Drop binarization algorithm

README.md

@@ -42,7 +42,6 @@ learning frameworks.
 | :--- | :---: | :---: |
 |[Quantization Aware Training](./docs/compression_algorithms/Quantization.md) | Supported | Supported |
 |[Mixed-Precision Quantization](./docs/compression_algorithms/Quantization.md#mixed-precision-quantization) | Supported | Not supported |
-|[Binarization](./docs/compression_algorithms/Binarization.md) | Supported | Not supported |
 |[Sparsity](./docs/compression_algorithms/Sparsity.md) | Supported | Supported |
 |[Filter pruning](./docs/compression_algorithms/Pruning.md) | Supported | Supported |
 |[Movement pruning](./nncf/experimental/torch/sparsity/movement/MovementSparsity.md) | Experimental | Not supported |

docs/Algorithms.md

@@ -12,9 +12,6 @@ Each compression method receives its own hyperparameters that are organized as a
   - Automatic bitwidth assignment based on HAWQ
   - Automatic quantization parameter selection and activation quantizer setup based on HW config preset
   - Automatic bitwidth assignment mode AutoQ, based on HAQ, a Deep Reinforcement Learning algorithm to select best mixed precision given quality metric and HW type.
-- [Binarization](./compression_algorithms/Binarization.md)
-  - XNOR, DoReFa weight binarization
-  - Scale/threshold based per-channel activation binarization
 - [Sparsity](./compression_algorithms/Sparsity.md)
   - Magnitude sparsity
   - Regularization-based (RB) sparsity

docs/ModelZoo.md

@@ -127,30 +127,6 @@ The applied quantization compression algorithms are divided into two broad categ
       <td><a href="../examples/torch/classification/configs/quantization/mobilenet_v3_small_imagenet_int8.json">Config</a></td>
       <td><a href="https://storage.openvinotoolkit.org/repositories/nncf/models/develop/torch/mobilenet_v3_small_imagenet_int8.pth">Download</a></td>
     </tr>
-    <tr>
-      <td align="left">ResNet-18</td>
-      <td align="left">-</td>
-      <td>ImageNet</td>
-      <td>69.76</td>
-      <td><a href="../examples/torch/classification/configs/binarization/resnet18_imagenet.json">Config</a></td>
-      <td>-</td>
-    </tr>
-    <tr>
-      <td align="left">ResNet-18</td>
-      <td align="left">• Binarization: XNOR (weights), scale/threshold (activations)</td>
-      <td>ImageNet</td>
-      <td>61.67 (8.09)</td>
-      <td><a href="../examples/torch/classification/configs/binarization/resnet18_imagenet_binarization_xnor.json">Config</a></td>
-      <td><a href="https://storage.openvinotoolkit.org/repositories/nncf/models/develop/torch/resnet18_imagenet_binarization_xnor.pth">Download</a></td>
-    </tr>
-    <tr>
-      <td align="left">ResNet-18</td>
-      <td align="left">• Binarization: DoReFa (weights), scale/threshold (activations)</td>
-      <td>ImageNet</td>
-      <td>61.63 (8.13)</td>
-      <td><a href="../examples/torch/classification/configs/binarization/resnet18_imagenet_binarization_dorefa.json">Config</a></td>
-      <td><a href="https://storage.openvinotoolkit.org/repositories/nncf/models/develop/torch/resnet18_imagenet_binarization_dorefa.pth">Download</a></td>
-    </tr>
     <tr>
       <td align="left">ResNet-18</td>
       <td align="left">• Filter pruning: 40%, magnitude criterion</td>

examples/torch/classification/staged_quantization_worker.py

@@ -52,7 +52,6 @@
 from nncf.config.schemata.defaults import STAGED_QUANTIZATION_BASE_LR
 from nncf.config.schemata.defaults import STAGED_QUANTIZATION_BASE_WD
 from nncf.torch import create_compressed_model
-from nncf.torch.binarization.algo import BinarizationController
 from nncf.torch.checkpoint_loading import load_state
 from nncf.torch.initialization import default_criterion_fn
 from nncf.torch.initialization import register_default_init_args
@@ -186,9 +185,9 @@ def autoq_eval_fn(model, eval_loader):
     if model_state_dict is not None:
         load_state(model, model_state_dict, is_resume=True)
 
-    if not isinstance(compression_ctrl, (BinarizationController, QuantizationController)):
+    if not isinstance(compression_ctrl, QuantizationController):
         raise nncf.InternalError(
-            "The stage quantization sample worker may only be run with the binarization and quantization algorithms!"
+            "The stage quantization sample worker may only be run with the quantization algorithms!"
         )
 
     model, _ = prepare_model_for_execution(model, config)

examples/torch/common/utils.py

@@ -166,9 +166,7 @@ def is_staged_quantization(config):
     if isinstance(compression_config, list):
         compression_config = compression_config[0]
     algo_type = compression_config.get("algorithm")
-    if algo_type is not None and algo_type == "binarization":
-        return True
-    if algo_type == "quantization" and compression_config.get("params", {}):
+    if algo_type is not None and algo_type == "quantization" and compression_config.get("params", {}):
         return True
     return False
 

examples/torch/object_detection/layers/modules/l2norm.py

@@ -19,9 +19,7 @@
 from nncf.torch.utils import no_jit_trace
 
 
-@register_module(
-    ignored_algorithms=["quantization", "binarization", "const_sparsity", "magnitude_sparsity", "rb_sparsity"]
-)
+@register_module(ignored_algorithms=["quantization", "const_sparsity", "magnitude_sparsity", "rb_sparsity"])
 class L2Norm(nn.Module):
     def __init__(self, n_channels, scale, eps, across_spatial=0, channel_shared=0):
         super().__init__()

nncf/common/statistics.py

@@ -94,10 +94,6 @@ def filter_pruning(self) -> Optional[FilterPruningStatistics]:
         """
         return self._storage.get("filter_pruning")
 
