mct-nightly 2.2.0.20241222.533__py3-none-any.whl → 2.2.0.20241223.525__py3-none-any.whl

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v3_lut/tp_model.py

@@ -19,7 +19,8 @@ import model_compression_toolkit.target_platform_capabilities.schema.mct_current
 from model_compression_toolkit.constants import FLOAT_BITWIDTH
 from model_compression_toolkit.target_platform_capabilities.constants import KERNEL_ATTR, BIAS_ATTR, WEIGHTS_N_BITS, \
     WEIGHTS_QUANTIZATION_METHOD, IMX500_TP_MODEL
-from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import TargetPlatformModel, Signedness, \
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import TargetPlatformModel, \
+    Signedness, \
     AttributeQuantizationConfig, OpQuantizationConfig
 
 tp = mct.target_platform
@@ -152,7 +153,7 @@ def generate_tp_model(default_config: OpQuantizationConfig,
     # of possible configurations to consider when quantizing a set of operations (in mixed-precision, for example).
     # If the QuantizationConfigOptions contains only one configuration,
     # this configuration will be used for the operation quantization:
-    default_configuration_options = schema.QuantizationConfigOptions([default_config])
+    default_configuration_options = schema.QuantizationConfigOptions(tuple([default_config]))
 
     # Create a QuantizationConfigOptions for quantizing constants in functional ops.
     # Constant configuration is similar to the default eight bit configuration except for PoT
@@ -163,7 +164,55 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         default_weight_attr_config=default_config.default_weight_attr_config.clone_and_edit(
             enable_weights_quantization=True, weights_per_channel_threshold=True,
             weights_quantization_method=tp.QuantizationMethod.POWER_OF_TWO))
-    const_configuration_options = schema.QuantizationConfigOptions([const_config])
+    const_configuration_options = schema.QuantizationConfigOptions(tuple([const_config]))
+
+    # Create Mixed-Precision quantization configuration options from the given list of OpQuantizationConfig objects
+    mixed_precision_configuration_options = schema.QuantizationConfigOptions(tuple(mixed_precision_cfg_list),
+                                                                             base_config=base_config)
+
+    # Create an OperatorsSet to represent a set of operations.
+    # Each OperatorsSet has a unique label.
+    # If a quantization configuration options is passed, these options will
+    # be used for operations that will be attached to this set's label.
+    # Otherwise, it will be a configure-less set (used in fusing):
+    operator_set = []
+    fusing_patterns = []
+    # May suit for operations like: Dropout, Reshape, etc.
+    operator_set.append(schema.OperatorsSet("NoQuantization",
+                                            default_configuration_options.clone_and_edit(
+                                                enable_activation_quantization=False)
+                                            .clone_and_edit_weight_attribute(enable_weights_quantization=False)))
+
+    # Define operator sets that use mixed_precision_configuration_options:
+    conv = schema.OperatorsSet("Conv", mixed_precision_configuration_options)
+    fc = schema.OperatorsSet("FullyConnected", mixed_precision_configuration_options)
+
+    # Define operations sets without quantization configuration
+    # options (useful for creating fusing patterns, for example):
+    any_relu = schema.OperatorsSet("AnyReLU")
+    add = schema.OperatorsSet("Add", const_configuration_options)
+    sub = schema.OperatorsSet("Sub", const_configuration_options)
+    mul = schema.OperatorsSet("Mul", const_configuration_options)
+    div = schema.OperatorsSet("Div", const_configuration_options)
+    prelu = schema.OperatorsSet("PReLU")
+    swish = schema.OperatorsSet("Swish")
+    sigmoid = schema.OperatorsSet("Sigmoid")
+    tanh = schema.OperatorsSet("Tanh")
+
+    operator_set.extend([conv, fc, any_relu, add, sub, mul, div, prelu, swish, sigmoid, tanh])
+    # Combine multiple operators into a single operator to avoid quantization between
+    # them. To do this we define fusing patterns using the OperatorsSets that were created.
+    # To group multiple sets with regard to fusing, an OperatorSetConcat can be created
+    activations_after_conv_to_fuse = schema.OperatorSetConcat([any_relu, swish, prelu, sigmoid, tanh])
+    activations_after_fc_to_fuse = schema.OperatorSetConcat([any_relu, swish, sigmoid])
+    any_binary = schema.OperatorSetConcat([add, sub, mul, div])
+
+    # ------------------- #
+    # Fusions
+    # ------------------- #
+    fusing_patterns.append(schema.Fusing((conv, activations_after_conv_to_fuse)))
+    fusing_patterns.append(schema.Fusing((fc, activations_after_fc_to_fuse)))
+    fusing_patterns.append(schema.Fusing((any_binary, any_relu)))
 
