Merge branch 'main' into nightly-workflow-reponame

Author: lapid92

model_compression_toolkit/core/common/framework_info.py

@@ -13,14 +13,10 @@
 # limitations under the License.
 # ==============================================================================
 
-
-from collections.abc import Callable
 from enum import Enum
 from typing import Dict, Any, Tuple, NamedTuple, Optional
 from abc import ABC, abstractmethod
 
-from mct_quantizers import QuantizationMethod
-
 
 class ChannelAxis(Enum):
     """
@@ -56,13 +52,11 @@ class FrameworkInfo(ABC):
         kernel_ops_attribute_mapping (Dict): Dictionary from a framework operator to its weight attribute to quantize.
         out_channel_axis_mapping (Dict): Dictionary of output channels of the model's layers (for computing statistics per-channel).
         _layer_min_max_mapping (Dict[Any, tuple]): Dictionary from a layer to its min/max output values.
-        activation_quantizer_factory_mapping: A mapping from QuantizationMethod to a factory function that accepts activation bitwidth and a dict of quantization params, and returns the corresponding quantization function.
     """
 
     kernel_ops_attribute_mapping: Dict[Any, str]
     kernel_channels_mapping: Dict[Any, ChannelAxisMapping]
     out_channel_axis_mapping: Dict[Any, int]
-    activation_quantizer_factory_mapping: Dict[QuantizationMethod, Callable[[int, dict], Callable]]
 
     _layer_min_max_mapping: Dict[Any, tuple]
     _default_channel_mapping = ChannelAxisMapping(None, None)

model_compression_toolkit/core/common/mixed_precision/configurable_quantizer_utils.py

@@ -92,22 +92,24 @@ def init_quantized_weights(node_q_cfg: List[CandidateNodeQuantizationConfig],
     return quantized_weights
 
 
-def init_activation_quantizers(node_q_cfg: List[CandidateNodeQuantizationConfig]) -> List:
+def init_activation_quantizers(node_q_cfg: List[CandidateNodeQuantizationConfig],
+                               get_activation_quantization_fn_factory: Callable) -> List:
     """
     Builds a list of quantizers for each of the bitwidth candidates for activation quantization,
     to be stored and used during MP search.
 
     Args:
         node_q_cfg: Quantization configuration candidates of the node that generated the layer that will
                     use this quantizer.
+        get_activation_quantization_fn_factory: activation quantization functions factory.
 
     Returns: a list of activation quantizers - for each bitwidth and layer's attribute to be quantized.
     """
 
     activation_quantizers = []
     for index, qc in enumerate(node_q_cfg):
         q_activation = node_q_cfg[index].activation_quantization_cfg
-        quantizer = get_activation_quantization_fn(q_activation)
+        quantizer = get_activation_quantization_fn(q_activation, get_activation_quantization_fn_factory)
         activation_quantizers.append(quantizer)
 
     return activation_quantizers

model_compression_toolkit/core/common/quantization/quantization_fn_selection.py

@@ -17,33 +17,21 @@
 
 from mct_quantizers import QuantizationMethod
 
-from model_compression_toolkit.core.common.framework_info import get_fw_info
 from model_compression_toolkit.core.common.quantization.node_quantization_config import NodeActivationQuantizationConfig
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.core.common.quantization.quantizers.lut_kmeans_quantizer import lut_kmeans_quantizer
 from model_compression_toolkit.core.common.quantization.quantizers.uniform_quantizers import power_of_two_quantizer, \
     symmetric_quantizer, uniform_quantizer
 
 
-def get_activation_quantization_fn_factory(quantization_method: QuantizationMethod) -> Callable[[int, dict], Callable]:
-    """
-    Get factory for activation quantizer.
-
-    Args:
-        quantization_method: quantization method for activation.
-
-    Returns:
-        Factory that accepts activation bitwidth and a dict of quantization params, and returns the quantizer.
-    """
-    return get_fw_info().activation_quantizer_factory_mapping[quantization_method]
-
-
-def get_activation_quantization_fn(activation_quantization_cfg: NodeActivationQuantizationConfig) -> Callable:
+def get_activation_quantization_fn(activation_quantization_cfg: NodeActivationQuantizationConfig,
+                                   get_activation_quantization_fn_factory: Callable) -> Callable:
     """
     Get activation quantizer based on activation quantization configuration.
 
