mct-nightly 2.0.0.20240409.404__tar.gz → 2.0.0.20240411.406__tar.gz

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.0.0.20240409.404
+Version: 2.0.0.20240411.406
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/mct_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.0.0.20240409.404
+Version: 2.0.0.20240411.406
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/mct_nightly.egg-info/SOURCES.txt

@@ -175,6 +175,7 @@ model_compression_toolkit/core/keras/graph_substitutions/substitutions/activatio
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/batchnorm_folding.py
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/batchnorm_reconstruction.py
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/batchnorm_refusing.py
+model_compression_toolkit/core/keras/graph_substitutions/substitutions/concat_threshold_update.py
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/dwconv_to_conv.py
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/input_scaling.py
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/linear_collapsing.py
@@ -237,6 +238,7 @@ model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/__init_
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/batchnorm_folding.py
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/batchnorm_reconstruction.py
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/batchnorm_refusing.py
+model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/concat_threshold_update.py
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/const_holder_conv.py
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/functional_batch_norm.py
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/functional_layer_norm.py

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/__init__.py

@@ -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__ = "2.0.0.20240409.000404"
+__version__ = "2.0.0.20240411.000406"

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/core/common/quantization/node_quantization_config.py

@@ -106,6 +106,7 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
         self.z_threshold = qc.z_threshold
         self.shift_negative_ratio = qc.shift_negative_ratio
         self.shift_negative_threshold_recalculation = qc.shift_negative_threshold_recalculation
+        self.concat_threshold_update = qc.concat_threshold_update
 
     def quantize_node_output(self,
                              tensors: Any) -> Any:
@@ -219,7 +220,7 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
                self.shift_negative_activation_correction == other.shift_negative_activation_correction and \
                self.z_threshold == other.z_threshold and \
                self.shift_negative_ratio == other.shift_negative_ratio and \
-               self.shift_negative_threshold_recalculation == other.shift_negative_threshold_recalculation
+               self.shift_negative_threshold_recalculation == other.shift_negative_threshold_recalculation 
 
     def __hash__(self):
         return hash((self.activation_quantization_fn,

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/core/common/quantization/quantization_config.py

@@ -62,7 +62,8 @@ class QuantizationConfig:
                  residual_collapsing: bool = True,
                  shift_negative_ratio: float = 0.05,
                  shift_negative_threshold_recalculation: bool = False,
-                 shift_negative_params_search: bool = False):
+                 shift_negative_params_search: bool = False,
+                 concat_threshold_update: bool = False):
         """
         Class to wrap all different parameters the library quantize the input model according to.
 
@@ -117,6 +118,7 @@ class QuantizationConfig:
         self.shift_negative_ratio = shift_negative_ratio
         self.shift_negative_threshold_recalculation = shift_negative_threshold_recalculation
         self.shift_negative_params_search = shift_negative_params_search
+        self.concat_threshold_update = concat_threshold_update
 
     def __repr__(self):
         return str(self.__dict__)

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/core/common/quantization/quantization_params_generation/qparams_computation.py

@@ -42,14 +42,14 @@ def calculate_quantization_params(graph: Graph,
 
     """
 
-    Logger.info(f"Running quantization parameters search. "
+    Logger.info(f"\nRunning quantization parameters search. "
                 f"This process might take some time, "
                 f"depending on the model size and the selected quantization methods.\n")
 
     # Create a list of nodes to compute their thresholds
     nodes_list: List[BaseNode] = nodes if specific_nodes else graph.nodes()
 
-    for n in tqdm(nodes_list, "Calculating quantization params"):  # iterate only nodes that we should compute their thresholds
+    for n in tqdm(nodes_list, "Calculating quantization parameters"):  # iterate only nodes that we should compute their thresholds
         for candidate_qc in n.candidates_quantization_cfg:
             for attr in n.get_node_weights_attributes():
                 if n.is_weights_quantization_enabled(attr):

mct-nightly-2.0.0.20240411.406/model_compression_toolkit/core/keras/graph_substitutions/substitutions/concat_threshold_update.py