     # Create a TargetPlatformModel and set its default quantization config.
     # This default configuration will be used for all operations
@@ -173,56 +222,9 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         tpc_minor_version=3,
         tpc_patch_version=0,
         tpc_platform_type=IMX500_TP_MODEL,
+        operator_set=tuple(operator_set),
+        fusing_patterns=tuple(fusing_patterns),
         add_metadata=True,
         name=name)
 
-    # To start defining the model's components (such as operator sets, and fusing patterns),
-    # use 'with' the TargetPlatformModel instance, and create them as below:
-    with generated_tpm:
-        # Create an OperatorsSet to represent a set of operations.
-        # Each OperatorsSet has a unique label.
-        # If a quantization configuration options is passed, these options will
-        # be used for operations that will be attached to this set's label.
-        # Otherwise, it will be a configure-less set (used in fusing):
-
-        # May suit for operations like: Dropout, Reshape, etc.
-        default_qco = tp.get_default_quantization_config_options()
-        schema.OperatorsSet("NoQuantization",
-                               default_qco.clone_and_edit(enable_activation_quantization=False)
-                               .clone_and_edit_weight_attribute(enable_weights_quantization=False))
-
-        # Create Mixed-Precision quantization configuration options from the given list of OpQuantizationConfig objects
-        mixed_precision_configuration_options = schema.QuantizationConfigOptions(mixed_precision_cfg_list,
-                                                                                    base_config=base_config)
-
-        # Define operator sets that use mixed_precision_configuration_options:
-        conv = schema.OperatorsSet("Conv", mixed_precision_configuration_options)
-        fc = schema.OperatorsSet("FullyConnected", mixed_precision_configuration_options)
-
-        # Define operations sets without quantization configuration
-        # options (useful for creating fusing patterns, for example):
-        any_relu = schema.OperatorsSet("AnyReLU")
-        add = schema.OperatorsSet("Add", const_configuration_options)
-        sub = schema.OperatorsSet("Sub", const_configuration_options)
-        mul = schema.OperatorsSet("Mul", const_configuration_options)
-        div = schema.OperatorsSet("Div", const_configuration_options)
-        prelu = schema.OperatorsSet("PReLU")
-        swish = schema.OperatorsSet("Swish")
-        sigmoid = schema.OperatorsSet("Sigmoid")
-        tanh = schema.OperatorsSet("Tanh")
-
-        # Combine multiple operators into a single operator to avoid quantization between
-        # them. To do this we define fusing patterns using the OperatorsSets that were created.
-        # To group multiple sets with regard to fusing, an OperatorSetConcat can be created
-        activations_after_conv_to_fuse = schema.OperatorSetConcat([any_relu, swish, prelu, sigmoid, tanh])
-        activations_after_fc_to_fuse = schema.OperatorSetConcat([any_relu, swish, sigmoid])
-        any_binary = schema.OperatorSetConcat([add, sub, mul, div])
-
-        # ------------------- #
-        # Fusions
-        # ------------------- #
-        schema.Fusing([conv, activations_after_conv_to_fuse])
-        schema.Fusing([fc, activations_after_fc_to_fuse])
-        schema.Fusing([any_binary, any_relu])
-
     return generated_tpm

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v4/tp_model.py

@@ -19,7 +19,8 @@ import model_compression_toolkit.target_platform_capabilities.schema.v1 as schem
 from model_compression_toolkit.constants import FLOAT_BITWIDTH
 from model_compression_toolkit.target_platform_capabilities.constants import KERNEL_ATTR, BIAS_ATTR, WEIGHTS_N_BITS, \
     IMX500_TP_MODEL
-from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import TargetPlatformModel, Signedness, \
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import TargetPlatformModel, \
+    Signedness, \
     AttributeQuantizationConfig, OpQuantizationConfig
 
 tp = mct.target_platform
@@ -87,7 +88,8 @@ def get_op_quantization_configs() -> \
         weights_quantization_method=tp.QuantizationMethod.POWER_OF_TWO,
         weights_n_bits=8,
         weights_per_channel_threshold=False,
-        enable_weights_quantization=False,  # TODO: this will changed to True once implementing multi-attributes quantization
+        enable_weights_quantization=False,
+        # TODO: this will changed to True once implementing multi-attributes quantization
         lut_values_bitwidth=None)
 
