mct-nightly 2.1.0.20240801.515__py3-none-any.whl → 2.1.0.20240803.442__py3-none-any.whl

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

@@ -0,0 +1,235 @@
+# 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.
+# ==============================================================================
+from typing import List, Tuple
+
+import model_compression_toolkit as mct
+from model_compression_toolkit.constants import FLOAT_BITWIDTH
+from model_compression_toolkit.target_platform_capabilities.constants import KERNEL_ATTR, BIAS_ATTR, WEIGHTS_N_BITS
+from model_compression_toolkit.target_platform_capabilities.target_platform import OpQuantizationConfig, \
+    TargetPlatformModel
+from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
+    AttributeQuantizationConfig
+
+tp = mct.target_platform
+
+
+def get_tp_model() -> TargetPlatformModel:
+    """
+    A method that generates a default target platform model, with base 8-bit quantization configuration and 8, 4, 2
+    bits configuration list for mixed-precision quantization.
+    NOTE: in order to generate a target platform model with different configurations but with the same Operators Sets
+    (for tests, experiments, etc.), use this method implementation as a test-case, i.e., override the
+    'get_op_quantization_configs' method and use its output to call 'generate_tp_model' with your configurations.
+    This version enables metadata by default.
+
+    Returns: A TargetPlatformModel object.
+
+    """
+    base_config, mixed_precision_cfg_list, default_config = get_op_quantization_configs()
+    return generate_tp_model(default_config=default_config,
+                             base_config=base_config,
+                             mixed_precision_cfg_list=mixed_precision_cfg_list,
+                             name='imx500_tp_model')
+
+
+def get_op_quantization_configs() -> \
+        Tuple[OpQuantizationConfig, List[OpQuantizationConfig], OpQuantizationConfig]:
+    """
+    Creates a default configuration object for 8-bit quantization, to be used to set a default TargetPlatformModel.
+    In addition, creates a default configuration objects list (with 8, 4 and 2 bit quantization) to be used as
+    default configuration for mixed-precision quantization.
+
+    Returns: An OpQuantizationConfig config object and a list of OpQuantizationConfig objects.
+
+    """
+
+    # TODO: currently, we don't want to quantize any attribute but the kernel by default,
+    #  to preserve the current behavior of MCT, so quantization is disabled for all other attributes.
+    #  Other quantization parameters are set to what we eventually want to quantize by default
+    #  when we enable multi-attributes quantization - THIS NEED TO BE MODIFIED IN ALL TP MODELS!
+
+    # define a default quantization config for all non-specified weights attributes.
+    default_weight_attr_config = AttributeQuantizationConfig(
+        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
+        lut_values_bitwidth=None)
+
+    # define a quantization config to quantize the kernel (for layers where there is a kernel attribute).
+    kernel_base_config = AttributeQuantizationConfig(
+        weights_quantization_method=tp.QuantizationMethod.SYMMETRIC,
+        weights_n_bits=8,
+        weights_per_channel_threshold=True,
+        enable_weights_quantization=True,
+        lut_values_bitwidth=None)
+
+    # define a quantization config to quantize the bias (for layers where there is a bias attribute).
+    bias_config = AttributeQuantizationConfig(
+        weights_quantization_method=tp.QuantizationMethod.POWER_OF_TWO,
+        weights_n_bits=FLOAT_BITWIDTH,
+        weights_per_channel_threshold=False,
+        enable_weights_quantization=False,
+        lut_values_bitwidth=None)
+
+    # Create a quantization config.
+    # A quantization configuration defines how an operator
+    # should be quantized on the modeled hardware:
+
+    # We define a default config for operation without kernel attribute.
+    # This is the default config that should be used for non-linear operations.
+    eight_bits_default = tp.OpQuantizationConfig(
+        default_weight_attr_config=default_weight_attr_config,
+        attr_weights_configs_mapping={},
+        activation_quantization_method=tp.QuantizationMethod.POWER_OF_TWO,
+        activation_n_bits=8,
+        supported_input_activation_n_bits=8,
+        enable_activation_quantization=True,
+        quantization_preserving=False,
+        fixed_scale=None,
+        fixed_zero_point=None,
+        simd_size=32)
+
+    # We define an 8-bit config for linear operations quantization, that include a kernel and bias attributes.
+    linear_eight_bits = tp.OpQuantizationConfig(
+        default_weight_attr_config=default_weight_attr_config,
+        attr_weights_configs_mapping={KERNEL_ATTR: kernel_base_config, BIAS_ATTR: bias_config},
+        activation_quantization_method=tp.QuantizationMethod.POWER_OF_TWO,
+        activation_n_bits=8,
+        supported_input_activation_n_bits=8,
+        enable_activation_quantization=True,
+        quantization_preserving=False,
+        fixed_scale=None,
+        fixed_zero_point=None,
+        simd_size=32)
+
+    # To quantize a model using mixed-precision, create
