcePoint" />
|
||||
<mxPoint x="1053" y="-585" as="targetPoint" />
|
||||
</mxGeometry>
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-13" value="Associative type" style="text;html=1;strokeColor=none;fillColor=none;align=left;verticalAlign=middle;whiteSpace=wrap;rounded=0;fontSize=16;fontStyle=1;spacingLeft=6;" vertex="1" parent="1">
|
||||
<mxGeometry x="1040" y="-600" width="140" height="30" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-14" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;strokeWidth=3;endArrow=open;endFill=0;endSize=14;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" target="0A-74KZi9Q9iTp0nixR_-15">
|
||||
<mxGeometry relative="1" as="geometry">
|
||||
<mxPoint x="973" y="-540" as="sourcePoint" />
|
||||
<mxPoint x="1023" y="-540.5" as="targetPoint" />
|
||||
</mxGeometry>
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-15" value="Implement trait" style="text;html=1;strokeColor=none;fillColor=none;align=left;verticalAlign=middle;whiteSpace=wrap;rounded=0;fontSize=16;fontStyle=1;spacingLeft=6;" vertex="1" parent="1">
|
||||
<mxGeometry x="1040" y="-555" width="140" height="30" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-16" value="<span><br></span>" style="rhombus;whiteSpace=wrap;html=1;strokeWidth=3;fontSize=18;fontStyle=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
|
||||
<mxGeometry x="990" y="-508" width="30" height="30" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-17" value="Trait" style="text;html=1;strokeColor=none;fillColor=none;align=left;verticalAlign=middle;whiteSpace=wrap;rounded=0;fontSize=16;fontStyle=1;spacingLeft=6;" vertex="1" parent="1">
|
||||
<mxGeometry x="1040" y="-508" width="140" height="30" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-18" value="" style="rounded=1;whiteSpace=wrap;html=1;fontSize=22;fontStyle=1;strokeWidth=3;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
|
||||
<mxGeometry x="985" y="-460" width="40" height="20" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-19" value="Type" style="text;html=1;strokeColor=none;fillColor=none;align=left;verticalAlign=middle;whiteSpace=wrap;rounded=0;fontSize=16;fontStyle=1;spacingLeft=6;" vertex="1" parent="1">
|
||||
<mxGeometry x="1040" y="-465" width="140" height="30" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-26" value="<span style="font-size: 18px;">Serialize<br style="font-size: 18px;"></span>" style="rhombus;whiteSpace=wrap;html=1;strokeWidth=3;fontSize=18;fontStyle=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
|
||||
<mxGeometry x="220" y="-545" width="120" height="120" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-27" value="<span style="font-size: 18px;">Deserialize<br style="font-size: 18px;"></span>" style="rhombus;whiteSpace=wrap;html=1;strokeWidth=3;fontSize=18;fontStyle=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
|
||||
<mxGeometry x="390" y="-545" width="120" height="120" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-31" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;strokeWidth=3;endArrow=open;endFill=0;endSize=14;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-29" target="0A-74KZi9Q9iTp0nixR_-26">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-32" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;entryX=0.5;entryY=0;entryDx=0;entryDy=0;strokeWidth=3;endArrow=open;endFill=0;endSize=14;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-29" target="0A-74KZi9Q9iTp0nixR_-27">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-29" value="Record Item<br style="font-size: 18px;">&lt;<font color="#b85451" style="font-size: 18px;">RecordSettings</font>&gt;" style="rounded=1;whiteSpace=wrap;html=1;fontSize=18;fontStyle=1;strokeWidth=3;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
|
||||
<mxGeometry x="240" y="-750" width="250" height="100" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-40" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;entryX=0;entryY=0.5;entryDx=0;entryDy=0;endArrow=diamondThin;endFill=0;endSize=20;strokeWidth=3;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-33" target="0A-74KZi9Q9iTp0nixR_-36">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-41" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;entryX=0;entryY=0.5;entryDx=0;entryDy=0;endArrow=diamondThin;endFill=0;endSize=20;strokeWidth=3;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-33" target="0A-74KZi9Q9iTp0nixR_-35">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-42" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;entryX=0;entryY=0.5;entryDx=0;entryDy=0;endArrow=diamondThin;endFill=0;endSize=20;strokeWidth=3;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-33" target="0A-74KZi9Q9iTp0nixR_-34">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-33" value="<span style="font-size: 18px;">Record<br style="font-size: 18px;">Settings<br style="font-size: 18px;"></span>" style="rhombus;whiteSpace=wrap;html=1;strokeWidth=3;fontSize=18;fontStyle=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
