IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[TF FE] Update TensorFlow Frontend developer guide (#17521)

Browse Source 
* [TF FE] Update TensorFlow Frontend developer guide

Signed-off-by: Kazantsev, Roman <roman.kazantsev@intel.com>

* Apply suggestions from code review

Co-authored-by: Tatiana Savina <tatiana.savina@intel.com>

---------

Signed-off-by: Kazantsev, Roman <roman.kazantsev@intel.com>
Co-authored-by: Tatiana Savina <tatiana.savina@intel.com>
This commit is contained in:
Roman Kazantsev 2023-05-15 13:14:36 +03:00 committed by GitHub
parent 49b0c7248a
commit ac3d339811
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 14 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
src/frontends/tensorflow/README.md
View File

@ -2,7 +2,8 @@
The TensorFlow Frontend (TF FE) is a C++ based OpenVINO Frontend component that is responsible for reading and converting a TensorFlow model to an `ov::Model` object
that further can be serialized into the Intermediate Representation (IR) format.
This is an internal API for OpenVINO that is used to implement user facing API such as Model Optimizer, `read_model` function, and OpenVINO Integration with TensorFlow.
This is an internal API for OpenVINO that is used to implement user-facing API such as MO tool, MO Python API, and OpenVINO Runtime `read_model` function
for reading TensorFlow models of the original format in run-time. Also, OpenVINO Model Server uses the frontend for serving models.
Regular users should not use the frontend directly.
```mermaid
@ -15,8 +16,8 @@ flowchart BT
style model3 fill:#427cb0
ov_model[(ov::Model)]
style ov_model fill:#427cb0
ovtf(OpenVINO Integration with TensorFlow)
style ovtf fill:#ffffc2
ovms(OpenVINO Model Server)
style ovms fill:#ffffc2
tf_fe(TensorFlow Frontend)
style tf_fe fill:#ee9a4d
fem(Frontend Manager)
@ -24,13 +25,13 @@ flowchart BT
ov_runtime(OpenVINO Runtime)
model --> mo --> fem --> tf_fe
model2 --> ov_runtime --> fem
model3 --> ovtf --> tf_fe
model3 --> ovms --> ov_runtime
tf_fe --> ov_model
click ovtf "https://github.com/openvinotoolkit/openvino_tensorflow"
click ovms "https://github.com/openvinotoolkit/model_server"
```
Currently, it is only used by [OpenVINO Integration with TensorFlow](https://github.com/openvinotoolkit/openvino_tensorflow).
Model Optimizer for now relies on the legacy [TensorFlow Frontend](https://docs.openvino.ai/latest/openvino_docs_MO_DG_prepare_model_convert_model_Convert_Model_From_TensorFlow.html) developed in Python.
The MO tool and MO Python API now use the TensorFlow Frontend as the default path for conversion to IR.
Known limitations of TF FE are described [here](https://docs.openvino.ai/latest/openvino_docs_MO_DG_TensorFlow_Frontend.html).
## Key contacts
@ -45,7 +46,7 @@ The structure of OpenVINO TensorFlow Frontend sources includes the following dir
* [src](./src/) folder contains the sources of the component.
* [tests](./tests) cover internal transformations.
Additionally, there is a shared [tensorflow common](../tensorflow_common) directory with same structure and purposes.
Additionally, there is a shared [TensorFlow Common](../tensorflow_common) directory with same structure and purposes.
Its content depend only on common FrontEnd APIs thus is free to use in other FrontEnds.
## Architecture
@ -56,8 +57,8 @@ The whole workflow can be split into two steps: model loading and conversion.
During loading, the `FrontEnd::load()` method creates `InputModel` that encapsulates the `GraphIterator` object.
`GraphIterator` is a reader that iterates through the graph nodes in the topological order.
`GraphIterator::get_decoder()` provides a decoder for the current graph node to read its attributes.
Each TensorFlow model format has its implementation of `GraphIterator`. Currently, the frontend supports only binary frozen format `.pb`,
and `GraphIteratorProto` is used for reading and parsing this format. The architecture of the loading step is shown in the picture below:
Each TensorFlow model format has its implementation of `GraphIterator`. Currently, the frontend supports SavedModel, MetaGraph (`.meta`), and frozen protobuf (`.pb` and `.pbtxt`) formats.
The base class `GraphIteratorProto` is used for reading and parsing these formats. The architecture of the loading step is shown in the picture below:
```mermaid
classDiagram
@ -70,6 +71,9 @@ classDiagram
Place --o "1..*" InputModel
DecoderBase "1" --o "1" Place
GraphIteratorProto ..|> GraphIterator
GraphIteratorProtoTxt ..|> GraphIterator
GraphIteratorMeta ..|> GraphIterator
GraphIteratorSavedModel ..|> GraphIterator
```
After the loading step, `InputModel` includes a container of topologically sorted operation `Place` objects.