Add hardsigmoid & hardswish

sony · Nov 19, 2024 · e1c1459 · e1c1459
1 parent 3d86447
commit e1c1459
Show file tree

Hide file tree

Showing 4 changed files with 16 additions and 4 deletions.
diff --git a/model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v4/tp_model.py b/model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v4/tp_model.py
@@ -43,6 +43,8 @@
 OPSET_SIGMOID = "Sigmoid"
 OPSET_TANH = "Tanh"
 OPSET_GELU = "Gelu"
+OPSET_HARDSIGMOID = "HardSigmoid"
+OPSET_HARDSWISH = "HardSwish"
 
 
 def get_tp_model() -> TargetPlatformModel:
@@ -278,12 +280,16 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         sigmoid = tp.OperatorsSet(OPSET_SIGMOID, default_config_options_16bit)
         tanh = tp.OperatorsSet(OPSET_TANH, default_config_options_16bit)
         gelu = tp.OperatorsSet(OPSET_GELU, default_config_options_16bit)
+        hardsigmoid = tp.OperatorsSet(OPSET_HARDSIGMOID, default_config_options_16bit)
+        hardswish = tp.OperatorsSet(OPSET_HARDSWISH, default_config_options_16bit)
 
         # 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, gelu)
-        activations_after_fc_to_fuse = tp.OperatorSetConcat(any_relu, swish, sigmoid, tanh, gelu)
+        activations_after_conv_to_fuse = tp.OperatorSetConcat(any_relu, swish, prelu, sigmoid,
+                                                              tanh, gelu, hardswish, hardsigmoid)
+        activations_after_fc_to_fuse = tp.OperatorSetConcat(any_relu, swish, sigmoid, tanh, gelu,
+                                                            hardswish, hardsigmoid)
         any_binary = tp.OperatorSetConcat(add, sub, mul, div)
 
         # ------------------- #

diff --git a/model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v4/tpc_keras.py b/model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v4/tpc_keras.py
@@ -38,7 +38,7 @@
 from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v4.tp_model import OPSET_NO_QUANTIZATION, \
     OPSET_QUANTIZATION_PRESERVING, OPSET_DIMENSION_MANIPULATION_OPS_WITH_WEIGHTS, OPSET_DIMENSION_MANIPULATION_OPS, \
     OPSET_MERGE_OPS, OPSET_CONV, OPSET_FULLY_CONNECTED, OPSET_ANY_RELU, OPSET_ADD, OPSET_SUB, OPSET_MUL, OPSET_DIV, \
-    OPSET_PRELU, OPSET_SWISH, OPSET_SIGMOID, OPSET_TANH, OPSET_GELU, OPSET_BATCH_NORM, OPSET_MIN_MAX
+    OPSET_PRELU, OPSET_SWISH, OPSET_SIGMOID, OPSET_TANH, OPSET_GELU, OPSET_BATCH_NORM, OPSET_MIN_MAX, OPSET_HARDSIGMOID
 
 tp = mct.target_platform
 
@@ -135,5 +135,7 @@ def generate_keras_tpc(name: str, tp_model: tp.TargetPlatformModel):
         tp.OperationsSetToLayers(OPSET_SIGMOID, [tf.nn.sigmoid, tp.LayerFilterParams(Activation, activation="sigmoid")])
         tp.OperationsSetToLayers(OPSET_TANH, [tf.nn.tanh, tp.LayerFilterParams(Activation, activation="tanh")])
         tp.OperationsSetToLayers(OPSET_GELU, [tf.nn.gelu, tp.LayerFilterParams(Activation, activation="gelu")])
+        tp.OperationsSetToLayers(OPSET_HARDSIGMOID, [tf.keras.activations.hard_sigmoid,
+                                                     tp.LayerFilterParams(Activation, activation="hard_sigmoid")])
 
     return keras_tpc
diff --git a/..._compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v4/tpc_pytorch.py b/..._compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v4/tpc_pytorch.py
@@ -34,7 +34,8 @@
 from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v4.tp_model import OPSET_NO_QUANTIZATION, \
     OPSET_QUANTIZATION_PRESERVING, OPSET_DIMENSION_MANIPULATION_OPS_WITH_WEIGHTS, OPSET_DIMENSION_MANIPULATION_OPS, \
     OPSET_MERGE_OPS, OPSET_CONV, OPSET_FULLY_CONNECTED, OPSET_ANY_RELU, OPSET_ADD, OPSET_SUB, OPSET_MUL, OPSET_DIV, \
-    OPSET_PRELU, OPSET_SWISH, OPSET_SIGMOID, OPSET_TANH, OPSET_GELU, OPSET_BATCH_NORM, OPSET_MIN_MAX
+    OPSET_PRELU, OPSET_SWISH, OPSET_SIGMOID, OPSET_TANH, OPSET_GELU, OPSET_BATCH_NORM, OPSET_MIN_MAX, OPSET_HARDSIGMOID, \
+    OPSET_HARDSWISH
 
 tp = mct.target_platform
 
@@ -118,5 +119,7 @@ def generate_pytorch_tpc(name: str, tp_model: tp.TargetPlatformModel):
         tp.OperationsSetToLayers(OPSET_SIGMOID, [Sigmoid, sigmoid, F.sigmoid])
         tp.OperationsSetToLayers(OPSET_TANH, [Tanh, tanh, F.tanh])
         tp.OperationsSetToLayers(OPSET_GELU, [GELU, gelu])
+        tp.OperationsSetToLayers(OPSET_HARDSIGMOID, [torch.nn.Hardsigmoid, torch.nn.functional.hardsigmoid])
+        tp.OperationsSetToLayers(OPSET_HARDSWISH, [torch.nn.Hardswish, torch.nn.functional.hardswish])
 
     return pytorch_tpc
diff --git a/tests/keras_tests/layer_tests/test_layers_runner.py b/tests/keras_tests/layer_tests/test_layers_runner.py
@@ -30,6 +30,7 @@ def test_activation(self):
         BaseKerasLayerTest(self,
                            [Activation('linear'),
                             Activation('hard_sigmoid'),
+                            tf.keras.activations.hard_sigmoid,
                             Activation('exponential')]).run_test()
 
     def test_softplus(self):