Skip to content

[RedSuns Review] Maintenance Tutorials(MCTWrapper) #1639

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ueno-hiroshi002 merged 10 commits into from
Jan 26, 2026
Merged

[RedSuns Review] Maintenance Tutorials(MCTWrapper) #1639

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
98be03f
Refactoring MCTWrapper tutorials.
ueno-hiroshi002 Jan 20, 2026
6a1a281
Remove local install.
ueno-hiroshi002 Jan 20, 2026
31cfe06
Modify Comments(Add parameter comment, standardized PTQ and Post-... …
ueno-hiroshi002 Jan 22, 2026
b29691a
Modify Comments(Add explanation of sdsp version. Standardize configur…
ueno-hiroshi002 Jan 22, 2026
1223dc6
modify comments(explanation of sdsp version)
ueno-hiroshi002 Jan 22, 2026
41ceaf6
Remove quantize overlap
ueno-hiroshi002 Jan 22, 2026
67b0e06
(Wrapper)Added QuantizationConfig parameters and updated MCTWrapper s…
KazunoriSumiya Jan 22, 2026
483b7ee
Add accuracy notes.
ueno-hiroshi002 Jan 26, 2026
2acb76f
change text
ueno-hiroshi002 Jan 26, 2026
3167487
change text
ueno-hiroshi002 Jan 26, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 16 additions & 9 deletions model_compression_toolkit/wrapper/constants.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,9 +17,19 @@
FW_NAME = 'fw_name'
SDSP_VERSION = 'sdsp_version'

# QuantizationConfig parameters
ACTIVATION_ERROR_METHOD = 'activation_error_method'
WEIGHTS_BIAS_CORRECTION = 'weights_bias_correction'
Z_THRESHOLD = 'z_threshold'
LINEAR_COLLAPSING = 'linear_collapsing'
RESIDUAL_COLLAPSING = 'residual_collapsing'

# MixedPrecisionQuantizationConfig parameters
DISTANCE_WEIGHTING_METHOD = 'distance_weighting_method'
NUM_OF_IMAGES = 'num_of_images'
USE_HESSIAN_BASED_SCORES = 'use_hessian_based_scores'

# ResourceUtilization parameters
WEIGHTS_COMPRESSION_RATIO = 'weights_compression_ratio'

# Resource utilization data parameters
Expand All @@ -32,14 +42,6 @@
TARGET_RESOURCE_UTILIZATION = 'target_resource_utilization'
IN_MODULE = 'in_module'

# QuantizationConfig parameters
ACTIVATION_ERROR_METHOD = 'activation_error_method'
WEIGHTS_ERROR_METHOD = 'weights_error_method'
WEIGHTS_BIAS_CORRECTION = 'weights_bias_correction'
Z_THRESHOLD = 'z_threshold'
LINEAR_COLLAPSING = 'linear_collapsing'
RESIDUAL_COLLAPSING = 'residual_collapsing'

# GPTQ specific parameters
GPTQ_CONFIG = 'gptq_config'
MODEL = 'model'
Expand All @@ -48,7 +50,12 @@
N_EPOCHS = 'n_epochs'
OPTIMIZER = 'optimizer'

