Add pytorch substitution to remove Identity layers (#1043)

* Add pytorch substitution to remove Identity layers.
* Adapt tests that used identity layers.

---------

Co-authored-by: reuvenp <[email protected]>

model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/remove_identity.py

@@ -0,0 +1,66 @@
+# 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 torch import reshape
+import torch
+
+from model_compression_toolkit.logger import Logger
+from model_compression_toolkit.core.common.graph.graph_matchers import NodeOperationMatcher
+from model_compression_toolkit.core import common
+from model_compression_toolkit.core.common.graph.base_graph import Graph
+from model_compression_toolkit.core.common.graph.base_node import BaseNode
+from model_compression_toolkit.core.pytorch.constants import BATCH_DIM_VALUE
+
+
+class RemoveIdentity(common.BaseSubstitution):
+    """
+    Remove `torch.nn.Identity` layers from the graph.
+    """
+
+    def __init__(self):
+        nodes = NodeOperationMatcher(torch.nn.Identity)
+        super().__init__(matcher_instance=nodes)
+
+    def substitute(self,
+                   graph: Graph,
+                   node: BaseNode) -> Graph:
+        """
+        The method to perform the substitution of the `torch.nn.Identity` node by
+        reconnecting its input directly to its output, effectively removing the node
+        from the graph.
+
+        Args:
+            graph: The current graph of operations where the node resides.
+            node: The specific `BaseNode` that is matched to be an Identity operation.
+
+        Returns:
+            Graph: The updated graph after removing the identity node.
+        """
+
+        # Retrieve the predecessor nodes of the identity node.
+        prev_identity_nodes = graph.get_prev_nodes(node)
+        # Ensure there is exactly one predecessor; otherwise, do nothing.
+        if len(prev_identity_nodes) != 1:
+            return graph
+
+        # Reconnect the output edges of the identity node to its predecessor,
+        # effectively bypassing the identity node.
+        graph.reconnect_out_edges(current_node=node, new_node=prev_identity_nodes[0])
+        # Remove the edge from the predecessor to the identity node.
+        graph.remove_edge(prev_identity_nodes[0], node)
+        # Remove the identity node from the graph.
+        graph.remove_node(node_to_remove=node)
+
+        return graph
+

model_compression_toolkit/core/pytorch/pytorch_implementation.py

@@ -58,6 +58,7 @@
     FunctionalConvSubstitution
 from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.relu_bound_to_power_of_2 import \
     ReLUBoundToPowerOfTwo
+from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.remove_identity import RemoveIdentity
 from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.reshape_with_static_shapes import \
     ReshapeWithStaticShapes
 from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.residual_collapsing import \
@@ -238,7 +239,8 @@ def get_substitutions_prepare_graph(self, fw_info: FrameworkInfo = None) -> List
                 PermuteCallMethod(),
                 FunctionalConvSubstitution(fw_info),
                 FunctionalBatchNorm(),
-                FunctionalLayerNorm()]
+                FunctionalLayerNorm(),
+                RemoveIdentity()]
 
     def get_substitutions_pre_statistics_collection(self,
                                                     quant_config: QuantizationConfig

tests/pytorch_tests/function_tests/test_activation_quantization_holder_gptq.py

@@ -1,3 +1,5 @@
+import copy
+
 import unittest
 import torch
 from mct_quantizers import PytorchActivationQuantizationHolder, PytorchQuantizationWrapper
@@ -23,11 +25,9 @@ def __init__(self, num_channels=3, kernel_size=1):
         super(BasicModel, self).__init__()
         self.conv1 = Conv2d(num_channels, num_channels, kernel_size=kernel_size, bias=False)
         self.conv2 = Conv2d(num_channels, num_channels, kernel_size=kernel_size, bias=False)
-        self.identity = torch.nn.Identity()
 
     def forward(self, inp):
         x = self.conv1(inp)
-        x = self.identity(x)
         x = self.conv2(x)
         return x
 
@@ -51,11 +51,9 @@ class ReuseModel(torch.nn.Module):
     def __init__(self, num_channels=3, kernel_size=1):
         super(ReuseModel, self).__init__()
         self.conv = Conv2d(num_channels, num_channels, kernel_size=kernel_size, bias=False)
-        self.identity = torch.nn.Identity()
 
     def forward(self, inp):
         x = self.conv(inp)
-        x = self.identity(x)
         x = self.conv(x)
         return x
 
