Agrippa
This python package is meant to assist in building/understanding/analyzing machine learning models. The core of the system is a markup language that can be used to specify a model architecture. This package contains utilities to convert that language into the ONNX format, which is compatible with a variety of deployment options and ML frameworks.
Installation
Agrippa can be installed with pip install agrippa
. The requirements.txt
file contains dependencies to run both the package and the tests found in the tests
folder.
Usage
The principal function is export, which takes a project folder and compiles the contents into a .onnx file.
import agrippa
model_dir = '../path/to/dir'
agrippa.export(model_dir, 'outfile_name.onnx')
The function header for export is:
def export(
infile, # the markup file
outfile=None, # the .onnx file name
producer="Unknown", # your name
graph_name="Unknown", # the name of the model
write_weights=True, # should this create a weights file, should none exist
suppress=False, # suppresses certain print statements
reinit=False, # reinitializes all weights from the weights file
bindings=None # a dictionary to bind variables present in the markup
):
Markup Language Spec
A project should be bundled into its own directory, which should have three files:
- One file with the extension
.agr
or .xml
specifying the model architecture - A
weights.pkl
file to specify the parameter values in the model (optional) - A
meta.json
file to define certain metadata, like the producer name (optional)
The architecture file is parsed like XML, so it should be well-formed XML. Recall that tags with no respective end-tag should end with \>
, and all attributes should be formatted like strings with quotes around them.
Markup Syntax
Every architecture file should be encased in a <model>
tag, ideally with an attribute script-version="0.0.0"
(the current version).
An example script that does only one matrix multiply might look like this:
<model script-version="0.0.1">
<import dim="[5, 1]" from="input" type="float32" />
<block title="Projection">
<import from="input" />
<node title="Linear" op="MatMul">
<params dim="[5, 5]" name="W" type="float32" />
<input dim="[var(features), 1]" src="input" />
<output dim="[5, 1]" name="y" />
</node>
<export from="y" />
</block>
<export dim="[5, 1]" from="y" type="float32" />
</model>
There are only three types of root-level tags that are allowed: <import>
, <export>
, and <block>
. The import and export tags specify the inputs and outputs of the entire model, respectively. There may be multiple of each type, but each type must appear at least once. Each import and export tag must have three attributes: dim
, from
, and type
. They are used like so:
<import dim="[3, 1]" from="input" type="float32" />
<export dim="[3, 1]" from="y" type="float32" />
The from
name for the export matches the name you should expect from ONNX runtime. It should also match the output of the last node from which you are piping output.
Most of the architechture should be contained inside <block>
tags. These tags take a title attribute, which does not need to be unique. Importantly, <node>
tags must be inside blocks. Block tags should contain <import>
and <export>
tags (with the attributes mentioned above) specifying all of the inputs/outputs the underlying nodes inside the block use.
Nodes define operations. Their op
attribute defines the ONNX op type they will be converted to. They must also have a title
attribute, which is unique. Nodes must also contain appropriate <input>
, <output>
, and <params>
tags. The <input>
and <params>
tags need to be in the order specified in the ONNX documentation for a particular node type. See an example node:
<node title="Linear" op="MatMul">
<params dim="[3, 3]" name="W" type="float32" shared="no" />
<input dim="[3, 1]" src="input" />
<output dim="[3, 1]" name="linear" />
</node>
Parameters, which are specified using the <params>
tag, take a name
attribute (unique only for non-shared parameters), a dim
attribute, a type
attribute, and an optional shared
attribute. The shared
attribute should equal "yes" or "no". It specifies whether a parameter name is meant to be unique; by default, parameters which share the same name (such as in a repitition) become independent values upon compilation.
Repetitions
Blocks may take a rep
attribute, which defines how many times a block should be stacked on top of itself. Its outputs are passed to its inputs and so on. The number of inputs and the number of outputs need not match (they are matched based on order; note that if you want to use intermediate outputs, you must account for name mangling in repeated blocks). Even though the names of the outputs are mangled during repetitions, you may use the outputs in your markup with consideration to that fact: simply refer back to the name you specified, which is automatically mapped to the last name in the repetition.
