mct-nightly 2.0.0.20240410.422__tar.gz → 2.0.0.20240412.408__tar.gz

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.0.0.20240410.422
+Version: 2.0.0.20240412.408
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/mct_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.0.0.20240410.422
+Version: 2.0.0.20240412.408
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/mct_nightly.egg-info/SOURCES.txt

@@ -11,6 +11,7 @@ model_compression_toolkit/__init__.py
 model_compression_toolkit/constants.py
 model_compression_toolkit/defaultdict.py
 model_compression_toolkit/logger.py
+model_compression_toolkit/metadata.py
 model_compression_toolkit/core/__init__.py
 model_compression_toolkit/core/analyzer.py
 model_compression_toolkit/core/graph_prep_runner.py
@@ -450,6 +451,14 @@ model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v1_
 model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v1_pot/tp_model.py
 model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v1_pot/tpc_keras.py
 model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v1_pot/tpc_pytorch.py
+model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v2/__init__.py
+model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v2/tp_model.py
+model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v2/tpc_keras.py
+model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v2/tpc_pytorch.py
+model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v2_lut/__init__.py
+model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v2_lut/tp_model.py
+model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v2_lut/tpc_keras.py
+model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v2_lut/tpc_pytorch.py
 model_compression_toolkit/target_platform_capabilities/tpc_models/qnnpack_tpc/__init__.py
 model_compression_toolkit/target_platform_capabilities/tpc_models/qnnpack_tpc/target_platform_capabilities.py
 model_compression_toolkit/target_platform_capabilities/tpc_models/qnnpack_tpc/latest/__init__.py

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/mct_nightly.egg-info/requires.txt

@@ -10,4 +10,4 @@ PuLP
 matplotlib
 scipy
 protobuf
-mct-quantizers==1.4
+mct-quantizers==1.5

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/__init__.py

@@ -27,4 +27,4 @@ from model_compression_toolkit import data_generation
 from model_compression_toolkit import pruning
 from model_compression_toolkit.trainable_infrastructure.keras.load_model import keras_load_quantized_model
 
-__version__ = "2.0.0.20240410.000422"
+__version__ = "2.0.0.20240412.000408"

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/constants.py

@@ -24,6 +24,10 @@ FOUND_ONNX = importlib.util.find_spec("onnx") is not None
 FOUND_ONNXRUNTIME = importlib.util.find_spec("onnxruntime") is not None
 FOUND_SONY_CUSTOM_LAYERS = importlib.util.find_spec('sony_custom_layers') is not None
 
+# Metadata fields
+MCT_VERSION = 'mct_version'
+TPC_VERSION = 'tpc_version'
+
 WEIGHTS_SIGNED = True
 # Minimal threshold to use for quantization ranges:
 MIN_THRESHOLD = (2 ** -16)

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/core/common/graph/base_graph.py

@@ -103,9 +103,10 @@ class Graph(nx.MultiDiGraph, GraphSearches):
             if n.is_custom:
                 if not is_node_in_tpc:
                     Logger.critical(f'MCT does not support optimizing Keras custom layers. Found a layer of type {n.type}. '
-                                 f' Please add the custom layer to Target Platform Capabilities (TPC), or file a feature request or an issue if you believe this should be supported.')
+                                    ' Please add the custom layer to Target Platform Capabilities (TPC), or file a feature '
+                                    'request or an issue if you believe this should be supported.')  # pragma: no cover
                 if any([qc.default_weight_attr_config.enable_weights_quantization for qc in n.get_qco(tpc).quantization_config_list]):
-                    Logger.critical(f'Layer identified: {n.type}. MCT does not support weight quantization for Keras custom layers.')
+                    Logger.critical(f'Layer identified: {n.type}. MCT does not support weight quantization for Keras custom layers.')  # pragma: no cover
 
         self.tpc = tpc
 

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/core/common/quantization/quantization_params_generation/qparams_computation.py

@@ -42,14 +42,14 @@ def calculate_quantization_params(graph: Graph,
 
     """
 
-    Logger.info(f"Running quantization parameters search. "
+    Logger.info(f"\nRunning quantization parameters search. "
                 f"This process might take some time, "
                 f"depending on the model size and the selected quantization methods.\n")
 
     # Create a list of nodes to compute their thresholds
     nodes_list: List[BaseNode] = nodes if specific_nodes else graph.nodes()
 