|
||||
<mxGeometry x="520" y="-1150" width="150" height="150" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-34" value="Float Elem" style="rounded=1;whiteSpace=wrap;html=1;fontSize=18;fontStyle=1;strokeWidth=3;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
|
||||
<mxGeometry x="670" y="-980" width="170" height="70" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-35" value="Int Elem" style="rounded=1;whiteSpace=wrap;html=1;fontSize=18;fontStyle=1;strokeWidth=3;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
|
||||
<mxGeometry x="670" y="-890" width="170" height="70" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-36" value="Recorder" style="rounded=1;whiteSpace=wrap;html=1;fontSize=18;fontStyle=1;strokeWidth=3;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
|
||||
<mxGeometry x="670" y="-800.5" width="170" height="70" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-39" value="<b><font style="font-size: 23px;">Legend</font></b>" style="text;html=1;strokeColor=none;fillColor=none;align=center;verticalAlign=middle;whiteSpace=wrap;rounded=0;" vertex="1" parent="1">
|
||||
<mxGeometry x="950" y="-650" width="260" height="30" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-44" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;entryX=0;entryY=0.5;entryDx=0;entryDy=0;endArrow=diamondThin;endFill=0;endSize=20;strokeWidth=3;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-47" target="0A-74KZi9Q9iTp0nixR_-50">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-45" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;entryX=0;entryY=0.5;entryDx=0;entryDy=0;endArrow=diamondThin;endFill=0;endSize=20;strokeWidth=3;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-47" target="0A-74KZi9Q9iTp0nixR_-49">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-46" style="edgeStyle=orthogonalEdgeStyle;rounded=0;orthogonalLoop=1;jettySize=auto;html=1;entryX=0;entryY=0.5;entryDx=0;entryDy=0;endArrow=diamondThin;endFill=0;endSize=20;strokeWidth=3;fontSize=18;" edge="1" parent="1" source="0A-74KZi9Q9iTp0nixR_-47" target="0A-74KZi9Q9iTp0nixR_-48">
|
||||
<mxGeometry relative="1" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-47" value="<span style="font-size: 18px;">Recorder<br style="font-size: 18px;"></span>" style="rhombus;whiteSpace=wrap;html=1;strokeWidth=3;fontSize=18;fontStyle=1;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
|
||||
<mxGeometry x="890" y="-1150" width="150" height="150" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-48" value="Record Args" style="rounded=1;whiteSpace=wrap;html=1;fontSize=18;fontStyle=1;strokeWidth=3;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
|
||||
<mxGeometry x="1040" y="-980" width="170" height="70" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-49" value="Record Output" style="rounded=1;whiteSpace=wrap;html=1;fontSize=18;fontStyle=1;strokeWidth=3;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
|
||||
<mxGeometry x="1040" y="-890" width="170" height="70" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-50" value="Load Args" style="rounded=1;whiteSpace=wrap;html=1;fontSize=18;fontStyle=1;strokeWidth=3;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1">
|
||||
<mxGeometry x="1040" y="-800.5" width="170" height="70" as="geometry" />
|
||||
</mxCell>
|
||||
<mxCell id="0A-74KZi9Q9iTp0nixR_-52" value="" style="edgeStyle=none;orthogonalLoop=1;jettySize=auto;html=1;rounded=0;" edge="1" parent="1">
|
||||
<mxGeometry width="80" relative="1" as="geometry">
|
||||
<mxPoint x="-660" y="-1310" as="sourcePoint" />
|
||||
<mxPoint x="-580" y="-1310" as="targetPoint" />
|
||||
<Array as="points" />
|
||||
</mxGeometry>
|
||||
</mxCell>
|
||||
</root>
|
||||
</mxGraphModel>
|
||||
</diagram>
|
||||
</mxfile>
|
||||
Loading…
Reference in New Issue