Refactor Target Platform Capabilities Design (#1276)

* Refactor Target Platform Capabilities Design

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

Additional Changes:
- 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.
- 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
- Remove unused files

---------

Co-authored-by: liord <lior.dikstein@altair-semi.com>

model_compression_toolkit/constants.py

@@ -18,9 +18,6 @@
 TENSORFLOW = 'tensorflow'
 PYTORCH = 'pytorch'
 
-# Metadata fields
-MCT_VERSION = 'mct_version'
-TPC_VERSION = 'tpc_version'
 
 WEIGHTS_SIGNED = True
 # Minimal threshold to use for quantization ranges:

model_compression_toolkit/core/common/graph/base_node.py

@@ -22,8 +22,9 @@
     ACTIVATION_N_BITS_ATTRIBUTE, FP32_BYTES_PER_PARAMETER
 from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig
 from model_compression_toolkit.logger import Logger
-from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationConfigOptions, \
-    TargetPlatformCapabilities, LayerFilterParams, OpQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import QuantizationConfigOptions, \
+    OpQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.target_platform import TargetPlatformCapabilities, LayerFilterParams
 
 
 class BaseNode:

model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_data.py

@@ -23,8 +23,8 @@
 from model_compression_toolkit.core.common.framework_implementation import FrameworkImplementation
 from model_compression_toolkit.core.common.graph.edge import EDGE_SINK_INDEX
 from model_compression_toolkit.core.graph_prep_runner import graph_preparation_runner
-from model_compression_toolkit.target_platform_capabilities.target_platform import TargetPlatformCapabilities, \
-    QuantizationConfigOptions
+from model_compression_toolkit.target_platform_capabilities.target_platform import TargetPlatformCapabilities
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import QuantizationConfigOptions
 
 
 def compute_resource_utilization_data(in_model: Any,

model_compression_toolkit/core/common/quantization/candidate_node_quantization_config.py

@@ -17,8 +17,8 @@
 from model_compression_toolkit.core import QuantizationConfig
 from model_compression_toolkit.core.common.quantization.node_quantization_config import BaseNodeQuantizationConfig, \
     NodeWeightsQuantizationConfig, NodeActivationQuantizationConfig
-from model_compression_toolkit.target_platform_capabilities.target_platform import OpQuantizationConfig, \
-    AttributeQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import AttributeQuantizationConfig, \
+    OpQuantizationConfig
 from model_compression_toolkit.logger import Logger
 
 

model_compression_toolkit/core/common/quantization/node_quantization_config.py

@@ -25,8 +25,8 @@
 
 from model_compression_toolkit.core.common.quantization.quantization_config import QuantizationConfig, \
     QuantizationErrorMethod
-from model_compression_toolkit.target_platform_capabilities.target_platform import OpQuantizationConfig, \
-    AttributeQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import AttributeQuantizationConfig, \
+    OpQuantizationConfig
 
 
 ##########################################

model_compression_toolkit/core/common/quantization/quantization_params_generation/qparams_activations_computation.py

@@ -15,7 +15,8 @@
 import numpy as np
 from typing import Dict, Union
 
-from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod, Signedness
+from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import Signedness
 from model_compression_toolkit.core.common.collectors.statistics_collector import BaseStatsCollector
 from model_compression_toolkit.core.common.quantization import quantization_params_generation
 from model_compression_toolkit.core.common.node_prior_info import NodePriorInfo

model_compression_toolkit/core/common/quantization/set_node_quantization_config.py

@@ -33,7 +33,7 @@
 from model_compression_toolkit.core.common.quantization.quantization_fn_selection import \
     get_weights_quantization_fn
 from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework import TargetPlatformCapabilities
-from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import OpQuantizationConfig, \
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import OpQuantizationConfig, \
     QuantizationConfigOptions
 
 

model_compression_toolkit/core/common/statistics_correction/apply_activation_bias_correction_to_graph.py

@@ -17,7 +17,7 @@
 from model_compression_toolkit.core.common import BaseNode, Graph
 from model_compression_toolkit.core.common.framework_implementation import FrameworkImplementation
 from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig
-from model_compression_toolkit.target_platform_capabilities.target_platform import AttributeQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import AttributeQuantizationConfig
 
 
 def apply_activation_bias_correction_to_graph(graph: Graph,

model_compression_toolkit/core/common/statistics_correction/apply_bias_correction_to_graph.py

@@ -19,7 +19,7 @@
 from model_compression_toolkit.core.common import Graph, BaseNode
 from model_compression_toolkit.core.common.framework_implementation import FrameworkImplementation
 from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig
-from model_compression_toolkit.target_platform_capabilities.target_platform import AttributeQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import AttributeQuantizationConfig
 
 
 def apply_bias_correction_to_graph(graph_to_apply_bias_correction: Graph,

model_compression_toolkit/core/common/substitutions/batchnorm_reconstruction.py

@@ -26,8 +26,8 @@
 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.core.common.graph.graph_matchers import NodeOperationMatcher
-from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod, \
-    AttributeQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import AttributeQuantizationConfig
 
 
 class BatchNormalizationReconstruction(common.BaseSubstitution):