     Args:
         activation_quantization_cfg: activation quantization configuration.
+        get_activation_quantization_fn_factory: activation quantization functions factory.
 
     Returns:
         Activation quantizer that accepts a tensor and returns a quantized tensor.

model_compression_toolkit/core/common/statistics_correction/compute_activation_bias_correction_of_graph.py

@@ -67,7 +67,8 @@ def compute_activation_bias_correction(graph: Graph,
                                        fw_impl: FrameworkImplementation,
                                        linear_node: BaseNode,
                                        prev_node: BaseNode,
-                                       kernel_size: str) -> Graph:
+                                       kernel_size: str,
+                                       get_activation_quantization_fn_factory: Callable) -> Graph:
     """
     Compute the activation bias correction term, and store it in the final activation
     quantization configuration.
@@ -79,6 +80,7 @@ def compute_activation_bias_correction(graph: Graph,
         linear_node: Node to compute the activation bias correction for.
         prev_node: Node to compute the activation error caused by his activation quantization.
         kernel_size: The framework specific attribute name of the convolution layer's kernel size.
+        get_activation_quantization_fn_factory: activation quantization functions factory.
 
     Returns:
         Graph with activation bias correction term for each node.
@@ -105,7 +107,8 @@ def compute_activation_bias_correction(graph: Graph,
     float_centers = calculate_bin_centers(float_bins)
 
     # Quantize the bin edges and calculate the centers of the quantized bins
-    activation_quantizer = get_activation_quantization_fn(prev_node_act_quant_cfg)
+    activation_quantizer = get_activation_quantization_fn(prev_node_act_quant_cfg,
+                                                          get_activation_quantization_fn_factory)
     quant_bins = activation_quantizer(fw_impl.to_tensor(float_bins))
     quant_bins = fw_impl.to_numpy(quant_bins)
     quant_centers = calculate_bin_centers(quant_bins)
@@ -150,7 +153,8 @@ def compute_activation_bias_correction_of_graph(graph: Graph,
                                                 quant_config: QuantizationConfig,
                                                 fw_impl: FrameworkImplementation,
                                                 activation_bias_correction_node_matchers: Callable,
-                                                kernel_size: str) -> Graph:
+                                                kernel_size: str,
+                                                get_activation_quantization_fn_factory: Callable) -> Graph:
     """
     Compute the activation bias correction term for the graph.
 
@@ -160,7 +164,7 @@ def compute_activation_bias_correction_of_graph(graph: Graph,
         fw_impl: FrameworkImplementation object with a specific framework methods implementation.
         activation_bias_correction_node_matchers: Function to match the layers for activation bias correction.
         kernel_size: The framework specific attribute name of the convolution layer's kernel size.
-
+        get_activation_quantization_fn_factory: activation quantization functions factory.
 
     Returns:
         Graph with activation bias correction term for each relevant node.
@@ -176,5 +180,6 @@ def compute_activation_bias_correction_of_graph(graph: Graph,
                                                            fw_impl=fw_impl,
                                                            linear_node=n,
                                                            prev_node=prev_node,
-                                                           kernel_size=kernel_size)
+                                                           kernel_size=kernel_size,
+                                                           get_activation_quantization_fn_factory=get_activation_quantization_fn_factory)
     return graph

