PyPI - mct-nightly - Versions diffs - 2.4.0.20250707.643__tar.gz → 2.4.0.20250709.616__tar.gz - Mend

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.4.0.20250707.643
+Version: 2.4.0.20250709.616
 Summary: A Model Compression Toolkit for neural networks
 Author-email: ssi-dnn-dev@sony.com
 Classifier: Programming Language :: Python :: 3

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.4.0.20250707.643
+Version: 2.4.0.20250709.616
 Summary: A Model Compression Toolkit for neural networks
 Author-email: ssi-dnn-dev@sony.com
 Classifier: Programming Language :: Python :: 3

@@ -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__ = "2.4.0.20250707.000643"
+__version__ = "2.4.0.20250709.000616"

@@ -57,19 +57,21 @@ def create_stats_collector_for_node(node: common.BaseNode,
 def create_tensor2node(graph: common.Graph,
-                       node: common.BaseNode):
+                       node: common.BaseNode,
+                       next_node_output_channel_axis: int):
     """
     Force statistic collector creation and assignment for a node.
     Args:
         graph: Graph of the node (for retrieving the current tensor).
         node: Node to create a tensor for.
+        next_node_output_channel_axis: channel output axis of next node.
     """
     current_sc = graph.get_out_stats_collector(node)
     is_list_nostat_collectors = isinstance(current_sc, list) and len(
         [sc for sc in current_sc if not isinstance(sc, common.NoStatsCollector)]) == 0
     if isinstance(current_sc, common.NoStatsCollector) or current_sc is None or is_list_nostat_collectors:
-        stats_collector = common.StatsCollector(node.out_channel_axis)
+        stats_collector = common.StatsCollector(next_node_output_channel_axis if node.out_channel_axis is None else node.out_channel_axis)
         graph.set_out_stats_collector_to_node(node, stats_collector)
@@ -175,7 +177,8 @@ class ModelCollector:
                 for ie in graph.incoming_edges(n):
                     input_node = ie.source_node
                     create_tensor2node(graph,
-                                       input_node)
+                                       input_node,
+                                       n.out_channel_axis)
             if sc is not None:
                 graph.set_out_stats_collector_to_node(n, sc)

@@ -98,6 +98,9 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
         self.activation_quantization_params = {}
         # TODO: computed by compute_activation_bias_correction. Probably shouldnt be here.
         self.activation_bias_correction_term = None
+        # Z-threshold is a global param from QuantizationConfig, however it can be overridden per node by NetworkEditor.
+        # Since activation qparams are re-computed in several places, it's easier to keep it here and update it once.
+        self.z_threshold = None
     @property
     def enable_activation_quantization(self):
@@ -119,7 +122,7 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
             activation_params: Dictionary that contains weight quantization params.
         """
-        assert self.quant_mode == ActivationQuantizationMode.QUANT
+        assert self.quant_mode == ActivationQuantizationMode.QUANT or self.quant_mode == ActivationQuantizationMode.FLN_QUANT
         for param_name, param_value in activation_params.items():
             self.activation_quantization_params[param_name] = param_value

@@ -47,9 +47,12 @@ def compute_activation_qparams(quant_cfg: QuantizationConfig,
         node_activation_quant_cfg.activation_quantization_method, no_clipping=node_prior_info.is_output_bounded())
     # Extract and filter histogram data from the statistics container.
+    z_threshold = quant_cfg.z_threshold
+    if node_activation_quant_cfg.z_threshold is not None:
+        z_threshold = node_activation_quant_cfg.z_threshold
     bins_values, bins_counts = _get_histogram_data(out_stats_container,
                                                    activation_error_method=quant_cfg.activation_error_method,
-                                                   z_threshold=quant_cfg.z_threshold)
+                                                   z_threshold=z_threshold)
     # Retrieve the minimum and maximum values from the statistics container.
     min_value, max_value = out_stats_container.get_min_max_values()

@@ -105,7 +105,7 @@ def calculate_quantization_params(graph: Graph,
                     attr_cfg.weights_channels_axis = ChannelAxisMapping(output_channels_axis, attr_cfg.weights_channels_axis.input)
                     attr_cfg.set_weights_quantization_param(weights_params)
-            if n.is_activation_quantization_enabled():
+            if n.is_activation_quantization_enabled() or n.is_fln_quantization():
                 # If node's activations should be quantized as well, we compute its activation quantization parameters
                 activation_params = compute_activation_qparams(quant_cfg=quant_cfg,
                                                                node_activation_quant_cfg=candidate_qc.activation_quantization_cfg,

mct-nightly 2.4.0.20250707.643__tar.gz → 2.4.0.20250709.616__tar.gz

mct-nightly 2.4.0.20250707.643tar.gz → 2.4.0.20250709.616tar.gz