     # define a quantization config to quantize the kernel (for layers where there is a kernel attribute).
@@ -176,13 +178,13 @@ def generate_tp_model(default_config: OpQuantizationConfig,
     # of possible configurations to consider when quantizing a set of operations (in mixed-precision, for example).
     # If the QuantizationConfigOptions contains only one configuration,
     # this configuration will be used for the operation quantization:
-    default_configuration_options = schema.QuantizationConfigOptions([default_config])
+    default_configuration_options = schema.QuantizationConfigOptions(tuple([default_config]))
     default_config_input16 = default_config.clone_and_edit(supported_input_activation_n_bits=(8, 16))
-    default_config_options_16bit = schema.QuantizationConfigOptions([default_config_input16,
-                                                                        default_config_input16.clone_and_edit(
-                                                                            activation_n_bits=16,
-                                                                            signedness=Signedness.SIGNED)],
-                                                                       base_config=default_config_input16)
+    default_config_options_16bit = schema.QuantizationConfigOptions(tuple([default_config_input16,
+                                                                     default_config_input16.clone_and_edit(
+                                                                         activation_n_bits=16,
+                                                                         signedness=Signedness.SIGNED)]),
+                                                                    base_config=default_config_input16)
 
     # Create a QuantizationConfigOptions for quantizing constants in functional ops.
     # Constant configuration is similar to the default eight bit configuration except for PoT
@@ -193,7 +195,7 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         default_weight_attr_config=default_config.default_weight_attr_config.clone_and_edit(
             enable_weights_quantization=True, weights_per_channel_threshold=True,
             weights_quantization_method=tp.QuantizationMethod.POWER_OF_TWO))
-    const_configuration_options = schema.QuantizationConfigOptions([const_config])
+    const_configuration_options = schema.QuantizationConfigOptions(tuple([const_config]))
 
     # 16 bits inputs and outputs. Currently, only defined for consts since they are used in operators that
     # support 16 bit as input and output.
@@ -201,9 +203,9 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         supported_input_activation_n_bits=(8, 16))
     const_config_input16_output16 = const_config_input16.clone_and_edit(
         activation_n_bits=16, signedness=Signedness.SIGNED)
-    const_configuration_options_inout16 = schema.QuantizationConfigOptions([const_config_input16_output16,
-                                                                               const_config_input16],
-                                                                                                                                     base_config=const_config_input16)
+    const_configuration_options_inout16 = schema.QuantizationConfigOptions(tuple([const_config_input16_output16,
+                                                                            const_config_input16]),
+                                                                           base_config=const_config_input16)
 
     const_config_input16_per_tensor = const_config.clone_and_edit(
         supported_input_activation_n_bits=(8, 16),
@@ -213,20 +215,91 @@ def generate_tp_model(default_config: OpQuantizationConfig,
     )
     const_config_input16_output16_per_tensor = const_config_input16_per_tensor.clone_and_edit(
         activation_n_bits=16, signedness=Signedness.SIGNED)
-    const_configuration_options_inout16_per_tensor = schema.QuantizationConfigOptions(
+    const_configuration_options_inout16_per_tensor = schema.QuantizationConfigOptions(tuple(
         [const_config_input16_output16_per_tensor,
-         const_config_input16_per_tensor],
+         const_config_input16_per_tensor]),
         base_config=const_config_input16_per_tensor)
 
     qpreserving_const_config = const_config.clone_and_edit(enable_activation_quantization=False,
                                                            quantization_preserving=True,
                                                            default_weight_attr_config=const_config.default_weight_attr_config.clone_and_edit(
                                                                weights_per_channel_threshold=False))
-    qpreserving_const_config_options = schema.QuantizationConfigOptions([qpreserving_const_config])
+    qpreserving_const_config_options = schema.QuantizationConfigOptions(tuple([qpreserving_const_config]))
 
     mp_cfg_list_16bit = [mp_cfg.clone_and_edit(activation_n_bits=16, signedness=Signedness.SIGNED)
                          for mp_cfg in mixed_precision_cfg_list]
 