model_compression_toolkit/core/common/substitutions/shift_negative_activation.py

@@ -18,8 +18,6 @@
 
 from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig, \
     ActivationQuantizationMode
-from model_compression_toolkit.core.common.quantization.quantization_fn_selection import \
-    get_activation_quantization_fn_factory
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.core.common import Graph, BaseNode
 from model_compression_toolkit.constants import THRESHOLD, SIGNED, SHIFT_NEGATIVE_NON_LINEAR_NUM_BITS
@@ -253,6 +251,7 @@ def shift_negative_function(graph: Graph,
                             padding_str: str,
                             bias_str: str,
                             bias_flag_str: str,
+                            get_activation_quantization_fn_factory: Callable,
                             zero_padding_node: BaseNode = None,
                             bypass_nodes: List = None,
                             params_search_quantization_fn: Callable = None
@@ -278,6 +277,7 @@ def shift_negative_function(graph: Graph,
         padding_str: The framework specific attribute name of the padding.
         bias_str: The framework specific attribute name of the bias.
         bias_flag_str: The framework specific attribute name of the bias flag.
+        get_activation_quantization_fn_factory: activation quantization functions factory.
         zero_padding_node: ZeroPadding2D node that may be in the graph before the linear layer.
         params_search_quantization_fn: Function to quantize np tensor using a framework (tf/torch) quantization method. Needed for better param_search estimating the expected loss.
 
@@ -335,7 +335,7 @@ def shift_negative_function(graph: Graph,
                 the histogram (which is a numpy object) is quantized using the non-linear node activation
                 quantization function (to estimate the expected mse comparing to the original histogram).
                 The quantization function is a framework function, which makes it fail since it
-                expects a fw tensor. The commmon part of SNC receives an argument which is a callable 
+                expects a fw tensor. The common part of SNC receives an argument which is a callable 
                 that receives two argument and returns one: it gets the fw activation quantization function
                 and the bins to quantize. The function (of each fw) responsible for doing (if needed) a preprocessing and postprocessing
                 to the bins which is a numpy object.
@@ -569,6 +569,7 @@ def apply_shift_negative_correction(graph: Graph,
                                     padding_str: str,
                                     bias_str: str,
                                     bias_flag_str: str,
+                                    get_activation_quantization_fn_factory: Callable,
                                     params_search_quantization_fn: Callable=None) -> Graph:
     """
     Apply the substitution even if the linear node is not immediately after
@@ -590,6 +591,9 @@ def apply_shift_negative_correction(graph: Graph,
         padding_str: The framework specific attribute name of the padding.
         bias_str: The framework specific attribute name of the bias.
         bias_flag_str: The framework specific attribute name of the bias flag.
+        get_activation_quantization_fn_factory: activation quantization functions factory.
+        params_search_quantization_fn: Function to quantize np tensor using a framework (tf/torch) quantization method. Needed for better param_search estimating the expected loss.
+
     Returns:
         Graph after applying shift negative on selected activations.
     """
@@ -620,6 +624,7 @@ def apply_shift_negative_correction(graph: Graph,
                                                 padding_str,
                                                 bias_str,
                                                 bias_flag_str,
+                                                get_activation_quantization_fn_factory,
                                                 zero_padding_node=pad_node,
                                                 bypass_nodes=bypass_nodes,
                                                 params_search_quantization_fn=params_search_quantization_fn)

model_compression_toolkit/core/keras/back2framework/quantized_model_builder.py

@@ -14,10 +14,10 @@
 # ==============================================================================
 from typing import List
 
-from model_compression_toolkit.core import FrameworkInfo
 from model_compression_toolkit.core import common
 from model_compression_toolkit.core.common import BaseNode
 from model_compression_toolkit.core.common.quantization.quantization_fn_selection import get_activation_quantization_fn
+from model_compression_toolkit.core.keras.quantization.activation_quantization_fn_factory import get_activation_quantization_fn_factory
 from model_compression_toolkit.core.keras.back2framework.keras_model_builder import KerasModelBuilder
 from tensorflow.python.util.object_identity import Reference as TFReference
 
@@ -57,5 +57,6 @@ def _quantize_node_activations(self,
             Output of the node.
 
