Fix tpcv4 torch (#1150)

* Add "multiply" & "subtract" torch functions to TPCv4.
* Set all OpQuantizationConfig inputs as args without default. Change is_signed to signedness with Enum.

model_compression_toolkit/core/common/quantization/node_quantization_config.py

@@ -96,7 +96,7 @@ def __init__(self,
         self.activation_n_bits = op_cfg.activation_n_bits
         self.relu_bound_to_power_of_2 = qc.relu_bound_to_power_of_2
         self.enable_activation_quantization = op_cfg.enable_activation_quantization
-        self.is_signed = op_cfg.is_signed
+        self.signedness = op_cfg.signedness
         self.activation_channel_equalization = qc.activation_channel_equalization
         self.input_scaling = qc.input_scaling
         self.min_threshold = qc.min_threshold

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

@@ -15,9 +15,8 @@
 import numpy as np
 from typing import Dict, Union
 
-from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
+from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod, Signedness
 from model_compression_toolkit.core.common.collectors.statistics_collector import BaseStatsCollector
-from model_compression_toolkit.constants import SIGNED
 from model_compression_toolkit.core.common.quantization import quantization_params_generation
 from model_compression_toolkit.core.common.node_prior_info import NodePriorInfo
 from model_compression_toolkit.core.common.quantization.node_quantization_config import NodeActivationQuantizationConfig
@@ -49,8 +48,8 @@ def get_activations_qparams(activation_quant_cfg: NodeActivationQuantizationConf
                                                                     bins_counts)
     min_value, max_value = out_stats_container.get_min_max_values()
 
-    if activation_quant_cfg.is_signed is not None:
-        signed = activation_quant_cfg.is_signed
+    if activation_quant_cfg.signedness in [Signedness.SIGNED, Signedness.UNSIGNED]:
+        signed = activation_quant_cfg.signedness == Signedness.SIGNED
     elif nodes_prior_info.is_output_bounded():
         signed = min_value < 0
     else:

model_compression_toolkit/target_platform_capabilities/target_platform/__init__.py

@@ -17,7 +17,8 @@
 from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework.attribute_filter import AttributeFilter
 from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework import TargetPlatformCapabilities, OperationsSetToLayers, Smaller, SmallerEq, NotEq, Eq, GreaterEq, Greater, LayerFilterParams, OperationsToLayers, get_current_tpc
 from model_compression_toolkit.target_platform_capabilities.target_platform.target_platform_model import get_default_quantization_config_options, TargetPlatformModel
-from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import OpQuantizationConfig, QuantizationConfigOptions, AttributeQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
+    OpQuantizationConfig, QuantizationConfigOptions, AttributeQuantizationConfig, Signedness
 from model_compression_toolkit.target_platform_capabilities.target_platform.operators import OperatorsSet, OperatorSetConcat
 
 from mct_quantizers import QuantizationMethod

model_compression_toolkit/target_platform_capabilities/target_platform/op_quantization_config.py

@@ -15,12 +15,26 @@
 
 import copy
 from typing import List, Dict, Union, Any, Tuple
+from enum import Enum
 
 from mct_quantizers import QuantizationMethod
 from model_compression_toolkit.constants import FLOAT_BITWIDTH
 from model_compression_toolkit.logger import Logger
 
 
+class Signedness(Enum):
+    """
+    An enum for choosing the signedness of the quantization method:
+
+    AUTO - Signedness decided automatically by quantization.
+    SIGNED - Force signed quantization.
+    UNSIGNED - Force unsigned quantization.
+    """
+    AUTO = 0
+    SIGNED = 1
+    UNSIGNED = 2
+
+
 def clone_and_edit_object_params(obj: Any, **kwargs: Dict) -> Any:
     """
     Clones the given object and edit some of its parameters.
@@ -120,7 +134,7 @@ def __init__(self,
                  fixed_scale: float,
                  fixed_zero_point: int,
                  simd_size: int,
-                 is_signed: bool = None
+                 signedness: Signedness
                  ):
         """
 
@@ -134,8 +148,8 @@ def __init__(self,
             quantization_preserving (bool): Whether quantization parameters should be the same for an operator's input and output.
             fixed_scale (float): Scale to use for an operator quantization parameters.
             fixed_zero_point (int): Zero-point to use for an operator quantization parameters.
-            is_signed (bool): Force activation quantization signedness (None means don't force).
             simd_size (int): Per op integer representing the Single Instruction, Multiple Data (SIMD) width of an operator. It indicates the number of data elements that can be fetched and processed simultaneously in a single instruction.
+            signedness (bool): Set activation quantization signedness.
 