+    # Create Mixed-Precision quantization configuration options from the given list of OpQuantizationConfig objects
+    mixed_precision_configuration_options = schema.QuantizationConfigOptions(tuple(
+        mixed_precision_cfg_list + mp_cfg_list_16bit),
+        base_config=base_config)
+
+    # Create an OperatorsSet to represent a set of operations.
+    # Each OperatorsSet has a unique label.
+    # If a quantization configuration options is passed, these options will
+    # be used for operations that will be attached to this set's label.
+    # Otherwise, it will be a configure-less set (used in fusing):
+    operator_set = []
+    fusing_patterns = []
+    # May suit for operations like: Dropout, Reshape, etc.
+    operator_set.append(schema.OperatorsSet(OPSET_NO_QUANTIZATION,
+                                            default_configuration_options.clone_and_edit(
+                                                enable_activation_quantization=False)
+                                            .clone_and_edit_weight_attribute(enable_weights_quantization=False)))
+    operator_set.append(schema.OperatorsSet(OPSET_QUANTIZATION_PRESERVING,
+                                            default_configuration_options.clone_and_edit(
+                                                enable_activation_quantization=False,
+                                                quantization_preserving=True)
+                                            .clone_and_edit_weight_attribute(enable_weights_quantization=False)))
+    operator_set.append(
+        schema.OperatorsSet(OPSET_DIMENSION_MANIPULATION_OPS_WITH_WEIGHTS, qpreserving_const_config_options))
+    operator_set.append(schema.OperatorsSet(OPSET_DIMENSION_MANIPULATION_OPS,
+                                            default_configuration_options.clone_and_edit(
+                                                enable_activation_quantization=False,
+                                                quantization_preserving=True,
+                                                supported_input_activation_n_bits=(8, 16))
+                                            .clone_and_edit_weight_attribute(enable_weights_quantization=False)))
+    operator_set.append(schema.OperatorsSet(OPSET_MERGE_OPS, const_configuration_options_inout16_per_tensor))
+
+    # Define operator sets that use mixed_precision_configuration_options:
+    conv = schema.OperatorsSet(OPSET_CONV, mixed_precision_configuration_options)
+    fc = schema.OperatorsSet(OPSET_FULLY_CONNECTED, mixed_precision_configuration_options)
+
+    operator_set.append(schema.OperatorsSet(OPSET_BATCH_NORM, default_config_options_16bit))
+
+    # Note: Operations sets without quantization configuration are useful for creating fusing patterns
+    any_relu = schema.OperatorsSet(OPSET_ANY_RELU, default_config_options_16bit)
+    add = schema.OperatorsSet(OPSET_ADD, const_configuration_options_inout16)
+    sub = schema.OperatorsSet(OPSET_SUB, const_configuration_options_inout16)
+    mul = schema.OperatorsSet(OPSET_MUL, const_configuration_options_inout16)
+    div = schema.OperatorsSet(OPSET_DIV, const_configuration_options)
+    min_max = schema.OperatorsSet(OPSET_MIN_MAX, const_configuration_options_inout16)
+    prelu = schema.OperatorsSet(OPSET_PRELU, default_config_options_16bit)
+    swish = schema.OperatorsSet(OPSET_SWISH, default_config_options_16bit)
+    sigmoid = schema.OperatorsSet(OPSET_SIGMOID, default_config_options_16bit)
+    tanh = schema.OperatorsSet(OPSET_TANH, default_config_options_16bit)
+    gelu = schema.OperatorsSet(OPSET_GELU, default_config_options_16bit)
+    hardsigmoid = schema.OperatorsSet(OPSET_HARDSIGMOID, default_config_options_16bit)
+    hardswish = schema.OperatorsSet(OPSET_HARDSWISH, default_config_options_16bit)
+
+    operator_set.extend(
+        [conv, fc, any_relu, add, sub, mul, div, prelu, swish, sigmoid, tanh, min_max, gelu, hardsigmoid, hardswish])
+    # Combine multiple operators into a single operator to avoid quantization between
+    # them. To do this we define fusing patterns using the OperatorsSets that were created.
+    # To group multiple sets with regard to fusing, an OperatorSetConcat can be created
+    activations_after_conv_to_fuse = schema.OperatorSetConcat([any_relu, swish, prelu, sigmoid,
+                                                               tanh, gelu, hardswish, hardsigmoid])
+    activations_after_fc_to_fuse = schema.OperatorSetConcat([any_relu, swish, sigmoid, tanh, gelu,
+                                                             hardswish, hardsigmoid])
+    any_binary = schema.OperatorSetConcat([add, sub, mul, div])
+
+    # ------------------- #
+    # Fusions
+    # ------------------- #
+    fusing_patterns.append(schema.Fusing((conv, activations_after_conv_to_fuse)))
+    fusing_patterns.append(schema.Fusing((fc, activations_after_fc_to_fuse)))
+    fusing_patterns.append(schema.Fusing((any_binary, any_relu)))
+
     # Create a TargetPlatformModel and set its default quantization config.
     # This default configuration will be used for all operations
     # unless specified otherwise (see OperatorsSet, for example):
@@ -235,76 +308,10 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         tpc_minor_version=4,
         tpc_patch_version=0,
         tpc_platform_type=IMX500_TP_MODEL,
+        operator_set=tuple(operator_set),
+        fusing_patterns=tuple(fusing_patterns),
         add_metadata=True,
         name=name,
         is_simd_padding=True)
 
