Merge branch 'develop' into onnx_bump_version

examples/llm_compression/openvino/tiny_llama_find_hyperparams/main.py

@@ -24,6 +24,7 @@
 
 import nncf
 from nncf.common.logging import nncf_logger
+from nncf.quantization.advanced_parameters import AdvancedCompressionParameters
 
 DataItem = TypeVar("DataItem")
 ModelInput = TypeVar("ModelInput")
@@ -63,6 +64,7 @@ def compress_model(
         group_size=group_size,
         awq=awq,
         sensitivity_metric=nncf.parameters.SensitivityMetric.MAX_ACTIVATION_VARIANCE,
+        advanced_parameters=AdvancedCompressionParameters(statistics_path="statistics"),
     )
     return optimized_ov_model
 

examples/post_training_quantization/onnx/yolov8_quantize_with_accuracy_control/deploy.py

@@ -42,7 +42,7 @@ def validate_ov_model(
 ) -> Tuple[Dict, int, int]:
     validator.seen = 0
     validator.jdict = []
-    validator.stats = []
+    validator.stats = dict(tp_m=[], tp=[], conf=[], pred_cls=[], target_cls=[], target_img=[])
     validator.batch_i = 1
     validator.confusion_matrix = ConfusionMatrix(nc=validator.nc)
     compiled_model = ov.compile_model(ov_model, device_name="CPU")

examples/post_training_quantization/onnx/yolov8_quantize_with_accuracy_control/main.py

@@ -38,7 +38,7 @@ def validate(
 ) -> Tuple[Dict, int, int]:
     validator.seen = 0
     validator.jdict = []
-    validator.stats = []
+    validator.stats = dict(tp_m=[], tp=[], conf=[], pred_cls=[], target_cls=[], target_img=[])
     validator.batch_i = 1
     validator.confusion_matrix = ConfusionMatrix(nc=validator.nc)
 
@@ -101,13 +101,12 @@ def print_statistics(stats: np.ndarray, total_images: int, total_objects: int) -
 
 
 def prepare_validation(model: YOLO, args: Any) -> Tuple[Validator, torch.utils.data.DataLoader]:
-    validator = model.smart_load("validator")(args)
+    validator = model.task_map[model.task]["validator"](args=args)
     validator.data = check_det_dataset(args.data)
+    validator.stride = 32
 
     data_loader = validator.get_dataloader(f"{DATASETS_DIR}/coco128-seg", 1)
 
-    validator = model.smart_load("validator")(args)
-
     validator.is_coco = True
     validator.class_map = coco80_to_coco91_class()
     validator.names = model.model.names
@@ -146,7 +145,7 @@ def validation_ac(
     ) -> float:
         validator.seen = 0
         validator.jdict = []
-        validator.stats = []
+        validator.stats = dict(tp_m=[], tp=[], conf=[], pred_cls=[], target_cls=[], target_img=[])
         validator.batch_i = 1
         validator.confusion_matrix = ConfusionMatrix(nc=validator.nc)
 

examples/post_training_quantization/onnx/yolov8_quantize_with_accuracy_control/requirements.txt

@@ -1,3 +1,3 @@
-ultralytics==8.0.170
+ultralytics==8.3.22
 onnx==1.17.0
 openvino==2024.4

examples/post_training_quantization/openvino/yolov8/main.py

@@ -36,7 +36,7 @@ def validate(
 ) -> Tuple[Dict, int, int]:
     validator.seen = 0
     validator.jdict = []
-    validator.stats = []
+    validator.stats = dict(tp=[], conf=[], pred_cls=[], target_cls=[], target_img=[])
     validator.confusion_matrix = ConfusionMatrix(nc=validator.nc)
     model.reshape({0: [1, 3, -1, -1]})
     compiled_model = ov.compile_model(model, device_name="CPU")
@@ -66,8 +66,9 @@ def print_statistics(stats: np.ndarray, total_images: int, total_objects: int) -
 
 
 def prepare_validation(model: YOLO, args: Any) -> Tuple[Validator, torch.utils.data.DataLoader]:
-    validator = model.smart_load("validator")(args)
+    validator = model.task_map[model.task]["validator"](args=args)
     validator.data = check_det_dataset(args.data)
+    validator.stride = 32
     dataset = validator.data["val"]
     print(f"{dataset}")
 

examples/post_training_quantization/openvino/yolov8/requirements.txt

@@ -1,3 +1,3 @@
-ultralytics==8.0.170
+ultralytics==8.3.22
 onnx>=1.12.0,<1.16.2
 openvino==2024.4

examples/post_training_quantization/openvino/yolov8_quantize_with_accuracy_control/main.py