+    # a list with more than one OpQuantizationConfig.
+    # In this example, we quantize some operations' weights
+    # using 2, 4 or 8 bits, and when using 2 or 4 bits, it's possible
+    # to quantize the operations' activations using LUT.
+    four_bits = linear_eight_bits.clone_and_edit(attr_to_edit={KERNEL_ATTR: {WEIGHTS_N_BITS: 4}},
+                                                 simd_size=linear_eight_bits.simd_size * 2)
+    two_bits = linear_eight_bits.clone_and_edit(attr_to_edit={KERNEL_ATTR: {WEIGHTS_N_BITS: 2}},
+                                                simd_size=linear_eight_bits.simd_size * 4)
+
+    mixed_precision_cfg_list = [linear_eight_bits, four_bits, two_bits]
+
+    return linear_eight_bits, mixed_precision_cfg_list, eight_bits_default
+
+
+def generate_tp_model(default_config: OpQuantizationConfig,
+                      base_config: OpQuantizationConfig,
+                      mixed_precision_cfg_list: List[OpQuantizationConfig],
+                      name: str) -> TargetPlatformModel:
+    """
+    Generates TargetPlatformModel with default defined Operators Sets, based on the given base configuration and
+    mixed-precision configurations options list.
+
+    Args
+        default_config: A default OpQuantizationConfig to set as the TP model default configuration.
+        base_config: An OpQuantizationConfig to set as the TargetPlatformModel base configuration for mixed-precision purposes only.
+        mixed_precision_cfg_list: A list of OpQuantizationConfig to be used as the TP model mixed-precision
+            quantization configuration options.
+        name: The name of the TargetPlatformModel.
+
+    Returns: A TargetPlatformModel object.
+
+    """
+    # Create a QuantizationConfigOptions, which defines a set
+    # 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 = tp.QuantizationConfigOptions([default_config])
+
+    # Create a QuantizationConfigOptions for quantizing constants in functional ops.
+    # Constant configuration is similar to the default eight bit configuration except for PoT
+    # quantization method for the constant.
+    # Since the constants are not named attributes of the layer, we use the default_weight_attr_config to
+    # define the desired quantization properties for them.
+    const_config = default_config.clone_and_edit(
+        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 = tp.QuantizationConfigOptions([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.
+    const_config_input16 = const_config.clone_and_edit(
+        supported_input_activation_n_bits=(8, 16))
+    const_config_input16_output16 = const_config_input16.clone_and_edit(
+        activation_n_bits=16, is_signed=True)
+    const_configuration_options_inout16 = tp.QuantizationConfigOptions([const_config_input16_output16,
+                                                                        const_config_input16],
+                                                                       base_config=const_config_input16)
+
+    # 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):
+    generated_tpm = tp.TargetPlatformModel(default_configuration_options, 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):
+
+        generated_tpm.set_simd_padding(is_simd_padding=True)
+
+        # May suit for operations like: Dropout, Reshape, etc.
+        default_qco = tp.get_default_quantization_config_options()
+        tp.OperatorsSet("NoQuantization",
+                        default_qco.clone_and_edit(enable_activation_quantization=False,
+                                                   supported_input_activation_n_bits=(8, 16))
+                        .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 = tp.QuantizationConfigOptions(mixed_precision_cfg_list,
+                                                                             base_config=base_config)
+
+        # Define operator sets that use mixed_precision_configuration_options:
+        conv = tp.OperatorsSet("Conv", mixed_precision_configuration_options)
+        fc = tp.OperatorsSet("FullyConnected", mixed_precision_configuration_options)
+
+        # Define operations sets without quantization configuration
+        # options (useful for creating fusing patterns, for example):
+        any_relu = tp.OperatorsSet("AnyReLU")
+        add = tp.OperatorsSet("Add", const_configuration_options_inout16)
+        sub = tp.OperatorsSet("Sub", const_configuration_options_inout16)
+        mul = tp.OperatorsSet("Mul", const_configuration_options_inout16)
+        div = tp.OperatorsSet("Div", const_configuration_options)
+        prelu = tp.OperatorsSet("PReLU")
+        swish = tp.OperatorsSet("Swish")
+        sigmoid = tp.OperatorsSet("Sigmoid")
+        tanh = tp.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 = tp.OperatorSetConcat(any_relu, swish, prelu, sigmoid, tanh)
+        activations_after_fc_to_fuse = tp.OperatorSetConcat(any_relu, swish, sigmoid)
+        any_binary = tp.OperatorSetConcat(add, sub, mul, div)
+
+        # ------------------- #
+        # Fusions
+        # ------------------- #
+        tp.Fusing([conv, activations_after_conv_to_fuse])
+        tp.Fusing([fc, activations_after_fc_to_fuse])
+        tp.Fusing([any_binary, any_relu])
+
+    return generated_tpm