@@ -0,0 +1,66 @@
+# Copyright 2024 Sony Semiconductor Israel, Inc. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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 tensorflow.keras.layers import Concatenate
+import tensorflow as tf
+
+from model_compression_toolkit.core import common
+from model_compression_toolkit.core.common import Graph, BaseNode
+from model_compression_toolkit.core.common.graph.graph_matchers import NodeOperationMatcher
+from model_compression_toolkit.constants import THRESHOLD
+
+
+
+class ConcatThresholdUpdate(common.BaseSubstitution):
+
+
+    """
+    Find concat layers and match their prior layers thresholds unless prior layer outputs to multiple layers.
+    """
+
+    def __init__(self):
+        """
+        Initialize a threshold_updater object.
+        """
+        concatination_node = NodeOperationMatcher(Concatenate) | \
+            NodeOperationMatcher(tf.concat) 
+        super().__init__(matcher_instance=concatination_node)
+
+    def substitute(self,
+                   graph: Graph,
+                   node: BaseNode) -> Graph:
+        """
+        Update previous layers thresholds to match concatinations quantization thresholds. No change if
+        previous layer outputs to multiple layers. No change in case of uniform quantization. 
+        No change in case of multiple quantization candidates (mixed precision).
+
+
+        Args:
+            graph: Graph we apply the substitution on.
+            node: Node refference to edit previous nodes thresholds.
+
+        Returns:
+            Graph after applying the substitution.
+        """
+
+        if len(node.candidates_quantization_cfg) == 1 and THRESHOLD in node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params:
+            concat_threshold = node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params[THRESHOLD]
+            prev_nodes = graph.get_prev_nodes(node)
+            for prev_node in prev_nodes:
+                if len(graph.get_next_nodes(prev_node))==1 and prev_node.type != Concatenate and prev_node.type != tf.concat:
+                    prev_node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params[THRESHOLD] = concat_threshold
+
+        return graph

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/core/keras/keras_implementation.py

@@ -80,7 +80,8 @@ from model_compression_toolkit.core.keras.graph_substitutions.substitutions.line
 from model_compression_toolkit.core.keras.graph_substitutions.substitutions.residual_collapsing import \
     keras_residual_collapsing
 from model_compression_toolkit.core.keras.graph_substitutions.substitutions.input_scaling import InputScaling, \
-    InputScalingWithPad
+    InputScalingWithPad 
+from model_compression_toolkit.core.keras.graph_substitutions.substitutions.concat_threshold_update import ConcatThresholdUpdate
 from model_compression_toolkit.core.keras.graph_substitutions.substitutions.relu_bound_to_power_of_2 import \
     ReLUBoundToPowerOfTwo
 from model_compression_toolkit.core.keras.graph_substitutions.substitutions.multi_head_attention_decomposition import \
@@ -300,8 +301,8 @@ class KerasImplementation(FrameworkImplementation):
         """
         return keras_op2d_add_const_collapsing()
 
-    def get_substitutions_post_statistics_collection(self, quant_config: QuantizationConfig) \
-            -> List[common.BaseSubstitution]:
+    def get_substitutions_post_statistics_collection(self, 
+                                                     quant_config: QuantizationConfig) -> List[common.BaseSubstitution]:
         """
         Return a list of the framework substitutions used after we collect statistics.
 
@@ -317,6 +318,8 @@ class KerasImplementation(FrameworkImplementation):
         if quant_config.input_scaling:
             substitutions_list.append(InputScaling())
             substitutions_list.append(InputScalingWithPad())
+        if quant_config.concat_threshold_update:
+            substitutions_list.append(ConcatThresholdUpdate())
         return substitutions_list
 
 

mct-nightly-2.0.0.20240411.406/model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/concat_threshold_update.py

@@ -0,0 +1,69 @@
+# Copyright 2024 Sony Semiconductor Israel, Inc. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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 typing import List
+
+import torch
+
+from model_compression_toolkit.core import common
+from model_compression_toolkit.core.common.graph.base_graph import Graph
+from model_compression_toolkit.core.common.graph.graph_matchers import NodeOperationMatcher
+from model_compression_toolkit.core.common.graph.base_node import BaseNode
+from model_compression_toolkit.constants import THRESHOLD
+
+
+class ConcatThresholdUpdate(common.BaseSubstitution):
+    """
+    Find concat layers and match their prior layers thresholds unless prior layer outputs to multiple layers.
+    """
+
+
+    def __init__(self):
+        """
+        Initialize a threshold_updater object.
+        """
+        concatination_node = NodeOperationMatcher(torch.cat) | \
+            NodeOperationMatcher(torch.concat) 
+        super().__init__(matcher_instance=concatination_node)
+
+    def substitute(self,
+                   graph: Graph,
+                   node: BaseNode) -> Graph:
+        """
+        Update previous layers thresholds to match concatinations quantization thresholds. No change if
+        previous layer outputs to multiple layers. No change in case of uniform quantization. 
+        No change in case of multiple quantization candidates (mixed precision).
+
+
+        Args:
+            graph: Graph we apply the substitution on.
+            node: Node refference to edit previous nodes thresholds.
+
+        Returns:
+            Graph after applying the substitution.
+        """
+    
+        if len(node.candidates_quantization_cfg) == 1 and THRESHOLD in node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params:
+            concat_threshold = node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params[THRESHOLD]
+            prev_nodes = graph.get_prev_nodes(node)
+            for prev_node in prev_nodes:
+                if len(graph.get_next_nodes(prev_node))==1 and prev_node.type != torch.cat and prev_node.type != torch.concat:
+                    prev_node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params[THRESHOLD] = concat_threshold
+
+        return graph
+
+
+
+

