Add 16 bit input\output to stack & concat operators in TPC.v4 (#1158)

Add 16 bit input\output to stack & concat operators in TPC.v4
sony · Aug 12, 2024 · dc28638 · dc28638
1 parent 27b0742
commit dc28638
Show file tree

Hide file tree

Showing 3 changed files with 6 additions and 2 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
@@ -199,6 +199,7 @@ def generate_tp_model(default_config: OpQuantizationConfig,
                         default_qco.clone_and_edit(enable_activation_quantization=False,
                                                    supported_input_activation_n_bits=(8, 16))
                         .clone_and_edit_weight_attribute(enable_weights_quantization=False))
+        tp.OperatorsSet("Default16BitInout", const_configuration_options_inout16)
 
         # Create Mixed-Precision quantization configuration options from the given list of OpQuantizationConfig objects
         mixed_precision_configuration_options = tp.QuantizationConfigOptions(mixed_precision_cfg_list,

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
@@ -26,11 +26,11 @@
 if version.parse(tf.__version__) >= version.parse("2.13"):
     from keras.src.layers import Conv2D, DepthwiseConv2D, Dense, Reshape, ZeroPadding2D, Dropout, \
         MaxPooling2D, Activation, ReLU, Add, Subtract, Multiply, PReLU, Flatten, Cropping2D, LeakyReLU, Permute, \
-        Conv2DTranspose, Identity
+        Conv2DTranspose, Identity, Concatenate
 else:
     from keras.layers import Conv2D, DepthwiseConv2D, Dense, Reshape, ZeroPadding2D, Dropout, \
         MaxPooling2D, Activation, ReLU, Add, Subtract, Multiply, PReLU, Flatten, Cropping2D, LeakyReLU, Permute, \
-        Conv2DTranspose, Identity
+        Conv2DTranspose, Identity, Concatenate
 
 from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v4.tp_model import get_tp_model
 import model_compression_toolkit as mct
@@ -93,6 +93,8 @@ def generate_keras_tpc(name: str, tp_model: tp.TargetPlatformModel):
 
     with keras_tpc:
         tp.OperationsSetToLayers("NoQuantization", no_quant_list)
+        tp.OperationsSetToLayers("Default16BitInout", [tf.stack,
+                                                       tf.concat, Concatenate])
         tp.OperationsSetToLayers("Conv",
                                  [Conv2D,
                                   DepthwiseConv2D,

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
@@ -85,6 +85,7 @@ def generate_pytorch_tpc(name: str, tp_model: tp.TargetPlatformModel):
                                                     topk,
                                                     squeeze,
                                                     MaxPool2d])
+        tp.OperationsSetToLayers("Default16BitInout", [torch.stack, torch.cat])
 
         tp.OperationsSetToLayers("Conv", [Conv2d, ConvTranspose2d],
                                  attr_mapping=pytorch_linear_attr_mapping)