-    for n in tqdm(nodes_list, "Calculating quantization params"):  # iterate only nodes that we should compute their thresholds
+    for n in tqdm(nodes_list, "Calculating quantization parameters"):  # iterate only nodes that we should compute their thresholds
         for candidate_qc in n.candidates_quantization_cfg:
             for attr in n.get_node_weights_attributes():
                 if n.is_weights_quantization_enabled(attr):

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/exporter/model_exporter/keras/fakely_quant_keras_exporter.py

@@ -100,7 +100,7 @@ class FakelyQuantKerasExporter(BaseKerasExporter):
                         weights_list.append(layer.get_quantized_weights()['kernel'])
                     else:
                         Logger.critical(f'KerasQuantizationWrapper should wrap only DepthwiseConv2D, Conv2D, Dense'
-                                     f' and Conv2DTranspose layers but wrapped layer is {layer.layer}')
+                                        f' and Conv2DTranspose layers but wrapped layer is {layer.layer}')
 
                     if layer.layer.bias is not None:
                         weights_list.append(layer.layer.bias)
@@ -121,6 +121,11 @@ class FakelyQuantKerasExporter(BaseKerasExporter):
 
             return layer
 
+        # Delete metadata layer if exists
+        if hasattr(self.model, 'metadata_layer'):
+            Logger.info('Metadata is not exported to FakelyQuant models.')
+            delattr(self.model, 'metadata_layer')
+
         # clone each layer in the model and apply _unwrap_quantize_wrapper to layers wrapped with a QuantizeWrapper.
         self.exported_model = tf.keras.models.clone_model(self.model,
                                                           input_tensors=None,

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/exporter/model_exporter/keras/fakely_quant_tflite_exporter.py

@@ -56,6 +56,11 @@ class FakelyQuantTFLiteExporter(FakelyQuantKerasExporter):
         (namely, weights that are in fake-quant format) and fake-quant layers for the activations.
 
         """
+        # Delete metadata layer if exists
+        if hasattr(self.model, 'metadata_layer'):
+            Logger.info('Metadata is not exported to TFLite models.')
+            delattr(self.model, 'metadata_layer')
+
         # Use Keras exporter to quantize model's weights before converting it to TFLite.
         # Since exporter saves the model, we use a tmp path for saving, and then we delete it.
         handle, tmp_file = tempfile.mkstemp(DEFAULT_KERAS_EXPORT_EXTENTION)

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/exporter/model_exporter/keras/int8_tflite_exporter.py

@@ -166,6 +166,11 @@ class INT8TFLiteExporter(FakelyQuantKerasExporter):
 
             return layer_to_substitue
 
+        # Delete metadata layer if exists
+        if hasattr(self.model, 'metadata_layer'):
+            Logger.info('Metadata is not exported to TFLite models.')
+            delattr(self.model, 'metadata_layer')
+
         # Transform the model to a new model that can be converted to int8 models.
         # For example: replace dense layers with point-wise layers (to support per-channel quantization)
         self.transformed_model = clone_model(self.model,

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/exporter/model_exporter/pytorch/fakely_quant_onnx_pytorch_exporter.py

@@ -13,17 +13,21 @@
 # limitations under the License.
 # ==============================================================================
 from typing import Callable
+from io import BytesIO
 
 import torch.nn
+import onnx
 
 from mct_quantizers import PytorchActivationQuantizationHolder, PytorchQuantizationWrapper
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.core.pytorch.utils import to_torch_tensor
 from model_compression_toolkit.exporter.model_exporter.pytorch.base_pytorch_exporter import BasePyTorchExporter
 from mct_quantizers import pytorch_quantizers
+from mct_quantizers.pytorch.metadata import add_onnx_metadata
 
 DEFAULT_ONNX_OPSET_VERSION=15
 
+
 class FakelyQuantONNXPyTorchExporter(BasePyTorchExporter):
     """
     Exporter for fakely-quant PyTorch models.
@@ -58,7 +62,6 @@ class FakelyQuantONNXPyTorchExporter(BasePyTorchExporter):
         self._use_onnx_custom_quantizer_ops = use_onnx_custom_quantizer_ops
         self._onnx_opset_version = onnx_opset_version
 