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/core/pytorch/pytorch_implementation.py

@@ -73,6 +73,8 @@ from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.vi
     VirtualActivationWeightsComposition
 from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.weights_activation_split import \
     WeightsActivationSplit
+from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.concat_threshold_update import \
+    ConcatThresholdUpdate
 from model_compression_toolkit.core.pytorch.hessian.activation_trace_hessian_calculator_pytorch import \
     ActivationTraceHessianCalculatorPytorch
 from model_compression_toolkit.core.pytorch.hessian.weights_trace_hessian_calculator_pytorch import \
@@ -302,6 +304,8 @@ class PytorchImplementation(FrameworkImplementation):
             substitutions_list.append(pytorch_softmax_shift())
         if quant_config.input_scaling:
             Logger.critical('Input scaling is currently not supported for Pytorch.')
+        if quant_config.concat_threshold_update:
+            substitutions_list.append(ConcatThresholdUpdate())
         return substitutions_list
 
 

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/exporter/model_wrapper/keras/builder/fully_quantized_model_builder.py

@@ -90,7 +90,7 @@ if FOUND_TF:
                                                                                         fw_impl=C.keras.keras_implementation.KerasImplementation())).build_model()
         exportable_model.trainable = False
 
-        Logger.info("Please run your accuracy evaluation on the exported quantized model to verify it's accuracy.\n"
+        Logger.info("\nPlease run your accuracy evaluation on the exported quantized model to verify it's accuracy.\n"
                     "Checkout the FAQ and Troubleshooting pages for resolving common issues and improving the quantized model accuracy:\n"
                     "FAQ: https://github.com/sony/model_optimization/tree/main/FAQ.md\n"
                     "Quantization Troubleshooting: https://github.com/sony/model_optimization/tree/main/quantization_troubleshooting.md")

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/exporter/model_wrapper/pytorch/builder/fully_quantized_model_builder.py

@@ -82,7 +82,7 @@ if FOUND_TORCH:
                                                           get_activation_quantizer_holder(n,
                                                                                           fw_impl=C.pytorch.pytorch_implementation.PytorchImplementation())).build_model()
 
-        Logger.info("Please run your accuracy evaluation on the exported quantized model to verify it's accuracy.\n"
+        Logger.info("\nPlease run your accuracy evaluation on the exported quantized model to verify it's accuracy.\n"
                     "Checkout the FAQ and Troubleshooting pages for resolving common issues and improving the quantized model accuracy:\n"
                     "FAQ: https://github.com/sony/model_optimization/tree/main/FAQ.md\n"
                     "Quantization Troubleshooting: https://github.com/sony/model_optimization/tree/main/quantization_troubleshooting.md")

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/gptq/keras/gptq_training.py

@@ -301,21 +301,23 @@ class KerasGPTQTrainer(GPTQTrainer):
         Returns: None
 
         """
-        for _ in tqdm(range(n_epochs)):
-            for data in tqdm(data_function()):
-                input_data = [d * self.input_scale for d in data]
-
-                loss_value_step, grads = self.nano_training_step(input_data, in_compute_gradients,
-                                                                 in_optimizer_with_param, is_training)
-                # Run one step of gradient descent by updating
-                # the value of the variables to minimize the loss.
-                for i, (o, p) in enumerate(in_optimizer_with_param):
-                    o.apply_gradients(zip(grads[i], p))
-                if self.gptq_config.log_function is not None:
-                    self.gptq_config.log_function(loss_value_step, grads[0], in_optimizer_with_param[0][-1],
-                                                  self.compare_points)
-                self.loss_list.append(loss_value_step.numpy())
-                Logger.debug(f'last loss value: {self.loss_list[-1]}')
+        with tqdm(range(n_epochs), "Running GPTQ optimization") as epochs_pbar:
+            for _ in epochs_pbar:
+                with tqdm(data_function(), position=1, leave=False) as data_pbar:
+                    for data in data_pbar:
+                        input_data = [d * self.input_scale for d in data]
+
+                        loss_value_step, grads = self.nano_training_step(input_data, in_compute_gradients,
+                                                                         in_optimizer_with_param, is_training)
+                        # Run one step of gradient descent by updating
+                        # the value of the variables to minimize the loss.
+                        for i, (o, p) in enumerate(in_optimizer_with_param):
+                            o.apply_gradients(zip(grads[i], p))
+                        if self.gptq_config.log_function is not None:
+                            self.gptq_config.log_function(loss_value_step, grads[0], in_optimizer_with_param[0][-1],
+                                                          self.compare_points)
+                        self.loss_list.append(loss_value_step.numpy())
+                        Logger.debug(f'last loss value: {self.loss_list[-1]}')
 
     def update_graph(self):
         """

