Renaming TPC

.github/labeler.yml

@@ -27,7 +27,7 @@ auto:qat:
 - model_compression_toolkit/qat/**
 
 auto:target_platform_capabilities:
-- model_compression_toolkit/target_platform_capabilities/**
+- model_compression_toolkit/framework_quantization_capabilities/**
 
 auto:trainable_infrastructure:
 - model_compression_toolkit/trainable_infrastructure/**

docsrc/source/api/api_docs/index.rst

@@ -106,9 +106,9 @@ keras_load_quantized_model
 - :ref:`keras_load_quantized_model<ug-keras_load_quantized_model>`: A function to load a quantized keras model.
 
 
-target_platform
-================
-- :ref:`target_platform<ug-target_platform>`: Module to create and model hardware-related settings to optimize the model according to, by the hardware the optimized model will use during inference.
+target_platform_capabilities
+==============================
+- :ref:`target_platform_capabilities<ug-target_platform_capabilities>`: Module to create and model hardware-related settings to optimize the model according to, by the hardware the optimized model will use during inference.
 - :ref:`get_target_platform_capabilities<ug-get_target_platform_capabilities>`: A function to get a target platform model for Tensorflow and Pytorch.
 - :ref:`DefaultDict<ug-DefaultDict>`: Util class for creating a TargetPlatformCapabilities.
 

docsrc/source/api/api_docs/methods/get_target_platform_capabilities.rst

@@ -4,7 +4,7 @@
 
 
 =======================================
-Get TargetPlatformCapabilities
+Get FrameworkQuantizationCapabilities
 =======================================
 
 .. autofunction:: model_compression_toolkit.get_target_platform_capabilities

docsrc/source/api/api_docs/modules/layer_filters.rst

@@ -15,30 +15,30 @@ one may use the next filters to check if a layer configuration holds the created
 Attribute Filters
 ==================
 
-.. autoclass:: model_compression_toolkit.target_platform.AttributeFilter
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.AttributeFilter
 
 |
 
-.. autoclass:: model_compression_toolkit.target_platform.Eq
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.Eq
 
 |
 
-.. autoclass:: model_compression_toolkit.target_platform.NotEq
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.NotEq
 
 |
 
-.. autoclass:: model_compression_toolkit.target_platform.Greater
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.Greater
 
 |
 
 
-.. autoclass:: model_compression_toolkit.target_platform.GreaterEq
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.GreaterEq
 
 |
 
 
-.. autoclass:: model_compression_toolkit.target_platform.Smaller
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.Smaller
 
 |
 
-.. autoclass:: model_compression_toolkit.target_platform.SmallerEq
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.SmallerEq

docsrc/source/api/api_docs/modules/target_platform.rst → docsrc/source/api/api_docs/modules/target_platform_capabilities.rst

@@ -1,11 +1,11 @@
 :orphan:
 
-.. _ug-target_platform:
+.. _ug-target_platform_capabilities:
 
 
-=================================
-target_platform Module
-=================================
+=====================================
+target_platform_capabilities Module
+=====================================
 
 MCT can be configured to quantize and optimize models for different hardware settings.
 For example, when using qnnpack backend for Pytorch model inference, Pytorch `quantization
@@ -14,7 +14,7 @@ uses `per-tensor weights quantization <https://github.com/pytorch/pytorch/blob/m
 for Conv2d, while when using tflite modeling, Tensorflow uses `per-channel weights quantization for
 Conv2D <https://www.tensorflow.org/lite/performance/quantization_spec#per-axis_vs_per-tensor>`_.
 
-This can be addressed in MCT by using the target_platform module, that can configure different
+This can be addressed in MCT by using the target_platform_capabilities module, that can configure different
 parameters that are hardware-related, and the optimization process will use this to optimize the model accordingly.
 Models for IMX500, TFLite and qnnpack can be observed `here <https://github.com/sony/model_optimization/tree/main/model_compression_toolkit/target_platform_capabilities>`_, and can be used using :ref:`get_target_platform_capabilities function<ug-get_target_platform_capabilities>`.
 
@@ -36,68 +36,43 @@ Now, we will detail about the different components.
 