-    # To start defining the model's components (such as operator sets, and fusing patterns),
-    # use 'with' the TargetPlatformModel instance, and create them as below:
-    with generated_tpm:
-        # Create an OperatorsSet to represent a set of operations.
-        # Each OperatorsSet has a unique label.
-        # If a quantization configuration options is passed, these options will
-        # be used for operations that will be attached to this set's label.
-        # Otherwise, it will be a configure-less set (used in fusing):
-
-        # May suit for operations like: Dropout, Reshape, etc.
-        default_qco = tp.get_default_quantization_config_options()
-        schema.OperatorsSet(OPSET_NO_QUANTIZATION,
-                               default_qco.clone_and_edit(enable_activation_quantization=False)
-                               .clone_and_edit_weight_attribute(enable_weights_quantization=False))
-        schema.OperatorsSet(OPSET_QUANTIZATION_PRESERVING,
-                               default_qco.clone_and_edit(enable_activation_quantization=False,
-                                                          quantization_preserving=True)
-                               .clone_and_edit_weight_attribute(enable_weights_quantization=False))
-        schema.OperatorsSet(OPSET_DIMENSION_MANIPULATION_OPS_WITH_WEIGHTS, qpreserving_const_config_options)
-        schema.OperatorsSet(OPSET_DIMENSION_MANIPULATION_OPS,
-                               default_qco.clone_and_edit(enable_activation_quantization=False,
-                                                          quantization_preserving=True,
-                                                          supported_input_activation_n_bits=(8, 16))
-                               .clone_and_edit_weight_attribute(enable_weights_quantization=False))
-        schema.OperatorsSet(OPSET_MERGE_OPS, const_configuration_options_inout16_per_tensor)
-
-        # Create Mixed-Precision quantization configuration options from the given list of OpQuantizationConfig objects
-        mixed_precision_configuration_options = schema.QuantizationConfigOptions(
-            mixed_precision_cfg_list + mp_cfg_list_16bit,
-            base_config=base_config)
-
-        # Define operator sets that use mixed_precision_configuration_options:
-        conv = schema.OperatorsSet(OPSET_CONV, mixed_precision_configuration_options)
-        fc = schema.OperatorsSet(OPSET_FULLY_CONNECTED, mixed_precision_configuration_options)
-
-        schema.OperatorsSet(OPSET_BATCH_NORM, default_config_options_16bit)
-
-        # Note: Operations sets without quantization configuration are useful for creating fusing patterns
-        any_relu = schema.OperatorsSet(OPSET_ANY_RELU, default_config_options_16bit)
-        add = schema.OperatorsSet(OPSET_ADD, const_configuration_options_inout16)
-        sub = schema.OperatorsSet(OPSET_SUB, const_configuration_options_inout16)
-        mul = schema.OperatorsSet(OPSET_MUL, const_configuration_options_inout16)
-        div = schema.OperatorsSet(OPSET_DIV, const_configuration_options)
-        schema.OperatorsSet(OPSET_MIN_MAX, const_configuration_options_inout16)
-        prelu = schema.OperatorsSet(OPSET_PRELU, default_config_options_16bit)
-        swish = schema.OperatorsSet(OPSET_SWISH, default_config_options_16bit)
-        sigmoid = schema.OperatorsSet(OPSET_SIGMOID, default_config_options_16bit)
-        tanh = schema.OperatorsSet(OPSET_TANH, default_config_options_16bit)
-        gelu = schema.OperatorsSet(OPSET_GELU, default_config_options_16bit)
-        hardsigmoid = schema.OperatorsSet(OPSET_HARDSIGMOID, default_config_options_16bit)
-        hardswish = schema.OperatorsSet(OPSET_HARDSWISH, default_config_options_16bit)
-
-        # Combine multiple operators into a single operator to avoid quantization between
-        # them. To do this we define fusing patterns using the OperatorsSets that were created.
-        # To group multiple sets with regard to fusing, an OperatorSetConcat can be created
-        activations_after_conv_to_fuse = schema.OperatorSetConcat([any_relu, swish, prelu, sigmoid,
-                                                                     tanh, gelu, hardswish, hardsigmoid])
-        activations_after_fc_to_fuse = schema.OperatorSetConcat([any_relu, swish, sigmoid, tanh, gelu,
-                                                                   hardswish, hardsigmoid])
-        any_binary = schema.OperatorSetConcat([add, sub, mul, div])
-
-        # ------------------- #
-        # Fusions
-        # ------------------- #
-        schema.Fusing([conv, activations_after_conv_to_fuse])
-        schema.Fusing([fc, activations_after_fc_to_fuse])
-        schema.Fusing([any_binary, any_relu])
-
     return generated_tpm