         """
 
@@ -154,7 +168,7 @@ def __init__(self,
         self.quantization_preserving = quantization_preserving
         self.fixed_scale = fixed_scale
         self.fixed_zero_point = fixed_zero_point
-        self.is_signed = is_signed
+        self.signedness = signedness
         self.simd_size = simd_size
 
     def get_info(self):
@@ -206,7 +220,7 @@ def __eq__(self, other):
             self.activation_n_bits == other.activation_n_bits and \
             self.supported_input_activation_n_bits == other.supported_input_activation_n_bits and \
             self.enable_activation_quantization == other.enable_activation_quantization and \
-            self.is_signed == other.is_signed and \
+            self.signedness == other.signedness and \
             self.simd_size == other.simd_size
 
     @property

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v1/tp_model.py

@@ -18,7 +18,7 @@
 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
+    TargetPlatformModel, Signedness
 from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
     AttributeQuantizationConfig
 
@@ -98,7 +98,8 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # We define an 8-bit config for linear operations quantization, that include a kernel and bias attributes.
     linear_eight_bits = tp.OpQuantizationConfig(
@@ -111,7 +112,8 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # To quantize a model using mixed-precision, create
     # a list with more than one OpQuantizationConfig.

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v1_lut/tp_model.py

@@ -19,7 +19,7 @@
 from model_compression_toolkit.target_platform_capabilities.constants import KERNEL_ATTR, BIAS_ATTR, WEIGHTS_N_BITS, \
     WEIGHTS_QUANTIZATION_METHOD
 from model_compression_toolkit.target_platform_capabilities.target_platform import OpQuantizationConfig, \
-    TargetPlatformModel
+    TargetPlatformModel, Signedness
 from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
     AttributeQuantizationConfig
 
@@ -94,7 +94,8 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # We define an 8-bit config for linear operations quantization, that include a kernel and bias attributes.
     linear_eight_bits = tp.OpQuantizationConfig(
@@ -107,7 +108,8 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # To quantize a model using mixed-precision, create
     # a list with more than one OpQuantizationConfig.

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v1_pot/tp_model.py

@@ -18,7 +18,7 @@
 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
+    TargetPlatformModel, Signedness
 from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
     AttributeQuantizationConfig
 
@@ -94,7 +94,8 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # We define an 8-bit config for linear operations quantization, that include a kernel and bias attributes.
     linear_eight_bits = tp.OpQuantizationConfig(
@@ -107,7 +108,8 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # To quantize a model using mixed-precision, create
     # a list with more than one OpQuantizationConfig.

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v2/tp_model.py

@@ -18,7 +18,7 @@
 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
+    TargetPlatformModel, Signedness
 from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
     AttributeQuantizationConfig
 
@@ -100,7 +100,8 @@ def get_op_quantization_configs() -> \
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # We define an 8-bit config for linear operations quantization, that include a kernel and bias attributes.
     linear_eight_bits = tp.OpQuantizationConfig(
@@ -113,7 +114,8 @@ def get_op_quantization_configs() -> \
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # To quantize a model using mixed-precision, create
     # a list with more than one OpQuantizationConfig.

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v2_lut/tp_model.py

@@ -19,7 +19,7 @@
 from model_compression_toolkit.target_platform_capabilities.constants import KERNEL_ATTR, BIAS_ATTR, WEIGHTS_N_BITS, \
     WEIGHTS_QUANTIZATION_METHOD
 from model_compression_toolkit.target_platform_capabilities.target_platform import OpQuantizationConfig, \
-    TargetPlatformModel
+    TargetPlatformModel, Signedness
 from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
     AttributeQuantizationConfig
 
@@ -96,7 +96,8 @@ def get_op_quantization_configs() -> \
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # We define an 8-bit config for linear operations quantization, that include a kernel and bias attributes.
     linear_eight_bits = tp.OpQuantizationConfig(
@@ -109,7 +110,8 @@ def get_op_quantization_configs() -> \
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # To quantize a model using mixed-precision, create
     # a list with more than one OpQuantizationConfig.