-
     def export(self) -> None:
         """
         Convert an exportable (fully-quantized) PyTorch model to a fakely-quant model
@@ -74,7 +77,7 @@ class FakelyQuantONNXPyTorchExporter(BasePyTorchExporter):
         # If _use_onnx_custom_quantizer_ops is set to True, the quantizer forward function will use
         # the custom implementation when exporting the operator into onnx model. If not, it removes the
         # wraps and quantizes the ops in place (for weights, for activation torch quantization function is
-        # exported since it's used during forward.
+        # exported since it's used during forward).
         if self._use_onnx_custom_quantizer_ops:
             self._enable_onnx_custom_ops_export()
         else:
@@ -87,15 +90,30 @@ class FakelyQuantONNXPyTorchExporter(BasePyTorchExporter):
 
         model_input = to_torch_tensor(next(self.repr_dataset())[0])
 
-        torch.onnx.export(self.model,
-                          model_input,
-                          self.save_model_path,
-                          opset_version=self._onnx_opset_version,
-                          verbose=False,
-                          input_names=['input'],
-                          output_names=['output'],
-                          dynamic_axes={'input': {0: 'batch_size'},
-                                        'output': {0: 'batch_size'}})
+        if hasattr(self.model, 'metadata'):
+            onnx_bytes = BytesIO()
+            torch.onnx.export(self.model,
+                              model_input,
+                              onnx_bytes,
+                              opset_version=self._onnx_opset_version,
+                              verbose=False,
+                              input_names=['input'],
+                              output_names=['output'],
+                              dynamic_axes={'input': {0: 'batch_size'},
+                                            'output': {0: 'batch_size'}})
+            onnx_model = onnx.load_from_string(onnx_bytes.getvalue())
+            onnx_model = add_onnx_metadata(onnx_model, self.model.metadata)
+            onnx.save_model(onnx_model, self.save_model_path)
+        else:
+            torch.onnx.export(self.model,
+                              model_input,
+                              self.save_model_path,
+                              opset_version=self._onnx_opset_version,
+                              verbose=False,
+                              input_names=['input'],
+                              output_names=['output'],
+                              dynamic_axes={'input': {0: 'batch_size'},
+                                            'output': {0: 'batch_size'}})
 
     def _enable_onnx_custom_ops_export(self):
         """

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/exporter/model_exporter/pytorch/pytorch_export_facade.py

@@ -40,7 +40,7 @@ if FOUND_TORCH:
                              repr_dataset: Callable,
                              is_layer_exportable_fn: Callable = is_pytorch_layer_exportable,
                              serialization_format: PytorchExportSerializationFormat = PytorchExportSerializationFormat.ONNX,
-                             quantization_format : QuantizationFormat = QuantizationFormat.MCTQ,
+                             quantization_format: QuantizationFormat = QuantizationFormat.MCTQ,
                              onnx_opset_version=DEFAULT_ONNX_OPSET_VERSION) -> None:
         """
         Export a PyTorch quantized model to a torchscript or onnx model.

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/exporter/model_wrapper/keras/builder/fully_quantized_model_builder.py

@@ -90,7 +90,7 @@ if FOUND_TF:
                                                                                         fw_impl=C.keras.keras_implementation.KerasImplementation())).build_model()
         exportable_model.trainable = False
 
-        Logger.info("Please run your accuracy evaluation on the exported quantized model to verify it's accuracy.\n"
+        Logger.info("\nPlease run your accuracy evaluation on the exported quantized model to verify it's accuracy.\n"
                     "Checkout the FAQ and Troubleshooting pages for resolving common issues and improving the quantized model accuracy:\n"
                     "FAQ: https://github.com/sony/model_optimization/tree/main/FAQ.md\n"
                     "Quantization Troubleshooting: https://github.com/sony/model_optimization/tree/main/quantization_troubleshooting.md")

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/exporter/model_wrapper/pytorch/builder/fully_quantized_model_builder.py

@@ -82,7 +82,7 @@ if FOUND_TORCH:
                                                           get_activation_quantizer_holder(n,
                                                                                           fw_impl=C.pytorch.pytorch_implementation.PytorchImplementation())).build_model()
 