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

@@ -18,7 +18,8 @@ import model_compression_toolkit as mct
 import model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema as schema
 from model_compression_toolkit.constants import FLOAT_BITWIDTH
 from model_compression_toolkit.target_platform_capabilities.constants import KERNEL_ATTR, BIAS_ATTR, QNNPACK_TP_MODEL
-from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import TargetPlatformModel, Signedness, \
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import TargetPlatformModel, \
+    Signedness, \
     AttributeQuantizationConfig, OpQuantizationConfig
 
 tp = mct.target_platform
@@ -138,8 +139,28 @@ def generate_tp_model(default_config: OpQuantizationConfig,
     # of possible configurations to consider when quantizing a set of operations (in mixed-precision, for example).
     # If the QuantizationConfigOptions contains only one configuration,
     # this configuration will be used for the operation quantization:
-    default_configuration_options = schema.QuantizationConfigOptions([default_config])
-
+    default_configuration_options = schema.QuantizationConfigOptions(tuple([default_config]))
+
+    # Combine operations/modules into a single module.
+    # Pytorch supports the next fusing patterns:
+    # [Conv, Relu], [Conv, BatchNorm], [Conv, BatchNorm, Relu], [Linear, Relu]
+    # Source: # https://pytorch.org/docs/stable/quantization.html#model-preparation-for-quantization-eager-mode
+    operator_set = []
+    fusing_patterns = []
+
+    conv = schema.OperatorsSet("Conv")
+    batchnorm = schema.OperatorsSet("BatchNorm")
+    relu = schema.OperatorsSet("Relu")
+    linear = schema.OperatorsSet("Linear")
+
+    operator_set.extend([conv, batchnorm, relu, linear])
+    # ------------------- #
+    # Fusions
+    # ------------------- #
+    fusing_patterns.append(schema.Fusing((conv, batchnorm, relu)))
+    fusing_patterns.append(schema.Fusing((conv, batchnorm)))
+    fusing_patterns.append(schema.Fusing((conv, relu)))
+    fusing_patterns.append(schema.Fusing((linear, relu)))
     # Create a TargetPlatformModel and set its default quantization config.
     # This default configuration will be used for all operations
     # unless specified otherwise (see OperatorsSet, for example):
@@ -148,27 +169,8 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         tpc_minor_version=1,
         tpc_patch_version=0,
         tpc_platform_type=QNNPACK_TP_MODEL,
+        operator_set=tuple(operator_set),
+        fusing_patterns=tuple(fusing_patterns),
         add_metadata=False,
         name=name)
-
-    # To start defining the model's components (such as operator sets, and fusing patterns),
-    # use 'with' the target platform model instance, and create them as below:
-    with generated_tpc:
-        # Combine operations/modules into a single module.
-        # Pytorch supports the next fusing patterns:
-        # [Conv, Relu], [Conv, BatchNorm], [Conv, BatchNorm, Relu], [Linear, Relu]
-        # Source: # https://pytorch.org/docs/stable/quantization.html#model-preparation-for-quantization-eager-mode
-        conv = schema.OperatorsSet("Conv")
-        batchnorm = schema.OperatorsSet("BatchNorm")
-        relu = schema.OperatorsSet("Relu")
-        linear = schema.OperatorsSet("Linear")
-
-        # ------------------- #
-        # Fusions
-        # ------------------- #
-        schema.Fusing([conv, batchnorm, relu])
-        schema.Fusing([conv, batchnorm])
-        schema.Fusing([conv, relu])
-        schema.Fusing([linear, relu])
-
     return generated_tpc