Variable Bindings
The agrippa.export
function takes an optional argument, bindings
. The bindings
parameter is meant to be a dictionary of variables, set by the user, to replace areas in the markup file where the var
function is used. For example, if an input tag has a dim
attribute set to "[var(image_width), var(image_height)]"
, a binding of {'image_width': 512, 'image_height': '256'}
would set all occurances of var(image_height)
to 512
and all occurances of var(image_height)
to 256
. Note that in all cases, strings are used, since xml attributes require strings; the values are type-casted upon compilation.
Expressions
Attributes also support expressions using expr()
. For example, in order to specify that a parameter should be initialized to the square of a variable (supplied in bindings), you could use:
<params name="squared" dim="[2, 2]" init="constant" init_args="[expr(my_var^2)]">
.
Note that the expression goes inside the list (expressions do not support lists). They support to following binary operators: ^
, *
, /
, %
, -
, +
.
Also note that expr(my_var)
and var(my_var)
are equivalent.
Weight Initialization
By default, weights are initialized with a standard normal distribution. There are ways to specify other initializations for each parameter, however. The params
tag takes an optional init
attribute along with an optional init_args
attribute. The init_args
attribute must always be some value (non-string), such as a list (e.g., init_args="[2, 3]"
). Recall that all attributes are specified with double quotation marks) The options for initialization are:
Value | Description | Arguments |
---|
normal | Normally distributed | A list of two numbers, the first defining the mean and the second defining the standard deviation. |
uni_random | Uniformly random in [a, b) | A list of two numbers, the first defining the a and the second defining b. |
zeros | All zeros | None |
ones | All ones | None |
constant | Initializes tensor to specified value | The first argument in the list is the value |
Frozen Parameters
In order to freeze a parameter, you can set the frozen
attribute equal to yes
. Internally, this option adds $constant
to the ONNX initialization names. When importing the parameter into PyTorch using the conversion tool, the $constant
indicates that the initializer should be added as a buffer (constant) rather than a parameter.
Importing From Other Files
Another file can be used in your model by using a block
tag with a src
attribute. Like so:
<block src="path/to/file.agr" name="imported_file" />
The name
attribute defines how you refer to imports/exports of the imported model. For example, if the linked model has a root level import with name inputs
, an output (or import) (in the original file) with name imported_file$inputs
will be automatically passed to the imported model. Likewise, an export can be referred to in the original file by specifying an input with name imported_file$out_name_from_imp_file
.
Other Rules
Names
Node titles are optional (a default, unique game is given to them upon compilation). Parameter names should be unique only when they are not shared parameters; parameters inside repeated blocks will have their names mangled so that they are unique. Name mangling affects parameters, node titles, and output/input names separately.
Types
Types by default are set to float32.
Dimensions
Specifying the dimensions of inputs and outputs are optional. Specifying the dimensions of imports and exports are only required at the root level, though it is recommended that you specify them for clarity.
Any behavior not mentioned here is undefined.
Supported Types
The only currently supported type is float32
.
Supported ONNX OpTypes
The currently supported op types are:
ONNX OpType | Tested ONNX Compile Support | Tested Training Support |
---|
Add | Yes | Yes |
Concat | Yes | Yes |
Identity | Yes | Yes |
LeakyRelu | Yes | Yes |
LpNormalization | Yes | Yes |
MatMul | Yes | Yes |
Mul | Yes | Yes |
Relu | Yes | Yes |
ReduceMean | Yes | Yes |
Softmax | Yes | Yes |
Sqrt | Yes | Yes |
Sub | Yes | Yes |
Transpose | Yes | Yes |
Additional notes on functionality that might differ from ONNX. For most details, see the Onnx documentation.
ONNX OpType | Notes |
---|
Transpose | Important difference with the Onnx documentation: by default, when imported into PyTorch, the transpose operator will keep the first dimension the same so as to support batching. The Onnx default behavior is to reverse all the dimensions. |
Syntax Highlighting in VSCode
If you'd like to use the extension .agr
for clarity, you can enable syntax highlighting in vscode by placing the following in a settings.json file:
"files.associations": {
"*.agr": "xml"
}
To create that settings file, use the command pallet (CTRL-SHIFT-P), type settings.json
, and choose the appropriate option.
Training
ONNX is built for inference. However, various supporters of ONNX, including Microsoft (via the onnxruntime), have tried to implement training support. As far as I can tell, Microsoft gave up on trying to support training ONNX files directly. Many of the training tools in onnxruntime are either experimental or scheduled to be depricated. What they did end up implementing was a tool to train PyTorch models (i.e. objects of classes that inherit from torch.nn.Module). Their tool is more narrowly for speeding up training that you could already do natively in PyTorch, and it is not used in this project.