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/gptq/keras/quantizer/regularization_factory.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+from tqdm import tqdm
 from typing import Callable
 
 from model_compression_toolkit.gptq import RoundingType, GradientPTQConfig, GradientPTQConfig
@@ -35,7 +36,7 @@ def get_regularization(gptq_config: GradientPTQConfig, representative_data_gen:
     if gptq_config.rounding_type == RoundingType.SoftQuantizer:
         # dry run on the representative dataset to count number of batches
         num_batches = 0
-        for _ in representative_data_gen():
+        for _ in tqdm(representative_data_gen(), "GPTQ initialization"):
             num_batches += 1
 
         return SoftQuantizerRegularization(total_gradient_steps=num_batches * gptq_config.n_epochs)

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/gptq/pytorch/gptq_training.py

@@ -248,22 +248,24 @@ class PytorchGPTQTrainer(GPTQTrainer):
             data_function: A callable function that give a batch of samples.
             n_epochs: Number of update iterations of representative dataset.
         """
-        for _ in tqdm(range(n_epochs)):
-            for data in tqdm(data_function()):
-                input_data = [d * self.input_scale for d in data]
-                input_tensor = to_torch_tensor(input_data)
-                y_float = self.float_model(input_tensor)  # running float model
-                loss_value, grads = self.compute_gradients(y_float, input_tensor)
-                # Run one step of gradient descent by updating the value of the variables to minimize the loss.
-                for (optimizer, _) in self.optimizer_with_param:
-                    optimizer.step()
-                    optimizer.zero_grad()
-                if self.gptq_config.log_function is not None:
-                    self.gptq_config.log_function(loss_value.item(),
-                                                  torch_tensor_to_numpy(grads),
-                                                  torch_tensor_to_numpy(self.optimizer_with_param[0][-1]))
-                self.loss_list.append(loss_value.item())
-                Logger.debug(f'last loss value: {self.loss_list[-1]}')
+        with tqdm(range(n_epochs), "Running GPTQ optimization") as epochs_pbar:
+            for _ in epochs_pbar:
+                with tqdm(data_function(), position=1, leave=False) as data_pbar:
+                    for data in data_pbar:
+                        input_data = [d * self.input_scale for d in data]
+                        input_tensor = to_torch_tensor(input_data)
+                        y_float = self.float_model(input_tensor)  # running float model
+                        loss_value, grads = self.compute_gradients(y_float, input_tensor)
+                        # Run one step of gradient descent by updating the value of the variables to minimize the loss.
+                        for (optimizer, _) in self.optimizer_with_param:
+                            optimizer.step()
+                            optimizer.zero_grad()
+                        if self.gptq_config.log_function is not None:
+                            self.gptq_config.log_function(loss_value.item(),
+                                                          torch_tensor_to_numpy(grads),
+                                                          torch_tensor_to_numpy(self.optimizer_with_param[0][-1]))
+                        self.loss_list.append(loss_value.item())
+                        Logger.debug(f'last loss value: {self.loss_list[-1]}')
 
     def update_graph(self) -> Graph:
         """

{mct-nightly-2.0.0.20240409.404 → mct-nightly-2.0.0.20240411.406}/model_compression_toolkit/gptq/pytorch/quantizer/regularization_factory.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+from tqdm import tqdm
 from typing import Callable
 
 from model_compression_toolkit.gptq import RoundingType, GradientPTQConfig, GradientPTQConfig
@@ -35,7 +36,7 @@ def get_regularization(gptq_config: GradientPTQConfig, representative_data_gen:
     if gptq_config.rounding_type == RoundingType.SoftQuantizer:
         # dry run on the representative dataset to count number of batches
         num_batches = 0
-        for _ in representative_data_gen():
+        for _ in tqdm(representative_data_gen(), "GPTQ initialization"):
             num_batches += 1
 
         return SoftQuantizerRegularization(total_gradient_steps=num_batches * gptq_config.n_epochs)