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

@@ -0,0 +1,132 @@
+# 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 tensorflow as tf
+from packaging import version
+
+from model_compression_toolkit.defaultdict import DefaultDict
+from model_compression_toolkit.constants import FOUND_SONY_CUSTOM_LAYERS
+from model_compression_toolkit.target_platform_capabilities.constants import KERNEL_ATTR, KERAS_DEPTHWISE_KERNEL, \
+    KERAS_KERNEL, BIAS_ATTR, BIAS
+
+if FOUND_SONY_CUSTOM_LAYERS:
+    from sony_custom_layers.keras.object_detection.ssd_post_process import SSDPostProcess
+
+if version.parse(tf.__version__) >= version.parse("2.13"):
+    from keras.src.layers import Conv2D, DepthwiseConv2D, Dense, Reshape, ZeroPadding2D, Dropout, \
+        MaxPooling2D, Activation, ReLU, Add, Subtract, Multiply, PReLU, Flatten, Cropping2D, LeakyReLU, Permute, \
+        Conv2DTranspose, Identity
+else:
+    from keras.layers import Conv2D, DepthwiseConv2D, Dense, Reshape, ZeroPadding2D, Dropout, \
+        MaxPooling2D, Activation, ReLU, Add, Subtract, Multiply, PReLU, Flatten, Cropping2D, LeakyReLU, Permute, \
+        Conv2DTranspose, Identity
+
+from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v4.tp_model import get_tp_model
+import model_compression_toolkit as mct
+from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v4 import __version__ as TPC_VERSION
+
+tp = mct.target_platform
+
+
+def get_keras_tpc() -> tp.TargetPlatformCapabilities:
+    """
+    get a Keras TargetPlatformCapabilities object with default operation sets to layers mapping.
+
+    Returns: a Keras TargetPlatformCapabilities object for the given TargetPlatformModel.
+    """
+    imx500_tpc_tp_model = get_tp_model()
+    return generate_keras_tpc(name='imx500_tpc_keras_tpc', tp_model=imx500_tpc_tp_model)
+
+
+def generate_keras_tpc(name: str, tp_model: tp.TargetPlatformModel):
+    """
+    Generates a TargetPlatformCapabilities object with default operation sets to layers mapping.
+
+    Args:
+        name: Name of the TargetPlatformCapabilities.
+        tp_model: TargetPlatformModel object.
+
+    Returns: a TargetPlatformCapabilities object for the given TargetPlatformModel.
+    """
+
+    keras_tpc = tp.TargetPlatformCapabilities(tp_model, name=name, version=TPC_VERSION)
+
+    no_quant_list = [Identity,
+                     tf.identity,
+                     Reshape,
+                     tf.reshape,
+                     Permute,
+                     tf.transpose,
+                     Flatten,
+                     Cropping2D,
+                     ZeroPadding2D,
+                     Dropout,
+                     MaxPooling2D,
+                     tf.split,
+                     tf.quantization.fake_quant_with_min_max_vars,
+                     tf.math.argmax,
+                     tf.shape,
+                     tf.math.equal,
+                     tf.gather,
+                     tf.cast,
+                     tf.unstack,
+                     tf.compat.v1.gather,
+                     tf.nn.top_k,
