Remove TPC IMX500v3 and rename TPC IMX500v4 into IMX500v3 (#1181)

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/target_platform_capabilities.py

@@ -42,8 +42,6 @@ def get_tpc_dict_by_fw(fw_name):
                 get_keras_tpc as get_keras_tpc_v3
             from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v3_lut.tpc_keras import \
                 get_keras_tpc as get_keras_tpc_v3_lut
-            from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v4.tpc_keras import \
-                get_keras_tpc as get_keras_tpc_v4
 
             # Keras: TPC versioning
             tpc_models_dict = {'v1': get_keras_tpc_v1,
@@ -53,7 +51,6 @@ def get_tpc_dict_by_fw(fw_name):
                                'v2_lut': get_keras_tpc_v2_lut,
                                'v3': get_keras_tpc_v3,
                                'v3_lut': get_keras_tpc_v3_lut,
-                               'v4': get_keras_tpc_v4,
                                LATEST: get_keras_tpc_latest}
     elif fw_name == PYTORCH:
         ###############################
@@ -76,8 +73,6 @@ def get_tpc_dict_by_fw(fw_name):
                 get_pytorch_tpc as get_pytorch_tpc_v3
             from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v3_lut.tpc_pytorch import \
                 get_pytorch_tpc as get_pytorch_tpc_v3_lut
-            from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v4.tpc_pytorch import \
-                get_pytorch_tpc as get_pytorch_tpc_v4
 
             # Pytorch: TPC versioning
             tpc_models_dict = {'v1': get_pytorch_tpc_v1,
@@ -87,7 +82,6 @@ def get_tpc_dict_by_fw(fw_name):
                                'v2_lut': get_pytorch_tpc_v2_lut,
                                'v3': get_pytorch_tpc_v3,
                                'v3_lut': get_pytorch_tpc_v3_lut,
-                               'v4': get_pytorch_tpc_v4,
                                LATEST: get_pytorch_tpc_latest}
     if tpc_models_dict is not None:
         return tpc_models_dict

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

@@ -167,6 +167,16 @@ def generate_tp_model(default_config: OpQuantizationConfig,
             weights_quantization_method=tp.QuantizationMethod.POWER_OF_TWO))
     const_configuration_options = tp.QuantizationConfigOptions([const_config])
 
+    # 16 bits inputs and outputs. Currently, only defined for consts since they are used in operators that
+    # support 16 bit as input and output.
+    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, signedness=Signedness.SIGNED)
+    const_configuration_options_inout16 = tp.QuantizationConfigOptions([const_config_input16_output16,
+                                                                        const_config_input16],
+                                                                       base_config=const_config_input16)
+
     # Create a TargetPlatformModel and set its default quantization config.
     # This default configuration will be used for all operations
     # unless specified otherwise (see OperatorsSet, for example):
@@ -186,8 +196,10 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         # May suit for operations like: Dropout, Reshape, etc.
         default_qco = tp.get_default_quantization_config_options()
         tp.OperatorsSet("NoQuantization",
-                        default_qco.clone_and_edit(enable_activation_quantization=False)
+                        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,
@@ -200,9 +212,9 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         # Define operations sets without quantization configuration
         # options (useful for creating fusing patterns, for example):
         any_relu = tp.OperatorsSet("AnyReLU")
-        add = tp.OperatorsSet("Add", const_configuration_options)
-        sub = tp.OperatorsSet("Sub", const_configuration_options)
-        mul = tp.OperatorsSet("Mul", const_configuration_options)
+        add = tp.OperatorsSet("Add", const_configuration_options_inout16)
+        sub = tp.OperatorsSet("Sub", const_configuration_options_inout16)
+        mul = tp.OperatorsSet("Mul", const_configuration_options_inout16)
         div = tp.OperatorsSet("Div", const_configuration_options)
         prelu = tp.OperatorsSet("PReLU")
         swish = tp.OperatorsSet("Swish")

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v3/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.v3.tp_model import get_tp_model
 import model_compression_toolkit as mct
@@ -84,6 +84,7 @@ def generate_keras_tpc(name: str, tp_model: tp.TargetPlatformModel):
                      tf.compat.v1.gather,
                      tf.nn.top_k,
                      tf.__operators__.getitem,
+                     tf.strided_slice,
                      tf.image.combined_non_max_suppression,
                      tf.compat.v1.shape]
 
@@ -92,6 +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,

model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v3/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
@@ -85,6 +85,8 @@ def generate_pytorch_tpc(name: str, tp_model: tp.TargetPlatformModel):
                                                     topk,
                                                     squeeze,
                                                     MaxPool2d])
+        tp.OperationsSetToLayers("Default16BitInout",
+                                 [torch.stack, torch.cat, torch.concat, torch.concatenate])
 
         tp.OperationsSetToLayers("Conv", [Conv2d, ConvTranspose2d],
                                  attr_mapping=pytorch_linear_attr_mapping)
@@ -101,8 +103,8 @@ 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])