-    @property
-    def binarization(self) -> None:
-        raise NotImplementedError
-
     def register(self, algorithm_name: str, stats: Statistics):
         """
         Registers statistics for the algorithm.
@@ -108,7 +104,6 @@ def register(self, algorithm_name: str, stats: Statistics):
             * const_sparsity
             * quantization
             * filter_pruning
-            * binarization
 
         :param stats: Statistics of the algorithm.
         """
@@ -120,7 +115,6 @@ def register(self, algorithm_name: str, stats: Statistics):
             "const_sparsity",
             "quantization",
             "filter_pruning",
-            "binarization",
         ]
         if algorithm_name not in available_algorithms:
             raise ValueError(

nncf/config/definitions.py

@@ -16,7 +16,6 @@
 EARLY_EXIT_TRAINING_MODE_NAME_IN_CONFIG = "early_exit"
 EXPERIMENTAL_QUANTIZATION_ALGO_NAME_IN_CONFIG = "experimental_quantization"
 BOOTSTRAP_NAS_ALGO_NAME_IN_CONFIG = "bootstrapNAS"
-BINARIZATION_ALGO_NAME_IN_CONFIG = "binarization"
 CONST_SPARSITY_ALGO_NAME_IN_CONFIG = "const_sparsity"
 FILTER_PRUNING_ALGO_NAME_IN_CONFIG = "filter_pruning"
 KNOWLEDGE_DISTILLATION_ALGO_NAME_IN_CONFIG = "knowledge_distillation"
@@ -32,5 +31,4 @@
     RB_SPARSITY_ALGO_NAME_IN_CONFIG: "docs/compression_algorithms/Sparsity.md",
     CONST_SPARSITY_ALGO_NAME_IN_CONFIG: "docs/compression_algorithms/Sparsity.md",
     KNOWLEDGE_DISTILLATION_ALGO_NAME_IN_CONFIG: "docs/compression_algorithms/KnowledgeDistillation.md",
-    BINARIZATION_ALGO_NAME_IN_CONFIG: "docs/compression_algorithms/Binarization.md",
 }

nncf/config/schema.py

@@ -15,7 +15,6 @@
 import jsonschema
 
 from nncf.config.definitions import ALGO_NAME_VS_README_URL
-from nncf.config.definitions import BINARIZATION_ALGO_NAME_IN_CONFIG
 from nncf.config.definitions import CONST_SPARSITY_ALGO_NAME_IN_CONFIG
 from nncf.config.definitions import FILTER_PRUNING_ALGO_NAME_IN_CONFIG
 from nncf.config.definitions import KNOWLEDGE_DISTILLATION_ALGO_NAME_IN_CONFIG
@@ -25,7 +24,6 @@
 from nncf.config.definitions import SCHEMA_VISUALIZATION_URL
 from nncf.config.schemata.accuracy_aware import ACCURACY_AWARE_MODES_VS_SCHEMA
 from nncf.config.schemata.accuracy_aware import ACCURACY_AWARE_TRAINING_SCHEMA
-from nncf.config.schemata.algo.binarization import BINARIZATION_SCHEMA
 from nncf.config.schemata.algo.const_sparsity import CONST_SPARSITY_SCHEMA
 from nncf.config.schemata.algo.filter_pruning import FILTER_PRUNING_SCHEMA
 from nncf.config.schemata.algo.knowledge_distillation import KNOWLEDGE_DISTILLATION_SCHEMA
@@ -50,7 +48,6 @@
     RB_SPARSITY_ALGO_NAME_IN_CONFIG: RB_SPARSITY_SCHEMA,
     KNOWLEDGE_DISTILLATION_ALGO_NAME_IN_CONFIG: KNOWLEDGE_DISTILLATION_SCHEMA,
     CONST_SPARSITY_ALGO_NAME_IN_CONFIG: CONST_SPARSITY_SCHEMA,
-    BINARIZATION_ALGO_NAME_IN_CONFIG: BINARIZATION_SCHEMA,
     **EXPERIMENTAL_REF_VS_ALGO_SCHEMA,
 }
 

nncf/config/schemata/defaults.py

@@ -43,8 +43,6 @@
 STAGED_QUANTIZATION_BASE_LR = 1e-3
 STAGED_QUANTIZATION_BASE_WD = 1e-5
 
-BINARIZATION_MODE = "xnor"
-
 PRUNING_INIT = 0.0
 PRUNING_SCHEDULE = "exponential"
 PRUNING_TARGET = 0.5

nncf/torch/__init__.py

@@ -35,7 +35,6 @@
 
 
 # Required for correct COMPRESSION_ALGORITHMS registry functioning
-from nncf.torch.binarization import algo as binarization_algo
 from nncf.torch.quantization import algo as quantization_algo
 from nncf.torch.sparsity.const import algo as const_sparsity_algo
 from nncf.torch.sparsity.magnitude import algo as magnitude_sparsity_algo