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

@@ -136,7 +136,61 @@ def generate_tp_model(default_config: OpQuantizationConfig,
     # of possible configurations to consider when quantizing a set of operations (in mixed-precision, for example).
     # If the QuantizationConfigOptions contains only one configuration,
     # this configuration will be used for the operation quantization:
-    default_configuration_options = schema.QuantizationConfigOptions([default_config])
+    default_configuration_options = schema.QuantizationConfigOptions(tuple([default_config]))
+
+    # In TFLite, the quantized operator specifications constraint operators quantization
+    # differently. For more details:
+    # https://www.tensorflow.org/lite/performance/quantization_spec#int8_quantized_operator_specifications
+    operator_set = []
+    fusing_patterns = []
+
+    operator_set.append(schema.OperatorsSet("NoQuantization",
+                           default_configuration_options.clone_and_edit(
+                               quantization_preserving=True)))
+
+    fc = schema.OperatorsSet("FullyConnected",
+                                default_configuration_options.clone_and_edit_weight_attribute(weights_per_channel_threshold=False))
+
+    operator_set.append(schema.OperatorsSet("L2Normalization",
+                           default_configuration_options.clone_and_edit(
+                               fixed_zero_point=0, fixed_scale=1 / 128)))
+    operator_set.append(schema.OperatorsSet("LogSoftmax",
+                           default_configuration_options.clone_and_edit(
+                               fixed_zero_point=127, fixed_scale=16 / 256)))
+    operator_set.append(schema.OperatorsSet("Tanh",
+                           default_configuration_options.clone_and_edit(
+                               fixed_zero_point=0, fixed_scale=1 / 128)))
+    operator_set.append(schema.OperatorsSet("Softmax",
+                           default_configuration_options.clone_and_edit(
+                               fixed_zero_point=-128, fixed_scale=1 / 256)))
+    operator_set.append(schema.OperatorsSet("Logistic",
+                           default_configuration_options.clone_and_edit(
+                               fixed_zero_point=-128, fixed_scale=1 / 256)))
+
+    conv2d = schema.OperatorsSet("Conv2d")
+    kernel = schema.OperatorSetConcat([conv2d, fc])
+
+    relu = schema.OperatorsSet("Relu")
+    elu = schema.OperatorsSet("Elu")
+    activations_to_fuse = schema.OperatorSetConcat([relu, elu])
+
+    batch_norm = schema.OperatorsSet("BatchNorm")
+    bias_add = schema.OperatorsSet("BiasAdd")
+    add = schema.OperatorsSet("Add")
+    squeeze = schema.OperatorsSet("Squeeze",
+                                     qc_options=default_configuration_options.clone_and_edit(
+                                         quantization_preserving=True))
+    operator_set.extend([fc, conv2d, kernel, relu, elu, batch_norm, bias_add, add, squeeze])
+    # ------------------- #
+    # Fusions
+    # ------------------- #
+    # Source: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/grappler/optimizers/remapper
+    fusing_patterns.append(schema.Fusing((kernel, bias_add)))
+    fusing_patterns.append(schema.Fusing((kernel, bias_add, activations_to_fuse)))
+    fusing_patterns.append(schema.Fusing((conv2d, batch_norm, activations_to_fuse)))
+    fusing_patterns.append(schema.Fusing((conv2d, squeeze, activations_to_fuse)))
+    fusing_patterns.append(schema.Fusing((batch_norm, activations_to_fuse)))
+    fusing_patterns.append(schema.Fusing((batch_norm, add, activations_to_fuse)))
 
     # Create a TargetPlatformModel and set its default quantization config.
     # This default configuration will be used for all operations
@@ -145,62 +199,10 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         default_configuration_options,
         tpc_minor_version=1,
         tpc_patch_version=0,
+        operator_set=tuple(operator_set),
+        fusing_patterns=tuple(fusing_patterns),
         tpc_platform_type=TFLITE_TP_MODEL,
         add_metadata=False,
         name=name)
 