@@ -38,7 +38,7 @@ def validate(
 ) -> Tuple[Dict, int, int]:
     validator.seen = 0
     validator.jdict = []
-    validator.stats = []
+    validator.stats = dict(tp_m=[], tp=[], conf=[], pred_cls=[], target_cls=[], target_img=[])
     validator.batch_i = 1
     validator.confusion_matrix = ConfusionMatrix(nc=validator.nc)
     model.reshape({0: [1, 3, -1, -1]})
@@ -92,15 +92,14 @@ def print_statistics(stats: np.ndarray, total_images: int, total_objects: int) -
 
 
 def prepare_validation(model: YOLO, args: Any) -> Tuple[Validator, torch.utils.data.DataLoader]:
-    validator = model.smart_load("validator")(args)
+    validator = model.task_map[model.task]["validator"](args=args)
     validator.data = check_det_dataset(args.data)
+    validator.stride = 32
     dataset = validator.data["val"]
     print(f"{dataset}")
 
     data_loader = validator.get_dataloader(f"{DATASETS_DIR}/coco128-seg", 1)
 
-    validator = model.smart_load("validator")(args)
-
     validator.is_coco = True
     validator.class_map = coco80_to_coco91_class()
     validator.names = model.model.names
@@ -146,7 +145,7 @@ def validation_ac(
     ) -> float:
         validator.seen = 0
         validator.jdict = []
-        validator.stats = []
+        validator.stats = dict(tp_m=[], tp=[], conf=[], pred_cls=[], target_cls=[], target_img=[])
         validator.batch_i = 1
         validator.confusion_matrix = ConfusionMatrix(nc=validator.nc)
         num_outputs = len(compiled_model.outputs)
@@ -211,6 +210,7 @@ def main():
     model = YOLO(f"{ROOT}/{MODEL_NAME}.pt")
     args = get_cfg(cfg=DEFAULT_CFG)
     args.data = "coco128-seg.yaml"
+    args.workers = 0
 
     # Prepare validation dataset and helper
     validator, data_loader = prepare_validation(model, args)

examples/post_training_quantization/openvino/yolov8_quantize_with_accuracy_control/requirements.txt

@@ -1,3 +1,3 @@
-ultralytics==8.0.170
+ultralytics==8.3.22
 onnx>=1.12.0,<1.16.2
 openvino==2024.4

examples/post_training_quantization/torch/mobilenet_v2/requirements.txt

@@ -3,3 +3,4 @@ openvino==2024.4
 scikit-learn
 torch==2.4.0
 torchvision==0.19.0
+setuptools<=72.1.0

examples/post_training_quantization/torch/ssd300_vgg16/requirements.txt

@@ -6,3 +6,4 @@ torch==2.4.0
 torchmetrics==1.0.1
 torchvision==0.19.0
 numpy<2
+setuptools<=72.1.0

examples/post_training_quantization/torch_fx/resnet18/main.py

@@ -24,11 +24,13 @@
 import torchvision.models as models
 import torchvision.transforms as transforms
 from fastdownload import FastDownload
+from torch._dynamo.exc import BackendCompilerFailed
 
 import nncf
 import nncf.torch
 from nncf.common.logging.track_progress import track
 from nncf.common.utils.helpers import create_table
+from nncf.common.utils.os import is_windows
 from nncf.torch import disable_patching
 
 IMAGE_SIZE = 64
@@ -205,7 +207,16 @@ def transform_fn(data_item):
     print("Benchmark FP32 model compiled with default backend ...")
     with disable_patching():
         compiled_model = torch.compile(model)
-        fp32_latency = measure_latency(compiled_model, example_inputs=example_input)
+        try:
+            fp32_latency = measure_latency(compiled_model, example_inputs=example_input)
+        except BackendCompilerFailed as exp:
+            if not is_windows():
+                raise exp
+            print(
+                "WARNING: Torch Inductor is currently unavailable on Windows. "
+                "For more information, visit https://github.com/pytorch/pytorch/issues/135954"
+            )
+            fp32_latency = float("nan")
     print(f"{fp32_latency:.3f} ms")
 
     print("Benchmark FP32 model compiled with openvino backend ...")

