Add new TPCv6.0 (torch.take) (#1570)

Author: kawakami-masaki0

model_compression_toolkit/target_platform_capabilities/schema/v2.py

@@ -93,6 +93,7 @@ class OperatorSetNames(str, Enum):
     EXP = "Exp"
     SIN = "Sin"
     COS = "Cos"
+    TAKE = "Take"
 
     @classmethod
     def get_values(cls):

model_compression_toolkit/target_platform_capabilities/targetplatform2framework/attach2keras.py

@@ -103,7 +103,8 @@ def __init__(self):
             OperatorSetNames.SSD_POST_PROCESS: [SSDPostProcess],
             OperatorSetNames.EXP: [tf.math.exp],
             OperatorSetNames.SIN: [tf.math.sin],
-            OperatorSetNames.COS: [tf.math.cos]
+            OperatorSetNames.COS: [tf.math.cos],
+            OperatorSetNames.TAKE: [],  # no such operator in tensorflow
         }
 
         self._opset2attr_mapping = {

model_compression_toolkit/target_platform_capabilities/targetplatform2framework/attach2pytorch.py

@@ -102,6 +102,7 @@ def __init__(self):
             OperatorSetNames.EXP: [torch.exp],
             OperatorSetNames.SIN: [torch.sin],
             OperatorSetNames.COS: [torch.cos],
+            OperatorSetNames.TAKE: [torch.take],
         }
 
         pytorch_linear_attr_mapping = {KERNEL_ATTR: DefaultDict(default_value=PYTORCH_KERNEL),

tests_pytest/_fw_tests_common_base/base_tpc_attach2fw_test.py

@@ -51,7 +51,7 @@ def setup_method(self):
 
     def test_attach2fw_init(self):
         # verify built-in opset to operator mapping structure
-        assert len(self.attach2fw._opset2layer) == 60  # number of built-in operator sets
+        assert len(self.attach2fw._opset2layer) == 61  # number of built-in operator sets
         assert all(opset in self.attach2fw._opset2layer for opset in list(schema.OperatorSetNames))
         assert all(isinstance(key, schema.OperatorSetNames) for key in self.attach2fw._opset2layer.keys())
         assert all(isinstance(value, list) for value in self.attach2fw._opset2layer.values())

tests_pytest/pytorch_tests/e2e_tests/test_take.py

@@ -0,0 +1,87 @@
+# Copyright 2025 Sony Semiconductor Solutions, 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 Iterator, List
+import torch
+import torch.nn as nn
+import model_compression_toolkit as mct
+from model_compression_toolkit.target_platform_capabilities import AttributeQuantizationConfig
+import model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema as schema
+from mct_quantizers import QuantizationMethod
+
+
+class TakeModel(nn.Module):
+
+    def __init__(self, indices):
+        super().__init__()
+        self.conv = nn.Conv2d(3, 16, kernel_size=3)
+        self.relu = nn.ReLU()
+        self.indices = torch.as_tensor(indices, dtype=torch.long)
+
+    def forward(self, x):
+        x = self.relu(self.conv(x))
+        output = torch.take(x, self.indices)
+        return output
+
+
+def get_representative_dataset(n_iter=1):
+
+    def representative_dataset() -> Iterator[List]:
+        for _ in range(n_iter):
+            yield [torch.randn(1, 3, 32, 32)]
+    return representative_dataset
+
+
+def get_edgemdt_tpc_v6():
+
+    default_config = schema.OpQuantizationConfig(
+        default_weight_attr_config=AttributeQuantizationConfig(),
+        attr_weights_configs_mapping={},
+        activation_quantization_method=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,
+        signedness=schema.Signedness.AUTO)
+
+    default_configuration_options = schema.QuantizationConfigOptions(quantization_configurations=tuple([default_config]))
+    dim_manipulation_config = (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 = []
+    operator_set.append(schema.OperatorsSet(name=schema.OperatorSetNames.TAKE, qc_options=dim_manipulation_config))
+    operator_set.append(schema.OperatorsSet(name=schema.OperatorSetNames.CONV, qc_options=default_configuration_options))
+    operator_set.append(schema.OperatorsSet(name=schema.OperatorSetNames.RELU, qc_options=default_configuration_options))
+
+    tpc = schema.TargetPlatformCapabilities(
+        default_qco=default_configuration_options,
+        operator_set=tuple(operator_set))
+    return tpc
+
+
+def test_take():
+
+    model = TakeModel(indices=[0, 100])
+    tpc = get_edgemdt_tpc_v6() # TPC equivalent to edgemdt-tpc v6.0
+    quantized_model, _ = mct.ptq.pytorch_post_training_quantization(model, 
+                                                                    get_representative_dataset(n_iter=1),
+                                                                    target_resource_utilization=None,
+                                                                    core_config=mct.core.CoreConfig(),
+                                                                    target_platform_capabilities=tpc)
+    assert hasattr(quantized_model, 'take')
+    assert not hasattr(quantized_model, 'take_activation_holder_quantizer')