Another option, besides trying to rely on existing ONNX training projects, would have been to make our own. It is actually relatively straightforward: the ONNX file itself is a highly expressive computational graph. We could build a separate graph for training, which has gradient nodes added. It could even take parameters as input and output new parameters while keeping all the data on a GPU. The key is having access to (or building from scratch) nodes that can compute the gradient of each operation (there are many, but they are relatively simple). I ultimately decided (like Microsoft) that this was not worth the pain.
Instead, we opt for converting onnx files to PyTorch. We provide utilities to do that and to use the training features of PyTorch.
Unfortunately, PyTorch does not natively support importing ONNX files. But there is a work-around: building on top of some community tools, we can make our own ONNX to PyTorch converter that is suitable for training. There is more information in the README.md under src/agrippa/onnx2torch for details on exactly how a particular community project was modified. It does not support all ONNX operations, but neither does our markup language.
Usage
The following code snippet takes a project directory, converts it to an onnx file, then uses the build-in ONNX-to-PyTorch converter to create a PyTorch model, which can be trained in the usual way.
import agrippa
proj_name = 'simple-project'
onnx_out = 'simple_testing.onnx'
agrippa.export(proj_name, onnx_out)
torch_model = agrippa.onnx_to_torch(onnx_out)
Utilities
Some utilities are available in agrippa.utils. This includes find_params
, which returns weight (parameter) names and values as a dictionary. It also includes save_torch_model
, which takes trained weights from a PyTorch model and saves them into a fresh weights.pkl
file.
Finding Parameters
The returned dictionary includes parameters whose names contain the name
argument (first argument) as a substring. Searching for weights in this way is recommended, since the names of parameters might be changed when the markup is compiled (for example, the names of weights that appear in repeated blocks). The find_params
function takes two mandatory parameters and one optional: the substring that will be matched (mandatory), the directory of the project (mandatory), and the path to the weights file name within that directory (optional).
Example
matches = agrippa.utils.find_params('bias', 'FNN')
print(matches)
The above code might print:
{'biases$1': array([[-0.77192398],
[-0.02351803],
[-0.00533084],
[ 0.13640493],
[-0.12087004]]), 'biases$2': array([[-0.18979854],
[-0.15769928],
[ 0.46656397],
[-0.10602235]])}
Saving Model from PyTorch
After importing your model to PyTorch using agrippa.onnx_to_torch
, you probably would like to save the trained weights. When imported into PyTorch, the names of the weights change slightly, so it is recommended that you save your models using agrippa.utils.save_torch_model
, which takes as parameters the PyTorch model, the project directory, and (optionally) the weights filename inside that directory. Under the hood, this function loops over the state_dict
of the PyTorch model, removes initializer.
from the parameter's name, and saves it inside a dictionary to weights.pkl
.
Example
# ... training loop
agrippa.utils.save_torch_model(torch_model, "my-project", "weights.pkl")
Examples
The following architecture is a simple feed forward network with five layers followed by a normalization. The architecture is organized into two blocks, the FFN and the norm layer. Inside the FFN is a FFN Layer block, which is repeated five times.
<model script-version="0.0.1">
<import dim="[3, 1]" from="input" type="float32" />
<block title="FFN">
<import from="input" />
<block title="FFN Layer" rep="5">
<import from="input" />
<node title="Linear" op="MatMul">
<params dim="[3, 3]" name="W" type="float32" />
<input dim="[3, 1]" src="input" />
<output dim="[3, 1]" name="linear" />
</node>
<node title="Bias" op="Add">
<params dim="[3, 1]" name="B" type="float32" />
<input dim="[3, 1]" src="linear" />
<output dim="[3, 1]" name="biased" />
</node>
<node title="ReLu" op="Relu">
<input dim="[3, 1]" src="biased" />
<output dim="[3, 1]" name="relu" />
</node>
<export from="relu" />
</block>
</block>
<block title="Norm">
<import from="relu" />
<node title="ReLu" op="LpNormalization" axis="0" p="1">
<input dim="[3, 1]" src="relu" />
<output dim="[3, 1]" name="y" />
</node>
<export from="y" />
</block>
<export dim="[3, 1]" from="y" type="float32" />
</model>
You can find more example projects inside the tests
folder.