Refactor Target Platform Capabilities Design #1276

lior-dikstein · 2024-11-24T15:15:12Z

Create a new schema package to house all target platform modeling classes
Introduce a new versioning system with minor and patch versions

Additional Changes:

Update existing target platform models to adhere to the new versioning convention
Add necessary metadata
Correct all import statements
Update and enhance tests to reflect the design changes

Pull Request Description:

Checklist before requesting a review:

I set the appropriate labels on the pull request.
I have added/updated the release note draft (if necessary).
I have updated the documentation to reflect my changes (if necessary).
All function and files are well documented.
All function and classes have type hints.
There is a licenses in all file.
The function and variable names are informative.
I have checked for code duplications.
I have added new unittest (if necessary).

- Create a new `schema` package to house all target platform modeling classes - Introduce a new versioning system with minor and patch versions Additional Changes: - Update existing target platform models to adhere to the new versioning convention - Add necessary metadata - Correct all import statements - Update and enhance tests to reflect the design changes

model_compression_toolkit/metadata.py

model_compression_toolkit/target_platform_capabilities/schema/v1.py

ofirgo · 2024-11-25T12:26:37Z

model_compression_toolkit/core/common/graph/base_node.py

-from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationConfigOptions, \
-    TargetPlatformCapabilities, LayerFilterParams, OpQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.target_platform import TargetPlatformCapabilities, LayerFilterParams
+from model_compression_toolkit.target_platform_capabilities.schema.v1 import OpQuantizationConfig, \


Given that the schema is going to be updated (maybe frequently), and MCT core would change the schema that it is using accordingly, we need to figure out how to modify these imports throughout the code without accessing ".v1" directly at each import.

Is there a way to "export" (as TPC package "API") a default schema that references the currently used version, such that all imports will point to it and when we want to change the used schema by MCT we'll only have to change in 1 place?

Maybe @irenaby would have an idea how it can be done?

Let's discuss this offline if needed.

BTW, this is not mandatory for this PR, but solving this here would be better, because it will save us editing all these files again in a separate PR

The only part of MCT that should be aware of the schema is the parser once we have it. It should parse the schema into whatever representation the rest of mct works with. If we want to reuse the same classes for now, we can add a proxy module that will only import the classes from the schema, and the rest of mct imports from that proxy module.

I agree with @irenaby , we should have some proxy model at this stage...

model_compression_toolkit/target_platform_capabilities/target_platform/fusing.py

...l_compression_toolkit/target_platform_capabilities/target_platform/op_quantization_config.py

model_compression_toolkit/target_platform_capabilities/target_platform/operators.py

model_compression_toolkit/target_platform_capabilities/target_platform/target_platform_model.py

...atform_capabilities/target_platform/targetplatform2framework/target_platform_capabilities.py

ofirgo · 2024-11-25T12:37:07Z

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v1/tp_model.py

@@ -102,7 +102,7 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
        signedness=Signedness.AUTO)

    # We define an 8-bit config for linear operations quantization, that include a kernel and bias attributes.
-    linear_eight_bits = tp.OpQuantizationConfig(
+    linear_eight_bits = model_compression_toolkit.target_platform_capabilities.schema.v1.OpQuantizationConfig(


maybe import the path as "schema_v1" (or create an alias at the beginning of the file) instead of repeating it everywhere?
or do we want it to be explicit in the TPC model description? @haihabi
even if we want it to, eventually this file will be a JSON and not written as code using imports, so these fields in the JSON won't include the entire schema path anyway...

But the JSON file should have the schema version from the constant in the schema main class.

But the JSON file should have the schema version from the constant in the schema main class.

So you agree? the path here can be shortened with an alias to improve file readability?
@haihabi

tests/common_tests/test_tp_model.py

irenaby · 2024-11-26T14:26:53Z

model_compression_toolkit/metadata.py

+    tpc_patch_version = f'{tpc.tp_model.tpc_patch_version}'
+    tpc_platform_type = f'{tpc.tp_model.tpc_platform_type}'
+    tpc_schema = f'{tpc.tp_model.SCHEMA_VERSION}'
+    return {MCT_VERSION: mct_version,


If you need to access the fields elsewhere, why not define class or named tuple? If this goes directly into the model, no need to define global consts.

1. Fixed imports from long strings to alias "schema". 2. Fixed some failing tests. 3. Removed unused files.

1. Created mct_current_schema.py for one place to update the schema version, and replaced all the imports in mct to work with this location. 2. Replaced metadata dictionary with dataclass.

ofirgo · 2024-12-02T11:52:17Z

model_compression_toolkit/qat/keras/quantizer/README.md

@@ -5,7 +5,7 @@
 Several training methods may be applied by the user to train the QAT ready model
 created by `keras_quantization_aware_training_init` method in [`keras/quantization_facade`](../quantization_facade.py).
 Each `TrainingMethod` (an enum defined in the [`qat_config`](../../common/qat_config.py)) 
-and [`QuantizationMethod`](../../../target_platform_capabilities/target_platform/op_quantization_config.py)
+and `QuantizationMethod`


fix the path according to the new file location

lior-dikstein requested review from ofirgo and irenaby November 24, 2024 15:15

github-actions bot added auto:core auto:target_platform_capabilities auto:tests labels Nov 24, 2024

ofirgo reviewed Nov 25, 2024

View reviewed changes

model_compression_toolkit/metadata.py Outdated Show resolved Hide resolved

ofirgo reviewed Nov 25, 2024

View reviewed changes

model_compression_toolkit/target_platform_capabilities/schema/v1.py Show resolved Hide resolved

Added tpc_platfrom_type to meta data for tp model

536dc92

ofirgo reviewed Nov 25, 2024

View reviewed changes

ofirgo requested changes Nov 25, 2024

View reviewed changes

irenaby reviewed Nov 26, 2024

View reviewed changes

PR fixes:

6e3bfb5

1. Fixed imports from long strings to alias "schema". 2. Fixed some failing tests. 3. Removed unused files.

github-actions bot added the auto:qat label Nov 27, 2024

PR fixes:

be71127

1. Created mct_current_schema.py for one place to update the schema version, and replaced all the imports in mct to work with this location. 2. Replaced metadata dictionary with dataclass.

ofirgo approved these changes Dec 2, 2024

View reviewed changes

lior-dikstein merged commit 28b461b into main Dec 2, 2024
42 checks passed

lior-dikstein deleted the tpc_version branch December 2, 2024 13:11

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Refactor Target Platform Capabilities Design #1276

Refactor Target Platform Capabilities Design #1276

lior-dikstein commented Nov 24, 2024 •

edited

Loading

ofirgo Nov 25, 2024

ofirgo Nov 25, 2024

irenaby Nov 26, 2024

haihabi Nov 26, 2024

ofirgo Nov 25, 2024

haihabi Nov 26, 2024

ofirgo Nov 28, 2024 •

edited

Loading

haihabi Nov 28, 2024

irenaby Nov 26, 2024

ofirgo Dec 2, 2024

Refactor Target Platform Capabilities Design #1276

Refactor Target Platform Capabilities Design #1276

Conversation

lior-dikstein commented Nov 24, 2024 • edited Loading

Pull Request Description:

Checklist before requesting a review:

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ofirgo Nov 28, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

lior-dikstein commented Nov 24, 2024 •

edited

Loading

ofirgo Nov 28, 2024 •

edited

Loading