-    # To start defining the model's components (such as operator sets, and fusing patterns),
-    # use 'with' the TargetPlatformModel instance, and create them as below:
-    with generated_tpc:
-        # In TFLite, the quantized operator specifications constraint operators quantization
-        # differently. For more details:
-        # https://www.tensorflow.org/lite/performance/quantization_spec#int8_quantized_operator_specifications
-        schema.OperatorsSet("NoQuantization",
-                               tp.get_default_quantization_config_options().clone_and_edit(
-                                   quantization_preserving=True))
-
-        fc_qco = tp.get_default_quantization_config_options()
-        fc = schema.OperatorsSet("FullyConnected",
-                                    fc_qco.clone_and_edit_weight_attribute(weights_per_channel_threshold=False))
-
-        schema.OperatorsSet("L2Normalization",
-                               tp.get_default_quantization_config_options().clone_and_edit(
-                                   fixed_zero_point=0, fixed_scale=1 / 128))
-        schema.OperatorsSet("LogSoftmax",
-                               tp.get_default_quantization_config_options().clone_and_edit(
-                                   fixed_zero_point=127, fixed_scale=16 / 256))
-        schema.OperatorsSet("Tanh",
-                               tp.get_default_quantization_config_options().clone_and_edit(
-                                   fixed_zero_point=0, fixed_scale=1 / 128))
-        schema.OperatorsSet("Softmax",
-                               tp.get_default_quantization_config_options().clone_and_edit(
-                                   fixed_zero_point=-128, fixed_scale=1 / 256))
-        schema.OperatorsSet("Logistic",
-                               tp.get_default_quantization_config_options().clone_and_edit(
-                                   fixed_zero_point=-128, fixed_scale=1 / 256))
-
-        conv2d = schema.OperatorsSet("Conv2d")
-        kernel = schema.OperatorSetConcat([conv2d, fc])
-
-        relu = schema.OperatorsSet("Relu")
-        elu = schema.OperatorsSet("Elu")
-        activations_to_fuse = schema.OperatorSetConcat([relu, elu])
-
-        batch_norm = schema.OperatorsSet("BatchNorm")
-        bias_add = schema.OperatorsSet("BiasAdd")
-        add = schema.OperatorsSet("Add")
-        squeeze = schema.OperatorsSet("Squeeze",
-                                         qc_options=tp.get_default_quantization_config_options().clone_and_edit(
-                                             quantization_preserving=True))
-        # ------------------- #
-        # Fusions
-        # ------------------- #
-        # Source: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/grappler/optimizers/remapper
-        schema.Fusing([kernel, bias_add])
-        schema.Fusing([kernel, bias_add, activations_to_fuse])
-        schema.Fusing([conv2d, batch_norm, activations_to_fuse])
-        schema.Fusing([conv2d, squeeze, activations_to_fuse])
-        schema.Fusing([batch_norm, activations_to_fuse])
-        schema.Fusing([batch_norm, add, activations_to_fuse])
-
     return generated_tpc

model_compression_toolkit/target_platform_capabilities/target_platform/current_tp_model.py

@@ -1,67 +0,0 @@
-# Copyright 2022 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.
-# ==============================================================================
-
-from model_compression_toolkit.logger import Logger
-
-def get_current_tp_model():
-    """
-
-    Returns: The current TargetPlatformModel that is being used and accessed.
-
-    """
-    return _current_tp_model.get()
-
-
-class CurrentTPModel:
-    """
-    Wrapper of the current TargetPlatformModel object that is being accessed and defined.
-    """
-
-    def __init__(self):
-        super(CurrentTPModel, self).__init__()
-        self.tp_model = None
-
-    def get(self):
-        """
-
-        Returns: The current TargetPlatformModel that is being defined.
-
-        """
-        if self.tp_model is None:
-            Logger.critical('Target platform model is not initialized.')  # pragma: no cover
-        return self.tp_model
-
-    def reset(self):
-        """
-
-        Reset the current TargetPlatformModel so a new TargetPlatformModel can be wrapped and
-        used as the current TargetPlatformModel object.
-
-        """
-        self.tp_model = None
-
-    def set(self, tp_model):
-        """
-        Set and wrap a TargetPlatformModel as the current TargetPlatformModel.
-
-        Args:
-            tp_model: TargetPlatformModel to set as the current TargetPlatformModel to access and use.
-
-        """
-        self.tp_model = tp_model
-
-
-# Use a single instance for the current model.
-_current_tp_model = CurrentTPModel()