+                     tf.__operators__.getitem,
+                     tf.strided_slice,
+                     tf.image.combined_non_max_suppression,
+                     tf.compat.v1.shape]
+
+    if FOUND_SONY_CUSTOM_LAYERS:
+        no_quant_list.append(SSDPostProcess)
+
+    with keras_tpc:
+        tp.OperationsSetToLayers("NoQuantization", no_quant_list)
+        tp.OperationsSetToLayers("Conv",
+                                 [Conv2D,
+                                  DepthwiseConv2D,
+                                  Conv2DTranspose,
+                                  tf.nn.conv2d,
+                                  tf.nn.depthwise_conv2d,
+                                  tf.nn.conv2d_transpose],
+                                 # we provide attributes mapping that maps each layer type in the operations set
+                                 # that has weights attributes with provided quantization config (in the tp model) to
+                                 # its framework-specific attribute name.
+                                 # note that a DefaultDict should be provided if not all the layer types in the
+                                 # operation set are provided separately in the mapping.
+                                 attr_mapping={
+                                     KERNEL_ATTR: DefaultDict({
+                                         DepthwiseConv2D: KERAS_DEPTHWISE_KERNEL,
+                                         tf.nn.depthwise_conv2d: KERAS_DEPTHWISE_KERNEL}, default_value=KERAS_KERNEL),
+                                     BIAS_ATTR: DefaultDict(default_value=BIAS)})
+        tp.OperationsSetToLayers("FullyConnected", [Dense],
+                                 attr_mapping={KERNEL_ATTR: DefaultDict(default_value=KERAS_KERNEL),
+                                               BIAS_ATTR: DefaultDict(default_value=BIAS)})
+        tp.OperationsSetToLayers("AnyReLU", [tf.nn.relu,
+                                             tf.nn.relu6,
+                                             tf.nn.leaky_relu,
+                                             ReLU,
+                                             LeakyReLU,
+                                             tp.LayerFilterParams(Activation, activation="relu"),
+                                             tp.LayerFilterParams(Activation, activation="leaky_relu")])
+        tp.OperationsSetToLayers("Add", [tf.add, Add])
+        tp.OperationsSetToLayers("Sub", [tf.subtract, Subtract])
+        tp.OperationsSetToLayers("Mul", [tf.math.multiply, Multiply])
+        tp.OperationsSetToLayers("Div", [tf.math.divide, tf.math.truediv])
+        tp.OperationsSetToLayers("PReLU", [PReLU])
+        tp.OperationsSetToLayers("Swish", [tf.nn.swish, tp.LayerFilterParams(Activation, activation="swish")])
+        tp.OperationsSetToLayers("Sigmoid", [tf.nn.sigmoid, tp.LayerFilterParams(Activation, activation="sigmoid")])
+        tp.OperationsSetToLayers("Tanh", [tf.nn.tanh, tp.LayerFilterParams(Activation, activation="tanh")])
+
+    return keras_tpc