@@ -73,7 +71,7 @@ def test_adding_holder_instead_quantize_wrapper(self):
         # the last module should be an activation quantization holder
         self.assertTrue(isinstance(last_module, PytorchActivationQuantizationHolder))
         # check that 4 activation quantization holders where generated
-        self.assertTrue(len(activation_quantization_holders_in_model) == 4)
+        self.assertTrue(len(activation_quantization_holders_in_model) == 3)
         for a in activation_quantization_holders_in_model:
             self.assertTrue(isinstance(a.activation_holder_quantizer, ActivationPOTInferableQuantizer))
         for name, module in gptq_model.named_modules():
@@ -102,7 +100,7 @@ def test_adding_holders_after_reuse(self):
         # the last module should be an activation quantization holder
         self.assertTrue(isinstance(last_module, PytorchActivationQuantizationHolder))
         # check that 4 activation quantization holders where generated
-        self.assertTrue(len(activation_quantization_holders_in_model) == 4)
+        self.assertTrue(len(activation_quantization_holders_in_model) == 3)
         for a in activation_quantization_holders_in_model:
             self.assertTrue(isinstance(a.activation_holder_quantizer, ActivationPOTInferableQuantizer))
         for name, module in gptq_model.named_modules():
@@ -115,13 +113,16 @@ def test_adding_holders_after_reuse(self):
 
     def _get_gptq_model(self, input_shape, in_model):
         pytorch_impl = GPTQPytorchImplemantation()
+        qc = copy.deepcopy(mct.core.DEFAULTCONFIG)
+        qc.linear_collapsing = False
         graph = prepare_graph_with_quantization_parameters(in_model,
                                                            pytorch_impl,
                                                            DEFAULT_PYTORCH_INFO,
                                                            representative_dataset,
                                                            generate_pytorch_tpc,
                                                            [1] + input_shape,
-                                                           mixed_precision_enabled=False)
+                                                           mixed_precision_enabled=False,
+                                                           qc=qc)
         graph = set_bit_widths(mixed_precision_enable=False,
                                graph=graph)
         trainer = PytorchGPTQTrainer(graph,

tests/pytorch_tests/model_tests/feature_models/output_in_the_middle_test.py

@@ -28,7 +28,7 @@ def __init__(self):
 
     def forward(self, x):
         x1 = self.conv1(x)
-        x2 = self.identity(x1)
+        x2 = torch.relu(x1)
         x3 = self.conv2(x2)
         x4 = torch.relu(x3)
         return x, x1, x2, x3, x4

tests/pytorch_tests/model_tests/feature_models/relu_bound_test.py

@@ -38,13 +38,12 @@ def __init__(self):
         self.conv3 = Conv2d(3, 3, kernel_size=1, stride=1)
         self.conv4 = Conv2d(3, 3, kernel_size=1, stride=1)
         self.relu2 = ReLU()
-        self.identity = torch.nn.Identity()
 
     def forward(self, inp):
         x = self.conv1(inp)
         x = self.relu1(x)
         x = self.conv2(x)
-        x = self.identity(x)
+        x = relu(x)
         x = self.conv3(x)
         x = relu6(x)
         x = self.conv4(x)

tests/pytorch_tests/model_tests/feature_models/remove_identity_test.py

@@ -0,0 +1,50 @@
+# 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 torch
+
+from tests.pytorch_tests.model_tests.base_pytorch_feature_test import BasePytorchFeatureNetworkTest
+
+
+class RemoveIdentityNet(torch.nn.Module):
+    def __init__(self):
+        super(RemoveIdentityNet, self).__init__()
+        self.conv1 = torch.nn.Conv2d(3, 3, kernel_size=1, stride=1)
+        self.identity = torch.nn.Identity()
+        self.bn1 = torch.nn.BatchNorm2d(3)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.identity(x)
+        x = self.bn1(x)
+        return x
+
+
+class RemoveIdentityTest(BasePytorchFeatureNetworkTest):
+
+    def __init__(self, unit_test):
+        super().__init__(unit_test)
+
+    def create_networks(self):
+        return RemoveIdentityNet()
+
+    def compare(self,
+                quantized_model,
+                float_model,
+                input_x=None,
+                quantization_info=None):
+        for n,m in quantized_model.named_modules():
+            # make sure identity was removed and bn was folded into the conv
+            self.unit_test.assertFalse(isinstance(m, torch.nn.Identity) or isinstance(m, torch.nn.BatchNorm2d))
+

tests/pytorch_tests/model_tests/test_feature_models_runner.py

@@ -89,9 +89,15 @@
 from tests.pytorch_tests.model_tests.feature_models.const_representation_test import ConstRepresentationTest, \
     ConstRepresentationMultiInputTest
 from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
+from tests.pytorch_tests.model_tests.feature_models.remove_identity_test import RemoveIdentityTest
 
 
 class FeatureModelsTestRunner(unittest.TestCase):
+    def test_remove_identity(self):
+        """
+        This test checks that identity layers are removed from the model.
+        """
+        RemoveIdentityTest(self).run_test()
 
     def test_single_layer_replacement(self):
         """