examples/quantization_aware_training/torch/anomalib/requirements.txt

@@ -1 +1,2 @@
-anomalib[core,openvino]==1.0.0
+anomalib[core,openvino]==1.0.1
+setuptools<=72.1.0

examples/quantization_aware_training/torch/resnet18/requirements.txt

@@ -2,3 +2,4 @@ fastdownload==0.0.7
 openvino==2024.4
 torch==2.4.0
 torchvision==0.19.0
+setuptools<=72.1.0

nncf/common/quantization/quantizer_propagation/graph.py

@@ -719,6 +719,7 @@ def remove_propagating_quantizer(
         if prop_quantizer.unified_scale_type is not None:
             gid = self._unified_scale_group_manager.get_group_id_by_propagating_quantizer_id(prop_quantizer.id)
             self._unified_scale_group_manager.remove_from_group(gid, prop_quantizer)
+        self._pqs_after_weight_dependent_output_quantized_nodes.pop(prop_quantizer, None)
 
     def propagate_quantizer_via_path(
         self, prop_quantizer: PropagatingQuantizer, path: PropagationPath

nncf/common/tensor_statistics/aggregator.py

@@ -8,22 +8,28 @@
 # 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 abc import ABC
 from abc import abstractmethod
 from itertools import islice
 from typing import Any, Dict, Optional, TypeVar
 
 import nncf
+import nncf.common.tensor_statistics.statistics_serializer as statistics_serializer
+import nncf.common.tensor_statistics.statistics_validator as statistics_validator
 from nncf.common import factory
 from nncf.common.graph.graph import NNCFGraph
+from nncf.common.graph.transformations.commands import TargetPoint
 from nncf.common.graph.transformations.layout import TransformationLayout
 from nncf.common.logging import nncf_logger
 from nncf.common.logging.track_progress import track
 from nncf.common.tensor import NNCFTensor
 from nncf.common.tensor_statistics.statistic_point import StatisticPointsContainer
+from nncf.common.utils.backend import BackendType
 from nncf.data.dataset import DataItem
 from nncf.data.dataset import Dataset
 from nncf.data.dataset import ModelInput
+from nncf.experimental.common.tensor_statistics.statistics import TensorStatistic
 
 TensorType = TypeVar("TensorType")
 TModel = TypeVar("TModel")
@@ -38,6 +44,8 @@ class StatisticsAggregator(ABC):
     Base class for statistics collection.
     """
 
+    BACKEND: BackendType
+
     def __init__(self, dataset: Dataset[DataItem, ModelInput]):
         self.dataset = dataset
         self.stat_subset_size = None
@@ -88,6 +96,56 @@ def collect_statistics(self, model: TModel, graph: NNCFGraph) -> None:
                 f"smaller than the requested subset size {self.stat_subset_size}."
             )
 
+    def load_statistics_from_dir(self, dir_path: str) -> None:
+        """
+        Loads statistics from a directory and populates the statistic points with the loaded data.
+
+        :param dir_path: The name of the directory from which to load the statistics.
+        """
+        loaded_data, metadata = statistics_serializer.load_from_dir(dir_path)
+        statistics_validator.validate_backend(metadata, self.BACKEND)
+        self._load_statistics(loaded_data)
+        nncf_logger.info(f"Statistics were successfully loaded from a directory {dir_path}.")
+
+    def _load_statistics(self, data: Dict[str, Any]) -> None:
+        """
+        Loads statistics into the registered statistic points from the given data.
+
+        :param data: A dictionary containing the statistics loaded from a file.
+        """
+        for _, statistic_point, tensor_collector in self.statistic_points.get_tensor_collectors():
+            statistics = tensor_collector.get_statistics()
+            statistics_key = self._get_statistics_key(statistics, statistic_point.target_point)
+            if statistics_key not in data:
+                raise nncf.ValidationError(f"Not found statistics for {statistics_key}")
+            statistics_container = tensor_collector.create_statistics_container(data[statistics_key])
+            tensor_collector.set_cache(statistics_container)
+
+    def dump_statistics(self, dir_path: str) -> None:
+        """
+        Dumps the current statistics to a directory in a compressed format.
+
+        :param dir_path: The path of the directory where the statistics will be saved.
+        """
+        data_to_dump = self._prepare_statistics()
+        metadata = {"backend": self.BACKEND.value, "subset_size": self.stat_subset_size}
+        statistics_serializer.dump_to_dir(data_to_dump, dir_path, metadata)
+        nncf_logger.info(f"Statistics were successfully saved to a directory {dir_path}.")
+
+    def _prepare_statistics(self) -> Dict[str, Any]:
+        """
+        Prepares the statistics data for dumping into a directory.
+
+        :return: A dictionary containing the statistics data to be dumped.
+        """
+        data_to_dump = {}
+        for _, statistic_point, tensor_collector in self.statistic_points.get_tensor_collectors():
+            statistics = tensor_collector.get_statistics()
+            statistics_key = self._get_statistics_key(statistics, statistic_point.target_point)
+            data = statistics.get_data()
+            data_to_dump[statistics_key] = data
+        return data_to_dump
+
     def register_statistic_points(self, statistic_points: StatisticPointsContainer) -> None:
         """
         Register statistic points for statistics collection and recalculates the maximum number samples
@@ -154,3 +212,13 @@ def _process_outputs(outputs: Any) -> Dict[str, NNCFTensor]:
         :param outputs: raw model outputs
         :return: processed model outputs in Dict[str, Tensor] format
         """
+
+    @abstractmethod
+    def _get_statistics_key(self, statistics: TensorStatistic, target_point: TargetPoint) -> str:
+        """
+        Returns key of statistics.
+
+        :param statistics: Statistics value.
+        :param target_point: Statistics target point.
+        :return: Statistics key.
+        """

