remove min_threshold from NodeActivationQuantizationConfig and from trainable activation quantizer config

Author: irenab

model_compression_toolkit/core/common/quantization/node_quantization_config.py

@@ -101,14 +101,6 @@ def __init__(self, op_cfg: OpQuantizationConfig):
         # TODO irena: computed by compute_activation_bias_correction. shouldnt really be here
         self.activation_bias_correction_term = None
 
-        # TODO irena remove along with set_qc. Keeping for eq and hash to work without set_qc being called
-        self.min_threshold = None
-
-    def set_qc(self, qc: QuantizationConfig):
-        """ TODO irena: temporary keep all the attributes as before not to break all code at once.
-             Eventually all of them should be removed from here. """
-        self.min_threshold = qc.min_threshold
-
     @property
     def enable_activation_quantization(self):
         return self.quant_mode == ActivationQuantizationMode.QUANT
@@ -148,14 +140,12 @@ def __eq__(self, other: Any) -> bool:
 
         return self.activation_quantization_method == other.activation_quantization_method and \
                self.activation_n_bits == other.activation_n_bits and \
-               self.quant_mode == other.quant_mode and \
-               self.min_threshold == other.min_threshold
+               self.quant_mode == other.quant_mode
 
     def __hash__(self):
         return hash((self.activation_quantization_method,
                      self.activation_n_bits,
-                     self.quant_mode,
-                     self.min_threshold))
+                     self.quant_mode))
 
 
 class WeightsAttrQuantizationConfig:

model_compression_toolkit/core/common/quantization/quantization_params_generation/qparams_activations_computation.py

@@ -64,7 +64,7 @@ def compute_activation_qparams(quant_cfg: QuantizationConfig,
         node_activation_quant_cfg.activation_n_bits,
         min_value,
         max_value,
-        min_threshold=node_activation_quant_cfg.min_threshold,
+        min_threshold=quant_cfg.min_threshold,
         quant_error_method=quant_cfg.activation_error_method,
         is_signed=signed
     )

model_compression_toolkit/core/graph_prep_runner.py

@@ -158,7 +158,6 @@ def get_finalized_graph(initial_graph: Graph,
     #  not to break all the code at once. Eventually we need to handle quant_config directly, without injecting into candidates.
     #  TODO 2: Also we adjust candidates for single precision, which we shouldn't do here.
     def update(qc):
-        qc.activation_quantization_cfg.set_qc(quant_config)
         qc.weights_quantization_cfg.set_qc(quant_config)
         for attr_cfg in qc.weights_quantization_cfg.get_all_weight_attrs_configs().values():
             attr_cfg.weights_error_method = quant_config.weights_error_method

model_compression_toolkit/trainable_infrastructure/common/get_quantizer_config.py

@@ -76,7 +76,6 @@ def get_trainable_quantizer_activation_config(
                                               final_cfg.activation_n_bits,
                                               final_cfg.activation_quantization_params,
                                               final_cfg.enable_activation_quantization,
-                                              final_cfg.min_threshold,
                                               activation_quantization_candidates)
 
 

model_compression_toolkit/trainable_infrastructure/common/trainable_quantizer_config.py

@@ -44,7 +44,6 @@ def __init__(self,
                  activation_n_bits: int,
                  activation_quantization_params: Dict,
                  enable_activation_quantization: bool,
-                 min_threshold: float,
                  activation_quantization_candidates: List[TrainableQuantizerCandidateConfig] = None,
                  ):
         """
@@ -55,13 +54,11 @@ def __init__(self,
             activation_n_bits (int): Number of bits to quantize the activations.
             activation_quantization_params (Dict): Dictionary that contains activation quantization params.
             enable_activation_quantization (bool): Whether to quantize the layer's activations or not.
-            min_threshold (float): Minimum threshold to use during thresholds selection.
         """
         self.activation_quantization_method = activation_quantization_method
         self.activation_n_bits = activation_n_bits
         self.activation_quantization_params = activation_quantization_params
         self.enable_activation_quantization = enable_activation_quantization
-        self.min_threshold = min_threshold
         self.activation_bits_candidates = activation_quantization_candidates
 
 

model_compression_toolkit/trainable_infrastructure/keras/config_serialization.py

@@ -83,7 +83,6 @@ def config_deserialization(in_config: dict) -> Union[TrainableQuantizerWeightsCo
         return TrainableQuantizerActivationConfig(activation_quantization_method=QuantizationMethod(in_config[C.ACTIVATION_QUANTIZATION_METHOD]),
                                                   activation_n_bits=in_config[C.ACTIVATION_N_BITS],
                                                   activation_quantization_params=in_config[C.ACTIVATION_QUANTIZATION_PARAMS],
-                                                  enable_activation_quantization=in_config[C.ENABLE_ACTIVATION_QUANTIZATION],
-                                                  min_threshold=in_config[C.MIN_THRESHOLD])
+                                                  enable_activation_quantization=in_config[C.ENABLE_ACTIVATION_QUANTIZATION])
     else:
         raise NotImplemented  # pragma: no cover

tests/keras_tests/non_parallel_tests/test_lp_search_bitwidth.py

@@ -74,7 +74,6 @@ def dummy_representative_dataset():
         for node in graph.nodes:
             # TODO irena remove set_qc:
             for c in node.quantization_cfg.candidates_quantization_cfg:
-                c.activation_quantization_cfg.set_qc(core_config.quantization_config)
                 c.weights_quantization_cfg.set_qc(core_config.quantization_config)
                 for attr_cfg in c.weights_quantization_cfg.get_all_weight_attrs_configs().values():
                     attr_cfg.weights_error_method = core_config.quantization_config.weights_error_method

tests/keras_tests/trainable_infrastructure_tests/base_keras_trainable_infra_test.py

@@ -123,5 +123,4 @@ def get_activation_quantization_config(self):
         return TrainableQuantizerActivationConfig(activation_quantization_method=QuantizationMethod.POWER_OF_TWO,
                                                   activation_n_bits=8,
                                                   activation_quantization_params={},
-                                                  enable_activation_quantization=True,
-                                                  min_threshold=0)
+                                                  enable_activation_quantization=True)

tests/keras_tests/trainable_infrastructure_tests/trainable_keras/test_keras_base_quantizer.py

@@ -27,9 +27,6 @@
 
 class TestKerasBaseWeightsQuantizer(BaseKerasTrainableInfrastructureTest):
 
-    def __init__(self, unit_test):
-        super().__init__(unit_test)
-
     def get_weights_quantization_config(self):
         return TrainableQuantizerWeightsConfig(weights_quantization_method=QuantizationMethod.UNIFORM,
                                                weights_n_bits=8,
@@ -68,8 +65,7 @@ def get_activation_quantization_config(self):
         return TrainableQuantizerActivationConfig(activation_quantization_method=QuantizationMethod.UNIFORM,
                                                   activation_n_bits=8,
                                                   activation_quantization_params={},
-                                                  enable_activation_quantization=True,
-                                                  min_threshold=0)
+                                                  enable_activation_quantization=True)
 
     def run_test(self):
         with self.unit_test.assertRaises(Exception) as e:

tests/pytorch_tests/trainable_infrastructure_tests/base_pytorch_trainable_infra_test.py

@@ -110,5 +110,5 @@ def get_activation_quantization_config(self, quant_method=QuantizationMethod.POW
         return TrainableQuantizerActivationConfig(activation_quantization_method=quant_method,
                                                   activation_n_bits=8,
                                                   activation_quantization_params=activation_quant_params or {},
-                                                  enable_activation_quantization=True,
-                                                  min_threshold=0)
+                                                  enable_activation_quantization=True)
+