 QuantizationMethod
 ==========================
-.. autoclass:: model_compression_toolkit.target_platform.QuantizationMethod
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.QuantizationMethod
 
 
 
 OpQuantizationConfig
 ======================
-.. autoclass:: model_compression_toolkit.target_platform.OpQuantizationConfig
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema.OpQuantizationConfig
 
 
 
 AttributeQuantizationConfig
 ============================
-.. autoclass:: model_compression_toolkit.target_platform.AttributeQuantizationConfig
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema.AttributeQuantizationConfig
 
 
 QuantizationConfigOptions
 ============================
-.. autoclass:: model_compression_toolkit.target_platform.QuantizationConfigOptions
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema.QuantizationConfigOptions
 
 
-TargetPlatformModel
-=======================
-.. autoclass:: model_compression_toolkit.target_platform.TargetPlatformModel
+TargetPlatformCapabilities
+============================
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema.TargetPlatformCapabilities
 
 
 OperatorsSet
 ================
-.. autoclass:: model_compression_toolkit.target_platform.OperatorsSet
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema.OperatorsSet
 
 
 
 Fusing
 ==============
-.. autoclass:: model_compression_toolkit.target_platform.Fusing
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema.Fusing
 
 
 
-OperatorSetConcat
+OperatorSetGroup
 ====================
-.. autoclass:: model_compression_toolkit.target_platform.OperatorSetConcat
-
-
-OperationsToLayers
-=====================
-.. autoclass:: model_compression_toolkit.target_platform.OperationsToLayers
-
-
-OperationsSetToLayers
-=========================
-.. autoclass:: model_compression_toolkit.target_platform.OperationsSetToLayers
-
-
-LayerFilterParams
-=========================
-.. autoclass:: model_compression_toolkit.target_platform.LayerFilterParams
-
-More filters and usage examples are detailed :ref:`here<ug-layer_filters>`.
-
-
-TargetPlatformCapabilities
-=============================
-.. autoclass:: model_compression_toolkit.target_platform.TargetPlatformCapabilities
-
-
-
+.. autoclass:: model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema.OperatorSetGroup

docsrc/source/api/api_docs/modules/trainable_infrastructure.rst

@@ -55,7 +55,7 @@ For example, we can set a trainable weights quantizer with the following configu
 
 .. code-block:: python
 
-    from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
+    from model_compression_toolkit.target_platform_capabilities.target_platform_capabilities import QuantizationMethod
     from model_compression_toolkit.constants import THRESHOLD, MIN_THRESHOLD
 
     TrainableQuantizerWeightsConfig(weights_quantization_method=QuantizationMethod.SYMMETRIC,
@@ -79,7 +79,7 @@ For example, we can set a trainable activation quantizer with the following conf
 
 .. code-block:: python
 
-    from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
+    from model_compression_toolkit.target_platform_capabilities.target_platform_capabilities import QuantizationMethod
     from model_compression_toolkit.constants import THRESHOLD, MIN_THRESHOLD
 
     TrainableQuantizerActivationConfig(activation_quantization_method=QuantizationMethod.UNIFORM,

docsrc/source/api/api_docs/notes/tpc_note.rst

@@ -1,7 +1,7 @@
 
 .. note::
-   For now, some fields of :class:`~model_compression_toolkit.target_platform.OpQuantizationConfig` are ignored during
+   For now, some fields of :class:`~model_compression_toolkit.target_platform_capabilities.OpQuantizationConfig` are ignored during
    the optimization process such as quantization_preserving, fixed_scale, and fixed_zero_point.
 
-   - MCT will use more information from :class:`~model_compression_toolkit.target_platform.OpQuantizationConfig`, in the future.
+   - MCT will use more information from :class:`~model_compression_toolkit.target_platform_capabilities.OpQuantizationConfig`, in the future.
 

model_compression_toolkit/__init__.py

@@ -14,7 +14,7 @@
 # ==============================================================================
 
 from model_compression_toolkit.defaultdict import DefaultDict
-from model_compression_toolkit.target_platform_capabilities import target_platform
+from model_compression_toolkit import target_platform_capabilities
 from model_compression_toolkit.target_platform_capabilities.tpc_models.get_target_platform_capabilities import get_target_platform_capabilities
 from model_compression_toolkit import core
 from model_compression_toolkit.logger import set_log_folder

model_compression_toolkit/core/common/framework_info.py

@@ -18,10 +18,8 @@
 from enum import Enum
 from typing import Dict, Any, List
 
-
-
+from mct_quantizers import QuantizationMethod
 from model_compression_toolkit.defaultdict import DefaultDict
-from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
 
 
 class ChannelAxis(Enum):

model_compression_toolkit/core/common/fusion/layer_fusing.py

@@ -16,8 +16,9 @@
 from typing import Any, List
 from model_compression_toolkit.core.common.graph.base_graph import Graph
 from model_compression_toolkit.core.common.graph.base_node import BaseNode
-from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework import TargetPlatformCapabilities
-from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework.layer_filter_params import LayerFilterParams
+from model_compression_toolkit.target_platform_capabilities.targetplatform2framework.framework_quantization_capabilities import \
+    FrameworkQuantizationCapabilities
+from model_compression_toolkit.target_platform_capabilities.targetplatform2framework.layer_filter_params import LayerFilterParams
 
 
 def filter_fusing_patterns(fusing_patterns: List[List[Any]], node: BaseNode, idx: int = 0) -> List[List[Any]]:
@@ -77,18 +78,18 @@ def disable_nodes_activation_quantization(nodes: List[BaseNode]):
             qc.activation_quantization_cfg.enable_activation_quantization = False
 