model_compression_toolkit/core/common/substitutions/shift_negative_activation.py

@@ -22,8 +22,8 @@
 from model_compression_toolkit.core.common import FrameworkInfo, Graph, BaseNode
 from model_compression_toolkit.constants import THRESHOLD, SIGNED, SHIFT_NEGATIVE_NON_LINEAR_NUM_BITS
 from model_compression_toolkit.core.common.graph.graph_matchers import NodeOperationMatcher
-from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod, \
-    AttributeQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import AttributeQuantizationConfig
 from model_compression_toolkit.core.common.quantization.set_node_quantization_config import create_node_activation_qc, \
     set_quantization_configs_to_node
 from model_compression_toolkit.core.common.quantization.core_config import CoreConfig

model_compression_toolkit/metadata.py

@@ -12,10 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+from dataclasses import dataclass, asdict
 
 from typing import Dict, Any
-from model_compression_toolkit.constants import MCT_VERSION, TPC_VERSION, OPERATORS_SCHEDULING, FUSED_NODES_MAPPING, \
-    CUTS, MAX_CUT, OP_ORDER, OP_RECORD, SHAPE, NODE_OUTPUT_INDEX, NODE_NAME, TOTAL_SIZE, MEM_ELEMENTS
+from model_compression_toolkit.constants import OPERATORS_SCHEDULING, FUSED_NODES_MAPPING, CUTS, MAX_CUT, OP_ORDER, \
+    OP_RECORD, SHAPE, NODE_OUTPUT_INDEX, NODE_NAME, TOTAL_SIZE, MEM_ELEMENTS
 from model_compression_toolkit.core.common.graph.memory_graph.compute_graph_max_cut import SchedulerInfo
 from model_compression_toolkit.target_platform_capabilities.target_platform import TargetPlatformCapabilities
 
@@ -43,13 +44,21 @@ def create_model_metadata(tpc: TargetPlatformCapabilities,
 def get_versions_dict(tpc) -> Dict:
     """
 
-    Returns: A dictionary with TPC and MCT versions.
+    Returns: A dictionary with TPC, MCT and TPC-Schema versions.
 
     """
     # imported inside to avoid circular import error
     from model_compression_toolkit import __version__ as mct_version
-    tpc_version = f'{tpc.name}.{tpc.version}'
-    return {MCT_VERSION: mct_version, TPC_VERSION: tpc_version}
+
+    @dataclass
+    class TPCVersions:
+        mct_version: str
+        tpc_minor_version: str = f'{tpc.tp_model.tpc_minor_version}'
+        tpc_patch_version: str = f'{tpc.tp_model.tpc_patch_version}'
+        tpc_platform_type: str = f'{tpc.tp_model.tpc_platform_type}'
+        tpc_schema: str = f'{tpc.tp_model.SCHEMA_VERSION}'
+
+    return asdict(TPCVersions(mct_version))
 
 
 def get_scheduler_metadata(scheduler_info: SchedulerInfo) -> Dict[str, Any]:

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`
 selects a quantizer for weights and a quantizer for activations.
 
 Currently, only the STE (straight through estimator) training method is implemented by the MCT.

model_compression_toolkit/qat/pytorch/quantizer/README.md

@@ -5,7 +5,7 @@
 Several training methods may be applied by the user to train the QAT ready model
 created by `pytorch_quantization_aware_training_init` method in [`pytorch/quantization_facade`](../quantization_facade.py).
 Each [`TrainingMethod`](../../../trainable_infrastructure/common/training_method.py) 
-and [`QuantizationMethod`](../../../target_platform_capabilities/target_platform/op_quantization_config.py)
+and `QuantizationMethod`
 selects a quantizer for weights and a quantizer for activations.
 
 ## Make your own training method

model_compression_toolkit/target_platform_capabilities/schema/__init__.py

@@ -0,0 +1,14 @@
+# Copyright 2024 Sony Semiconductor Israel, Inc. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================

model_compression_toolkit/target_platform_capabilities/schema/mct_current_schema.py

@@ -0,0 +1,11 @@
+import model_compression_toolkit.target_platform_capabilities.schema.v1 as schema
+
+Signedness = schema.Signedness
+AttributeQuantizationConfig = schema.AttributeQuantizationConfig
+OpQuantizationConfig = schema.OpQuantizationConfig
+QuantizationConfigOptions = schema.QuantizationConfigOptions
+OperatorsSetBase = schema.OperatorsSetBase
+OperatorsSet = schema.OperatorsSet
+OperatorSetConcat= schema.OperatorSetConcat
+Fusing = schema.Fusing
+TargetPlatformModel = schema.TargetPlatformModel

model_compression_toolkit/target_platform_capabilities/schema/schema_functions.py

@@ -0,0 +1,37 @@
+# Copyright 2024 Sony Semiconductor Israel, Inc. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+import copy
+from typing import Any, Dict
+
+
+def clone_and_edit_object_params(obj: Any, **kwargs: Dict) -> Any:
+    """
+    Clones the given object and edit some of its parameters.
+
+    Args:
+        obj: An object to clone.
+        **kwargs: Keyword arguments to edit in the cloned object.
+
+    Returns:
+        Edited copy of the given object.
+    """
+
+    obj_copy = copy.deepcopy(obj)
+    for k, v in kwargs.items():
+        assert hasattr(obj_copy,
+                       k), f'Edit parameter is possible only for existing parameters in the given object, ' \
+                           f'but {k} is not a parameter of {obj_copy}.'
+        setattr(obj_copy, k, v)
+    return obj_copy