-        Logger.info("Please run your accuracy evaluation on the exported quantized model to verify it's accuracy.\n"
+        Logger.info("\nPlease run your accuracy evaluation on the exported quantized model to verify it's accuracy.\n"
                     "Checkout the FAQ and Troubleshooting pages for resolving common issues and improving the quantized model accuracy:\n"
                     "FAQ: https://github.com/sony/model_optimization/tree/main/FAQ.md\n"
                     "Quantization Troubleshooting: https://github.com/sony/model_optimization/tree/main/quantization_troubleshooting.md")
@@ -93,4 +93,4 @@ if FOUND_TORCH:
 else:
     def get_exportable_pytorch_model(*args, **kwargs):
         Logger.critical("PyTorch must be installed to use 'get_exportable_pytorch_model'. "
-                        "The 'torch' package is missing.")  # pragma: no cover
+                        "The 'torch' package is missing.")  # pragma: no cover

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/gptq/keras/gptq_training.py

@@ -301,21 +301,23 @@ class KerasGPTQTrainer(GPTQTrainer):
         Returns: None
 
         """
-        for _ in tqdm(range(n_epochs)):
-            for data in tqdm(data_function()):
-                input_data = [d * self.input_scale for d in data]
-
-                loss_value_step, grads = self.nano_training_step(input_data, in_compute_gradients,
-                                                                 in_optimizer_with_param, is_training)
-                # Run one step of gradient descent by updating
-                # the value of the variables to minimize the loss.
-                for i, (o, p) in enumerate(in_optimizer_with_param):
-                    o.apply_gradients(zip(grads[i], p))
-                if self.gptq_config.log_function is not None:
-                    self.gptq_config.log_function(loss_value_step, grads[0], in_optimizer_with_param[0][-1],
-                                                  self.compare_points)
-                self.loss_list.append(loss_value_step.numpy())
-                Logger.debug(f'last loss value: {self.loss_list[-1]}')
+        with tqdm(range(n_epochs), "Running GPTQ optimization") as epochs_pbar:
+            for _ in epochs_pbar:
+                with tqdm(data_function(), position=1, leave=False) as data_pbar:
+                    for data in data_pbar:
+                        input_data = [d * self.input_scale for d in data]
+
+                        loss_value_step, grads = self.nano_training_step(input_data, in_compute_gradients,
+                                                                         in_optimizer_with_param, is_training)
+                        # Run one step of gradient descent by updating
+                        # the value of the variables to minimize the loss.
+                        for i, (o, p) in enumerate(in_optimizer_with_param):
+                            o.apply_gradients(zip(grads[i], p))
+                        if self.gptq_config.log_function is not None:
+                            self.gptq_config.log_function(loss_value_step, grads[0], in_optimizer_with_param[0][-1],
+                                                          self.compare_points)
+                        self.loss_list.append(loss_value_step.numpy())
+                        Logger.debug(f'last loss value: {self.loss_list[-1]}')
 
     def update_graph(self):
         """

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/gptq/keras/quantization_facade.py

@@ -31,6 +31,7 @@ from model_compression_toolkit.core.runner import core_runner
 from model_compression_toolkit.gptq.runner import gptq_runner
 from model_compression_toolkit.core.analyzer import analyzer_model_quantization
 from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework import TargetPlatformCapabilities
+from model_compression_toolkit.metadata import get_versions_dict
 
 LR_DEFAULT = 0.15
 LR_REST_DEFAULT = 1e-4
@@ -48,6 +49,7 @@ if FOUND_TF:
     from model_compression_toolkit.target_platform_capabilities.constants import DEFAULT_TP_MODEL
     from model_compression_toolkit.exporter.model_wrapper import get_exportable_keras_model
     from model_compression_toolkit import get_target_platform_capabilities
+    from mct_quantizers.keras.metadata import add_metadata
 
     # As from TF2.9 optimizers package is changed
     if version.parse(tf.__version__) < version.parse("2.9"):
@@ -234,7 +236,10 @@ if FOUND_TF:
                                         fw_impl,
                                         DEFAULT_KERAS_INFO)
 
-        return get_exportable_keras_model(tg_gptq)
+        exportable_model, user_info = get_exportable_keras_model(tg_gptq)
+        if target_platform_capabilities.tp_model.add_metadata:
+            exportable_model = add_metadata(exportable_model, get_versions_dict(target_platform_capabilities))
+        return exportable_model, user_info
 
 else:
     # If tensorflow is not installed,

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/gptq/keras/quantizer/regularization_factory.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+from tqdm import tqdm
 from typing import Callable
 
 from model_compression_toolkit.gptq import RoundingType, GradientPTQConfig, GradientPTQConfig
@@ -35,7 +36,7 @@ def get_regularization(gptq_config: GradientPTQConfig, representative_data_gen:
     if gptq_config.rounding_type == RoundingType.SoftQuantizer:
         # dry run on the representative dataset to count number of batches
         num_batches = 0
-        for _ in representative_data_gen():
+        for _ in tqdm(representative_data_gen(), "GPTQ initialization"):
             num_batches += 1
 
         return SoftQuantizerRegularization(total_gradient_steps=num_batches * gptq_config.n_epochs)