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

@@ -0,0 +1,112 @@
+# 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 operator
+
+import torch
+from torch import add, sub, mul, div, flatten, reshape, split, unsqueeze, dropout, sigmoid, tanh, chunk, unbind, topk, \
+    gather, equal, transpose, permute, argmax, squeeze
+from torch.nn import Conv2d, Linear, ConvTranspose2d, MaxPool2d
+from torch.nn import Dropout, Flatten, Hardtanh, Identity
+from torch.nn import ReLU, ReLU6, PReLU, SiLU, Sigmoid, Tanh, Hardswish, LeakyReLU
+from torch.nn.functional import relu, relu6, prelu, silu, hardtanh, hardswish, leaky_relu
+
+from model_compression_toolkit.defaultdict import DefaultDict
+from model_compression_toolkit.target_platform_capabilities.constants import KERNEL_ATTR, BIAS_ATTR, PYTORCH_KERNEL, \
+    BIAS
+from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v4.tp_model import get_tp_model
+import model_compression_toolkit as mct
+from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v4 import __version__ as TPC_VERSION
+
+tp = mct.target_platform
+
+
+def get_pytorch_tpc() -> tp.TargetPlatformCapabilities:
+    """
+    get a Pytorch TargetPlatformCapabilities object with default operation sets to layers mapping.
+
+    Returns: a Pytorch TargetPlatformCapabilities object for the given TargetPlatformModel.
+    """
+    imx500_tpc_tp_model = get_tp_model()
+    return generate_pytorch_tpc(name='imx500_tpc_pytorch_tpc', tp_model=imx500_tpc_tp_model)
+
+
+def generate_pytorch_tpc(name: str, tp_model: tp.TargetPlatformModel):
+    """
+    Generates a TargetPlatformCapabilities object with default operation sets to layers mapping.
+    Args:
+        name: Name of the TargetPlatformModel.
+        tp_model: TargetPlatformModel object.
+    Returns: a TargetPlatformCapabilities object for the given TargetPlatformModel.
+    """
+
+    pytorch_tpc = tp.TargetPlatformCapabilities(tp_model,
+                                                name=name,
+                                                version=TPC_VERSION)
+
+    # we provide attributes mapping that maps each layer type in the operations set
+    # that has weights attributes with provided quantization config (in the tp model) to
+    # its framework-specific attribute name.
+    # note that a DefaultDict should be provided if not all the layer types in the
+    # operation set are provided separately in the mapping.
+    pytorch_linear_attr_mapping = {KERNEL_ATTR: DefaultDict(default_value=PYTORCH_KERNEL),
+                                   BIAS_ATTR: DefaultDict(default_value=BIAS)}
+
+    with pytorch_tpc:
+        tp.OperationsSetToLayers("NoQuantization", [Identity,
+                                                    Dropout,
+                                                    Flatten,
+                                                    dropout,
+                                                    flatten,
+                                                    split,
+                                                    operator.getitem,
+                                                    reshape,
+                                                    unsqueeze,
+                                                    chunk,
+                                                    unbind,
+                                                    torch.Tensor.size,
+                                                    permute,
+                                                    transpose,
+                                                    equal,
+                                                    argmax,
+                                                    gather,
+                                                    topk,
+                                                    squeeze,
+                                                    MaxPool2d])
+
+        tp.OperationsSetToLayers("Conv", [Conv2d, ConvTranspose2d],
+                                 attr_mapping=pytorch_linear_attr_mapping)
+        tp.OperationsSetToLayers("FullyConnected", [Linear],
+                                 attr_mapping=pytorch_linear_attr_mapping)
+        tp.OperationsSetToLayers("AnyReLU", [torch.relu,
+                                             ReLU,
+                                             ReLU6,
+                                             LeakyReLU,
+                                             relu,
+                                             relu6,
+                                             leaky_relu,
+                                             tp.LayerFilterParams(Hardtanh, min_val=0),
+                                             tp.LayerFilterParams(hardtanh, min_val=0)])
+
+        tp.OperationsSetToLayers("Add", [operator.add, add])
+        tp.OperationsSetToLayers("Sub", [operator.sub, sub])
+        tp.OperationsSetToLayers("Mul", [operator.mul, mul])
+        tp.OperationsSetToLayers("Div", [operator.truediv, div])
+        tp.OperationsSetToLayers("PReLU", [PReLU, prelu])
+        tp.OperationsSetToLayers("Swish", [SiLU, silu, Hardswish, hardswish])
+        tp.OperationsSetToLayers("Sigmoid", [Sigmoid, sigmoid])
+        tp.OperationsSetToLayers("Tanh", [Tanh, tanh])
+
+    return pytorch_tpc

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

@@ -90,6 +90,7 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         attr_weights_configs_mapping={},
         activation_quantization_method=tp.QuantizationMethod.POWER_OF_TWO,
         activation_n_bits=8,
+        supported_input_activation_n_bits=8,
         enable_activation_quantization=True,
         quantization_preserving=False,
         fixed_scale=None,
@@ -102,6 +103,7 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         default_weight_attr_config=default_weight_attr_config,
         attr_weights_configs_mapping={KERNEL_ATTR: kernel_base_config, BIAS_ATTR: bias_config},
         activation_n_bits=8,
+        supported_input_activation_n_bits=8,
         enable_activation_quantization=True,
         quantization_preserving=False,
         fixed_scale=None,

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

@@ -88,6 +88,7 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         attr_weights_configs_mapping={},
         activation_quantization_method=tp.QuantizationMethod.POWER_OF_TWO,
         activation_n_bits=8,
+        supported_input_activation_n_bits=8,
         enable_activation_quantization=True,
         quantization_preserving=False,
         fixed_scale=None,
@@ -100,6 +101,7 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         default_weight_attr_config=default_weight_attr_config,
         attr_weights_configs_mapping={KERNEL_ATTR: kernel_base_config, BIAS_ATTR: bias_config},
         activation_n_bits=8,
+        supported_input_activation_n_bits=8,
         enable_activation_quantization=True,
         quantization_preserving=False,
         fixed_scale=None,