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v3/tp_model.py

@@ -18,7 +18,7 @@
 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
+    TargetPlatformModel, Signedness
 from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
     AttributeQuantizationConfig
 
@@ -100,7 +100,8 @@ def get_op_quantization_configs() -> \
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # We define an 8-bit config for linear operations quantization, that include a kernel and bias attributes.
     linear_eight_bits = tp.OpQuantizationConfig(
@@ -113,7 +114,8 @@ def get_op_quantization_configs() -> \
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # To quantize a model using mixed-precision, create
     # a list with more than one OpQuantizationConfig.

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

@@ -19,7 +19,7 @@
 from model_compression_toolkit.target_platform_capabilities.constants import KERNEL_ATTR, BIAS_ATTR, WEIGHTS_N_BITS, \
     WEIGHTS_QUANTIZATION_METHOD
 from model_compression_toolkit.target_platform_capabilities.target_platform import OpQuantizationConfig, \
-    TargetPlatformModel
+    TargetPlatformModel, Signedness
 from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
     AttributeQuantizationConfig
 
@@ -96,7 +96,8 @@ def get_op_quantization_configs() -> \
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # We define an 8-bit config for linear operations quantization, that include a kernel and bias attributes.
     linear_eight_bits = tp.OpQuantizationConfig(
@@ -109,7 +110,8 @@ def get_op_quantization_configs() -> \
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # To quantize a model using mixed-precision, create
     # a list with more than one OpQuantizationConfig.

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

@@ -18,7 +18,7 @@
 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
+    TargetPlatformModel, Signedness
 from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
     AttributeQuantizationConfig
 
@@ -100,7 +100,8 @@ def get_op_quantization_configs() -> \
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # We define an 8-bit config for linear operations quantization, that include a kernel and bias attributes.
     linear_eight_bits = tp.OpQuantizationConfig(
@@ -113,7 +114,8 @@ def get_op_quantization_configs() -> \
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # To quantize a model using mixed-precision, create
     # a list with more than one OpQuantizationConfig.
@@ -170,7 +172,7 @@ def generate_tp_model(default_config: OpQuantizationConfig,
     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)
+        activation_n_bits=16, signedness=Signedness.SIGNED)
     const_configuration_options_inout16 = tp.QuantizationConfigOptions([const_config_input16_output16,
                                                                         const_config_input16],
                                                                        base_config=const_config_input16)

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

@@ -17,7 +17,7 @@
 
 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
+    gather, equal, transpose, permute, argmax, squeeze, multiply, subtract
 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
@@ -101,8 +101,8 @@ def generate_pytorch_tpc(name: str, tp_model: tp.TargetPlatformModel):
                                              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("Sub", [operator.sub, sub, subtract])
+        tp.OperationsSetToLayers("Mul", [operator.mul, mul, multiply])
         tp.OperationsSetToLayers("Div", [operator.truediv, div])
         tp.OperationsSetToLayers("PReLU", [PReLU, prelu])
         tp.OperationsSetToLayers("Swish", [SiLU, silu, Hardswish, hardswish])

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

@@ -18,7 +18,7 @@
 from model_compression_toolkit.constants import FLOAT_BITWIDTH
 from model_compression_toolkit.target_platform_capabilities.constants import KERNEL_ATTR, BIAS_ATTR
 from model_compression_toolkit.target_platform_capabilities.target_platform import OpQuantizationConfig, \
-    TargetPlatformModel
+    TargetPlatformModel, Signedness
 from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
     AttributeQuantizationConfig
 
@@ -95,7 +95,8 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=32)
+        simd_size=32,
+        signedness=Signedness.AUTO)
 
     # We define an 8-bit config for linear operations quantization, that include a kernel and bias attributes.
     linear_eight_bits = tp.OpQuantizationConfig(
@@ -108,8 +109,8 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         quantization_preserving=False,
         fixed_scale=None,
         fixed_zero_point=None,
-        simd_size=None
-    )
+        simd_size=None,
+        signedness=Signedness.AUTO)
 
     mixed_precision_cfg_list = []  # No mixed precision