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/gptq/pytorch/gptq_training.py

@@ -248,22 +248,24 @@ class PytorchGPTQTrainer(GPTQTrainer):
             data_function: A callable function that give a batch of samples.
             n_epochs: Number of update iterations of representative dataset.
         """
-        for _ in tqdm(range(n_epochs)):
-            for data in tqdm(data_function()):
-                input_data = [d * self.input_scale for d in data]
-                input_tensor = to_torch_tensor(input_data)
-                y_float = self.float_model(input_tensor)  # running float model
-                loss_value, grads = self.compute_gradients(y_float, input_tensor)
-                # Run one step of gradient descent by updating the value of the variables to minimize the loss.
-                for (optimizer, _) in self.optimizer_with_param:
-                    optimizer.step()
-                    optimizer.zero_grad()
-                if self.gptq_config.log_function is not None:
-                    self.gptq_config.log_function(loss_value.item(),
-                                                  torch_tensor_to_numpy(grads),
-                                                  torch_tensor_to_numpy(self.optimizer_with_param[0][-1]))
-                self.loss_list.append(loss_value.item())
-                Logger.debug(f'last loss value: {self.loss_list[-1]}')
+        with tqdm(range(n_epochs), "Running GPTQ optimization") as epochs_pbar:
+            for _ in epochs_pbar:
+                with tqdm(data_function(), position=1, leave=False) as data_pbar:
+                    for data in data_pbar:
+                        input_data = [d * self.input_scale for d in data]
+                        input_tensor = to_torch_tensor(input_data)
+                        y_float = self.float_model(input_tensor)  # running float model
+                        loss_value, grads = self.compute_gradients(y_float, input_tensor)
+                        # Run one step of gradient descent by updating the value of the variables to minimize the loss.
+                        for (optimizer, _) in self.optimizer_with_param:
+                            optimizer.step()
+                            optimizer.zero_grad()
+                        if self.gptq_config.log_function is not None:
+                            self.gptq_config.log_function(loss_value.item(),
+                                                          torch_tensor_to_numpy(grads),
+                                                          torch_tensor_to_numpy(self.optimizer_with_param[0][-1]))
+                        self.loss_list.append(loss_value.item())
+                        Logger.debug(f'last loss value: {self.loss_list[-1]}')
 
     def update_graph(self) -> Graph:
         """

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/gptq/pytorch/quantization_facade.py

@@ -31,6 +31,7 @@ from model_compression_toolkit.core.analyzer import analyzer_model_quantization
 from model_compression_toolkit.core import CoreConfig
 from model_compression_toolkit.core.common.mixed_precision.mixed_precision_quantization_config import \
     MixedPrecisionQuantizationConfig
+from model_compression_toolkit.metadata import get_versions_dict
 
 LR_DEFAULT = 1e-4
 LR_REST_DEFAULT = 1e-4
@@ -47,6 +48,7 @@ if FOUND_TORCH:
     from torch.nn import Module
     from torch.optim import Adam, Optimizer
     from model_compression_toolkit import get_target_platform_capabilities
+    from mct_quantizers.pytorch.metadata import add_metadata
     DEFAULT_PYTORCH_TPC = get_target_platform_capabilities(PYTORCH, DEFAULT_TP_MODEL)
 
     def get_pytorch_gptq_config(n_epochs: int,
@@ -202,7 +204,10 @@ if FOUND_TORCH:
                                         fw_impl,
                                         DEFAULT_PYTORCH_INFO)
 
-        return get_exportable_pytorch_model(graph_gptq)
+        exportable_model, user_info = get_exportable_pytorch_model(graph_gptq)
+        if target_platform_capabilities.tp_model.add_metadata:
+            exportable_model = add_metadata(exportable_model, get_versions_dict(target_platform_capabilities))
+        return exportable_model, user_info
 
 
 else:

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/gptq/pytorch/quantizer/regularization_factory.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+from tqdm import tqdm
 from typing import Callable
 
 from model_compression_toolkit.gptq import RoundingType, GradientPTQConfig, GradientPTQConfig
@@ -35,7 +36,7 @@ def get_regularization(gptq_config: GradientPTQConfig, representative_data_gen:
     if gptq_config.rounding_type == RoundingType.SoftQuantizer:
         # dry run on the representative dataset to count number of batches
         num_batches = 0
-        for _ in representative_data_gen():
+        for _ in tqdm(representative_data_gen(), "GPTQ initialization"):
             num_batches += 1
 
         return SoftQuantizerRegularization(total_gradient_steps=num_batches * gptq_config.n_epochs)

mct-nightly-2.0.0.20240412.408/model_compression_toolkit/metadata.py

@@ -0,0 +1,29 @@
+# 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 typing import Dict
+from model_compression_toolkit.constants import MCT_VERSION, TPC_VERSION
+
+
+def get_versions_dict(tpc) -> Dict:
+    """
+
+    Returns: A dictionary with TPC and MCT versions.
+
+    """
+    # imported inside to avoid circular import error
+    from model_compression_toolkit import __version__ as mct_version
+    tpc_version = f'{tpc.name}.{tpc.version}'
+    return {MCT_VERSION: mct_version, TPC_VERSION: tpc_version}

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/ptq/keras/quantization_facade.py

@@ -28,6 +28,7 @@ from model_compression_toolkit.core.common.mixed_precision.mixed_precision_quant
 from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework import TargetPlatformCapabilities
 from model_compression_toolkit.core.runner import core_runner
 from model_compression_toolkit.ptq.runner import ptq_runner
+from model_compression_toolkit.metadata import get_versions_dict
 
 if FOUND_TF:
     from model_compression_toolkit.core.keras.default_framework_info import DEFAULT_KERAS_INFO
@@ -38,6 +39,7 @@ if FOUND_TF:
     from model_compression_toolkit.exporter.model_wrapper import get_exportable_keras_model
 
     from model_compression_toolkit import get_target_platform_capabilities
+    from mct_quantizers.keras.metadata import add_metadata
     DEFAULT_KERAS_TPC = get_target_platform_capabilities(TENSORFLOW, DEFAULT_TP_MODEL)
 
 
@@ -164,8 +166,10 @@ if FOUND_TF:
                                         fw_impl,
                                         fw_info)
 
-        return get_exportable_keras_model(graph_with_stats_correction)
-
+        exportable_model, user_info = get_exportable_keras_model(graph_with_stats_correction)
+        if target_platform_capabilities.tp_model.add_metadata:
+            exportable_model = add_metadata(exportable_model, get_versions_dict(target_platform_capabilities))
+        return exportable_model, user_info
 
 
 else:

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/ptq/pytorch/quantization_facade.py

@@ -29,6 +29,7 @@ from model_compression_toolkit.core.runner import core_runner
 from model_compression_toolkit.ptq.runner import ptq_runner
 from model_compression_toolkit.core.analyzer import analyzer_model_quantization
 from model_compression_toolkit.core.common.quantization.quantize_graph_weights import quantize_graph_weights
+from model_compression_toolkit.metadata import get_versions_dict
 
 
 if FOUND_TORCH:
@@ -38,6 +39,7 @@ if FOUND_TORCH:
     from torch.nn import Module
     from model_compression_toolkit.exporter.model_wrapper.pytorch.builder.fully_quantized_model_builder import get_exportable_pytorch_model
     from model_compression_toolkit import get_target_platform_capabilities
+    from mct_quantizers.pytorch.metadata import add_metadata
 
     DEFAULT_PYTORCH_TPC = get_target_platform_capabilities(PYTORCH, DEFAULT_TP_MODEL)
 
@@ -139,7 +141,10 @@ if FOUND_TORCH:
                                         fw_impl,
                                         fw_info)
 
-        return get_exportable_pytorch_model(graph_with_stats_correction)
+        exportable_model, user_info = get_exportable_pytorch_model(graph_with_stats_correction)
+        if target_platform_capabilities.tp_model.add_metadata:
+            exportable_model = add_metadata(exportable_model, get_versions_dict(target_platform_capabilities))
+        return exportable_model, user_info
 
 
 else:

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/target_platform_capabilities/target_platform/target_platform_model.py