 
-def fusion(graph: Graph, tpc: TargetPlatformCapabilities) -> Graph:
+def fusion(graph: Graph, fqc: FrameworkQuantizationCapabilities) -> Graph:
     """
     Fusing defines a list of operators that should be combined and treated as a single operator,
     hence no quantization is applied between them when they appear in the graph.
     This function search and disable quantization for such patterns.
     Args:
         graph: Graph we apply the fusion on.
-        tpc: TargetPlatformCapabilities object that describes the desired inference target platform (includes fusing patterns MCT should handle).
+        fqc: FrameworkQuantizationCapabilities object that describes the desired inference target platform (includes fusing patterns MCT should handle).
     Returns:
         Graph after applying fusion activation marking.
     """
-    fusing_patterns = tpc.get_fusing_patterns()
+    fusing_patterns = fqc.get_fusing_patterns()
     if len(fusing_patterns) == 0:
         return graph
 

model_compression_toolkit/core/common/graph/base_graph.py

@@ -32,8 +32,9 @@
 from model_compression_toolkit.core.common.pruning.pruning_section import PruningSection
 from model_compression_toolkit.core.common.user_info import UserInformation
 from model_compression_toolkit.logger import Logger
-from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework import \
-    TargetPlatformCapabilities, LayerFilterParams
+from model_compression_toolkit.target_platform_capabilities.targetplatform2framework import LayerFilterParams
+from model_compression_toolkit.target_platform_capabilities.targetplatform2framework.framework_quantization_capabilities import \
+    FrameworkQuantizationCapabilities
 
 OutTensor = namedtuple('OutTensor', 'node node_out_index')
 
@@ -86,29 +87,29 @@ def set_fw_info(self,
 
         self.fw_info = fw_info
 
-    def set_tpc(self,
-                tpc: TargetPlatformCapabilities):
+    def set_fqc(self,
+                fqc: FrameworkQuantizationCapabilities):
         """
-        Set the graph's TPC.
+        Set the graph's FQC.
         Args:
-            tpc: TargetPlatformCapabilities object.
+            fqc: FrameworkQuantizationCapabilities object.
         """
-        # validate graph nodes are either from the framework or a custom layer defined in the TPC
-        # Validate graph nodes are either built-in layers from the framework or custom layers defined in the TPC
-        tpc_layers = tpc.op_sets_to_layers.get_layers()
-        tpc_filtered_layers = [layer for layer in tpc_layers if isinstance(layer, LayerFilterParams)]
+        # validate graph nodes are either from the framework or a custom layer defined in the FQC
+        # Validate graph nodes are either built-in layers from the framework or custom layers defined in the FQC
+        fqc_layers = fqc.op_sets_to_layers.get_layers()
+        fqc_filtered_layers = [layer for layer in fqc_layers if isinstance(layer, LayerFilterParams)]
         for n in self.nodes:
-            is_node_in_tpc = any([n.is_match_type(_type) for _type in tpc_layers]) or \
-                             any([n.is_match_filter_params(filtered_layer) for filtered_layer in tpc_filtered_layers])
+            is_node_in_fqc = any([n.is_match_type(_type) for _type in fqc_layers]) or \
+                             any([n.is_match_filter_params(filtered_layer) for filtered_layer in fqc_filtered_layers])
             if n.is_custom:
-                if not is_node_in_tpc:
+                if not is_node_in_fqc:
                     Logger.critical(f'MCT does not support optimizing Keras custom layers. Found a layer of type {n.type}. '
-                                    ' Please add the custom layer to Target Platform Capabilities (TPC), or file a feature '
+                                    ' Please add the custom layer to Framework Quantization Capabilities (FQC), or file a feature '
                                     'request or an issue if you believe this should be supported.')  # pragma: no cover
-                if any([qc.default_weight_attr_config.enable_weights_quantization for qc in n.get_qco(tpc).quantization_configurations]):
+                if any([qc.default_weight_attr_config.enable_weights_quantization for qc in n.get_qco(fqc).quantization_configurations]):
                     Logger.critical(f'Layer identified: {n.type}. MCT does not support weight quantization for Keras custom layers.')  # pragma: no cover
 
-        self.tpc = tpc
+        self.fqc = fqc
 
     def get_topo_sorted_nodes(self):
         """