nncf/common/tensor_statistics/collectors.py

@@ -59,14 +59,14 @@ def register_input(self, x: TensorType) -> TensorType:
     def _register_input(self, x: TensorType) -> None:
         pass
 
-    def get_statistics(self) -> None:
+    def get_statistics(self) -> Any:
         """Returns collected statistics, if present."""
         if self._collected_samples == 0:
             raise StatisticsNotCollectedError()
         return self._get_statistics()
 
     @abstractmethod
-    def _get_statistics(self) -> None:
+    def _get_statistics(self) -> Any:
         pass
 
     def enable(self) -> None:

nncf/common/tensor_statistics/statistic_point.py

@@ -13,8 +13,7 @@
 from typing import Any, Callable, Generator, Optional, Tuple, cast
 
 from nncf.common.graph.transformations.commands import TargetPoint
-from nncf.common.tensor import NNCFTensor
-from nncf.common.tensor_statistics.collectors import TensorStatisticCollectorBase
+from nncf.experimental.common.tensor_statistics.collectors import TensorCollector
 
 
 class StatisticPoint:
@@ -25,7 +24,7 @@ class StatisticPoint:
     algorithm implies on what algorithm nedeed this statistics.
     """
 
-    def __init__(self, target_point: TargetPoint, tensor_collector: TensorStatisticCollectorBase, algorithm: str):
+    def __init__(self, target_point: TargetPoint, tensor_collector: TensorCollector, algorithm: str):
         self.target_point = target_point
         self.algorithm_to_tensor_collectors = {algorithm: [tensor_collector]}
 
@@ -36,11 +35,6 @@ def __eq__(self, other: Any) -> bool:
             and self.algorithm_to_tensor_collectors == other.self.algorithm_to_tensor_collectors,
         )
 
-    def register_tensor(self, x: NNCFTensor) -> None:
-        for tensor_collectors in self.algorithm_to_tensor_collectors.values():
-            for tensor_collector in tensor_collectors:
-                tensor_collector.register_input(x)
-
 
 class StatisticPointsContainer(UserDict):  # type: ignore
     """
@@ -88,7 +82,7 @@ def iter_through_statistic_points_in_target_node(
 
     def get_tensor_collectors(
         self, filter_fn: Optional[Callable[[StatisticPoint], bool]] = None
-    ) -> Generator[Tuple[str, StatisticPoint, TensorStatisticCollectorBase], None, None]:
+    ) -> Generator[Tuple[str, StatisticPoint, TensorCollector], None, None]:
         """
         Returns iterable through all tensor collectors.
 
@@ -115,7 +109,7 @@ def get_algo_statistics_for_node(
         target_node_name: str,
         filter_fn: Callable[[StatisticPoint], bool],
         algorithm: str,
-    ) -> Generator[TensorStatisticCollectorBase, None, None]:
+    ) -> Generator[TensorCollector, None, None]:
         """
         Returns iterable through all statistic collectors in node with target_node_name.