@@ -60,15 +60,18 @@ class TargetPlatformModel(ImmutableClass):
 
     def __init__(self,
                  default_qco: QuantizationConfigOptions,
+                 add_metadata: bool = False,
                  name="default_tp_model"):
         """
 
         Args:
             default_qco (QuantizationConfigOptions): Default QuantizationConfigOptions to use for operators that their QuantizationConfigOptions are not defined in the model.
+            add_metadata (bool): Whether to add metadata to the model or not.
             name (str): Name of the model.
         """
 
         super().__init__()
+        self.add_metadata = add_metadata
         self.name = name
         self.operator_set = []
         assert isinstance(default_qco, QuantizationConfigOptions)
@@ -191,7 +194,7 @@ class TargetPlatformModel(ImmutableClass):
 
         """
         opsets_names = [op.name for op in self.operator_set]
-        if (len(set(opsets_names)) != len(opsets_names)):
+        if len(set(opsets_names)) != len(opsets_names):
             Logger.critical(f'Operator Sets must have unique names.')
 
     def get_default_config(self) -> OpQuantizationConfig:

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/target_platform_capabilities/target_platform/targetplatform2framework/target_platform_capabilities.py

@@ -29,6 +29,7 @@ from model_compression_toolkit.target_platform_capabilities.target_platform.op_q
 from model_compression_toolkit.target_platform_capabilities.target_platform.operators import OperatorsSetBase
 from model_compression_toolkit.target_platform_capabilities.target_platform.target_platform_model import TargetPlatformModel
 from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework.current_tpc import _current_tpc
+from model_compression_toolkit.constants import MCT_VERSION, TPC_VERSION
 
 
 class TargetPlatformCapabilities(ImmutableClass):

{mct-nightly-2.0.0.20240410.422 → mct-nightly-2.0.0.20240412.408}/model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/target_platform_capabilities.py

@@ -25,11 +25,15 @@ if FOUND_TF:
     from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v1.tpc_keras import get_keras_tpc as get_keras_tpc_v1
     from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v1_lut.tpc_keras import get_keras_tpc as get_keras_tpc_v1_lut
     from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v1_pot.tpc_keras import get_keras_tpc as get_keras_tpc_v1_pot
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v2.tpc_keras import get_keras_tpc as get_keras_tpc_v2
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v2_lut.tpc_keras import get_keras_tpc as get_keras_tpc_v2_lut
 
     # Keras: TPC versioning
     keras_tpc_models_dict = {'v1': get_keras_tpc_v1(),
                              'v1_lut': get_keras_tpc_v1_lut(),
                              'v1_pot': get_keras_tpc_v1_pot(),
+                             'v2': get_keras_tpc_v2(),
+                             'v2_lut': get_keras_tpc_v2_lut(),
                              LATEST: get_keras_tpc_latest()}
 
 ###############################
@@ -42,13 +46,19 @@ if FOUND_TORCH:
         get_pytorch_tpc as get_pytorch_tpc_v1
     from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v1_pot.tpc_pytorch import \
         get_pytorch_tpc as get_pytorch_tpc_v1_pot
-    from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v1_lut.tpc_pytorch import get_pytorch_tpc as get_pytorch_tpc_v1_lut
-
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v1_lut.tpc_pytorch import \
+        get_pytorch_tpc as get_pytorch_tpc_v1_lut
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v2.tpc_pytorch import \
+        get_pytorch_tpc as get_pytorch_tpc_v2
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.imx500_tpc.v2_lut.tpc_pytorch import \
+        get_pytorch_tpc as get_pytorch_tpc_v2_lut
 
     # Pytorch: TPC versioning
     pytorch_tpc_models_dict = {'v1': get_pytorch_tpc_v1(),
                                'v1_lut': get_pytorch_tpc_v1_lut(),
                                'v1_pot': get_pytorch_tpc_v1_pot(),
+                               'v2': get_pytorch_tpc_v2(),
+                               'v2_lut': get_pytorch_tpc_v2_lut(),
                                LATEST: get_pytorch_tpc_latest()}
 
 tpc_dict = {TENSORFLOW: keras_tpc_models_dict,

mct-nightly-2.0.0.20240412.408/model_compression_toolkit/target_platform_capabilities/tpc_models/imx500_tpc/v2/__init__.py

@@ -0,0 +1,16 @@
+# 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.
+# ==============================================================================
+
+__version__ = 'v2'