         """
-        activation_quantizer = get_activation_quantization_fn(node.final_activation_quantization_cfg)
+        activation_quantizer = get_activation_quantization_fn(node.final_activation_quantization_cfg,
+                                                              get_activation_quantization_fn_factory)
         return activation_quantizer(input_tensors)

model_compression_toolkit/core/keras/default_framework_info.py

@@ -18,19 +18,16 @@
 from typing import Tuple, Any, Dict
 from functools import wraps
 
-from model_compression_toolkit.core.keras.quantizer.lut_fake_quant import activation_lut_kmean_quantizer
 from packaging import version
 
 if version.parse(tf.__version__) >= version.parse("2.13"):
     from keras.src.layers import Conv2D, DepthwiseConv2D, Dense, Conv2DTranspose, Softmax, ELU, Activation
 else:
     from keras.layers import Conv2D, DepthwiseConv2D, Dense, Conv2DTranspose, Softmax, ELU, Activation  # pragma: no cover
 from model_compression_toolkit.core.common.framework_info import FrameworkInfo, set_fw_info, ChannelAxisMapping
-from mct_quantizers import QuantizationMethod
 from model_compression_toolkit.constants import SOFTMAX_THRESHOLD, ACTIVATION
 from model_compression_toolkit.core.keras.constants import SOFTMAX, LINEAR, RELU, SWISH, SIGMOID, IDENTITY, TANH, SELU, \
     KERNEL, DEPTHWISE_KERNEL, GELU
-from model_compression_toolkit.core.keras.quantizer.fake_quant_builder import power_of_two_quantization, symmetric_quantization, uniform_quantization
 
 
 class KerasInfo(FrameworkInfo):
@@ -69,16 +66,6 @@ class KerasInfo(FrameworkInfo):
                                 Dense: -1,
                                 Conv2DTranspose: -1}
 
-    """
-    Mapping from a QuantizationMethod to an activation quantizer function.
-    """
-    activation_quantizer_factory_mapping = {
-        QuantizationMethod.POWER_OF_TWO: power_of_two_quantization,
-        QuantizationMethod.SYMMETRIC: symmetric_quantization,
-        QuantizationMethod.UNIFORM: uniform_quantization,
-        QuantizationMethod.LUT_POT_QUANTIZER: activation_lut_kmean_quantizer
-    }
-
     """
     Map from an activation function name to its min/max output values (if known).
     The values are used for tensor min/max values initialization.

model_compression_toolkit/core/keras/graph_substitutions/substitutions/shift_negative_activation.py

@@ -34,6 +34,7 @@
     NodeFrameworkAttrMatcher
 from model_compression_toolkit.core.common.substitutions.shift_negative_activation import \
     apply_shift_negative_correction
+from model_compression_toolkit.core.keras.quantization.activation_quantization_fn_factory import get_activation_quantization_fn_factory
 from model_compression_toolkit.core.keras.constants import KERNEL_SIZE, STRIDES, ACTIVATION, SWISH, \
     SELU, GELU, FUNCTION, ADD, PAD
 from model_compression_toolkit.core.keras.constants import NEGATIVE_SLOPE, PADDING, PAD_SAME, PAD_VALID, BIAS, USE_BIAS
@@ -252,5 +253,6 @@ def keras_apply_shift_negative_correction(graph: Graph,
                                            is_padding_node_and_node_has_padding,
                                            PADDING,
                                            BIAS,
-                                           USE_BIAS
+                                           USE_BIAS,
+                                           get_activation_quantization_fn_factory
                                            )

model_compression_toolkit/core/keras/mixed_precision/configurable_activation_quantizer.py

@@ -23,6 +23,7 @@
     verify_candidates_descending_order, init_activation_quantizers
 from model_compression_toolkit.core.common.quantization.candidate_node_quantization_config import \
     CandidateNodeQuantizationConfig
+from model_compression_toolkit.core.keras.quantization.activation_quantization_fn_factory import get_activation_quantization_fn_factory
 from model_compression_toolkit.logger import Logger
 
 import tensorflow as tf
@@ -67,7 +68,7 @@ def __init__(self,
             if qc.activation_quantization_cfg.quant_mode != node_q_cfg[0].activation_quantization_cfg.quant_mode:
                 Logger.critical("Unsupported configuration: Mixing candidates with differing activation quantization states (enabled/disabled).")  # pragma: no cover
 
-        self.activation_quantizers = init_activation_quantizers(self.node_q_cfg)
+        self.activation_quantizers = init_activation_quantizers(self.node_q_cfg, get_activation_quantization_fn_factory)
         self.active_quantization_config_index = max_candidate_idx  # initialize with first config as default
 
     def set_active_activation_quantizer(self, index: Optional[int]):