mct-nightly 1.10.0.20231127.post419__tar.gz → 1.10.0.20231129.post414__tar.gz

{mct-nightly-1.10.0.20231127.post419 → mct-nightly-1.10.0.20231129.post414}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 1.10.0.20231127.post419
+Version: 1.10.0.20231129.post414
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct-nightly-1.10.0.20231127.post419 → mct-nightly-1.10.0.20231129.post414}/mct_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 1.10.0.20231127.post419
+Version: 1.10.0.20231129.post414
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct-nightly-1.10.0.20231127.post419 → mct-nightly-1.10.0.20231129.post414}/model_compression_toolkit/core/keras/constants.py

@@ -109,3 +109,6 @@ RELU_POT_BOUND = 8.0
 # TFOpLambda functions:
 ADD = 'add'
 PAD = 'pad'
+
+# Special/Custom layers strings
+COMBINED_NMS = 'combined_non_max_suppression'

{mct-nightly-1.10.0.20231127.post419 → mct-nightly-1.10.0.20231129.post414}/model_compression_toolkit/core/keras/hessian/activation_trace_hessian_calculator_keras.py

@@ -152,7 +152,7 @@ class ActivationTraceHessianCalculatorKeras(TraceHessianCalculatorKeras):
                     # Compute the final approximation for each output index
                     num_node_outputs = len(interest_point_scores[0])
                     for output_idx in range(num_node_outputs):
-                        final_approx_per_output.append(2 * tf.reduce_mean([x[output_idx] for x in interest_point_scores]) / output.shape[-1])
+                        final_approx_per_output.append(tf.reduce_mean([x[output_idx] for x in interest_point_scores]))
 
                     # final_approx_per_output is a list of all approximations (one per output), thus we average them to
                     # get the final score of a node.

{mct-nightly-1.10.0.20231127.post419 → mct-nightly-1.10.0.20231129.post414}/model_compression_toolkit/core/keras/keras_implementation.py

@@ -32,7 +32,8 @@ from model_compression_toolkit.trainable_infrastructure.keras.quantize_wrapper i
 from model_compression_toolkit.core.common.mixed_precision.sensitivity_evaluation import SensitivityEvaluation
 from model_compression_toolkit.core.common.mixed_precision.set_layer_to_bitwidth import set_layer_to_bitwidth
 from model_compression_toolkit.core.common.similarity_analyzer import compute_kl_divergence, compute_cs, compute_mse
-from model_compression_toolkit.core.keras.constants import ACTIVATION, SOFTMAX, SIGMOID, ARGMAX, LAYER_NAME
+from model_compression_toolkit.core.keras.constants import ACTIVATION, SOFTMAX, SIGMOID, ARGMAX, LAYER_NAME, \
+    COMBINED_NMS
 from model_compression_toolkit.core.keras.graph_substitutions.substitutions.batchnorm_reconstruction import \
     keras_batchnorm_reconstruction
 from model_compression_toolkit.core.keras.graph_substitutions.substitutions.virtual_activation_weights_composition import \
@@ -514,7 +515,8 @@ class KerasImplementation(FrameworkImplementation):
 
         if node.layer_class == TFOpLambda:
             node_attr = getattr(node, 'framework_attr', None)
-            if node_attr is not None and (ARGMAX in node_attr[LAYER_NAME]):
+            if node_attr is not None and (ARGMAX in node_attr[LAYER_NAME]
+                                          or COMBINED_NMS in node_attr[LAYER_NAME]):
                 return False
         elif node.layer_class in [tf.math.argmax]:
             return False

{mct-nightly-1.10.0.20231127.post419 → mct-nightly-1.10.0.20231129.post414}/model_compression_toolkit/core/pytorch/hessian/activation_trace_hessian_calculator_pytorch.py

@@ -131,8 +131,8 @@ class ActivationTraceHessianCalculatorPytorch(TraceHessianCalculatorPytorch):
                             break
 
                     trace_hv.append(hessian_trace_approx)
-                ipts_hessian_trace_approx.append(2 * torch.mean(torch.stack(trace_hv)) / output.shape[
-                    -1])  # Get averaged Hessian trace approximation
+
+                ipts_hessian_trace_approx.append(torch.mean(torch.stack(trace_hv)))  # Get averaged Hessian trace approximation
 
             # If a node has multiple outputs, it means that multiple approximations were computed
             # (one per output since granularity is per-tensor). In this case we average the approximations.

{mct-nightly-1.10.0.20231127.post419 → mct-nightly-1.10.0.20231129.post414}/model_compression_toolkit/exporter/model_exporter/keras/fakely_quant_tflite_exporter.py

@@ -57,14 +57,13 @@ class FakelyQuantTFLiteExporter(FakelyQuantKerasExporter):
 
         """
         # 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.
-        _, tmp_file = tempfile.mkstemp(TMP_KERAS_EXPORT_FORMAT)
-        custom_objects = FakelyQuantKerasExporter(self.model,
-                                                  self.is_layer_exportable_fn,
-                                                  tmp_file).export()
+        # Since exporter saves the model, we use a tmp path for saving, and then we delete it automatically.
+        with tempfile.NamedTemporaryFile(suffix=TMP_KERAS_EXPORT_FORMAT) as tmp_file:
+            custom_objects = FakelyQuantKerasExporter(self.model,
+                                                      self.is_layer_exportable_fn,
+                                                      tmp_file.name).export()
 
-        model = keras_load_quantized_model(tmp_file)
-        os.remove(tmp_file)
+            model = keras_load_quantized_model(tmp_file.name)
 
         self.exported_model = tf.lite.TFLiteConverter.from_keras_model(model).convert()
         Logger.info(f'Exporting FQ tflite model to: {self.save_model_path}')

{mct-nightly-1.10.0.20231127.post419 → mct-nightly-1.10.0.20231129.post414}/setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 description-file = README.md
-version = 1.10.0.20231127-000419
+version = 1.10.0.20231129-000414
 
 [egg_info]
 tag_build =