# Export parameters
# low_bit_quantizer_ptq
CONVERTER_VER = 'converter_ver'
LEARNING_RATE = 'learning_rate'

# Export parameters
SAVE_MODEL_PATH = 'save_model_path'

# default compression ratio
DEFAULT_COMPRESSION_RATIO = 0.75
134 changes: 92 additions & 42 deletions model_compression_toolkit/wrapper/mct_wrapper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,13 +17,13 @@
import model_compression_toolkit as mct
from model_compression_toolkit.logger import Logger
from model_compression_toolkit.wrapper.constants import (
REPRESENTATIVE_DATA_GEN, CORE_CONFIG, FW_NAME, SDSP_VERSION,
NUM_OF_IMAGES, USE_HESSIAN_BASED_SCORES, IN_MODEL, IN_MODULE, MODEL,
TARGET_PLATFORM_CAPABILITIES, TARGET_RESOURCE_UTILIZATION,
ACTIVATION_ERROR_METHOD, WEIGHTS_ERROR_METHOD, WEIGHTS_BIAS_CORRECTION,
Z_THRESHOLD, LINEAR_COLLAPSING, RESIDUAL_COLLAPSING, GPTQ_CONFIG,
WEIGHTS_COMPRESSION_RATIO, N_EPOCHS, OPTIMIZER, LEARNING_RATE,
CONVERTER_VER, SAVE_MODEL_PATH
FW_NAME, SDSP_VERSION, ACTIVATION_ERROR_METHOD, WEIGHTS_BIAS_CORRECTION,
Z_THRESHOLD, LINEAR_COLLAPSING, RESIDUAL_COLLAPSING,
DISTANCE_WEIGHTING_METHOD, NUM_OF_IMAGES,
USE_HESSIAN_BASED_SCORES, WEIGHTS_COMPRESSION_RATIO,
IN_MODEL, REPRESENTATIVE_DATA_GEN, CORE_CONFIG, TARGET_PLATFORM_CAPABILITIES,
TARGET_RESOURCE_UTILIZATION, IN_MODULE, GPTQ_CONFIG, MODEL,
N_EPOCHS, OPTIMIZER, LEARNING_RATE, CONVERTER_VER, SAVE_MODEL_PATH, DEFAULT_COMPRESSION_RATIO
)


Expand Down Expand Up @@ -55,11 +55,11 @@ def __init__(self):
:widths: 30, 30, 40

"sdsp_version", "'3.14'", "SDSP version for TPC"
"activation_error_method", "mct.core.QuantizationErrorMethod.MSE", "Activation quantization error method"
"weights_bias_correction", "True", "Enable weights bias correction"
"z_threshold", "float('inf')", "Z-threshold for quantization"
"linear_collapsing", "True", "Enable linear layer collapsing"
"residual_collapsing", "True", "Enable residual connection collapsing"
"activation_error_method", "mct.core.QuantizationErrorMethod.MSE", "Activation quantization error method (low priority)"
"weights_bias_correction", "True", "Enable weights bias correction (low priority)"
"z_threshold", "float('inf')", "Z-threshold for quantization (low priority)"
"linear_collapsing", "True", "Enable linear layer collapsing (low priority)"
"residual_collapsing", "True", "Enable residual connection collapsing (low priority)"
"save_model_path", "'./qmodel.keras' / './qmodel.onnx'", "Path to save quantized model (Keras/Pytorch)"

**PTQ, mixed_precision**
Expand All @@ -69,9 +69,15 @@ def __init__(self):
:widths: 30, 30, 40

"sdsp_version", "'3.14'", "SDSP version for TPC"
"activation_error_method", "mct.core.QuantizationErrorMethod.MSE", "Activation quantization error method (low priority)"
"weights_bias_correction", "True", "Enable weights bias correction (low priority)"
"z_threshold", "float('inf')", "Z-threshold for quantization (low priority)"
"linear_collapsing", "True", "Enable linear layer collapsing (low priority)"
"residual_collapsing", "True", "Enable residual connection collapsing (low priority)"
"distance_weighting_method", "See `MixedPrecisionQuantizationConfig <https://sonysemiconductorsolutions.github.io/mct-model-optimization/api/api_docs/classes/MixedPrecisionQuantizationConfig.html>`_", "Distance weighting method for mixed precision (low priority)"
"num_of_images", "5", "Number of images for mixed precision"
"use_hessian_based_scores", "False", "Use Hessian-based scores for mixed precision"
"weights_compression_ratio", "None", "Weights compression ratio for resource util"
"use_hessian_based_scores", "False", "Use Hessian-based scores for mixed precision (low priority)"
"weights_compression_ratio", "0.75", "Weights compression ratio for resource util (0.0～1.0)"
"save_model_path", "'./qmodel.keras' / './qmodel.onnx'", "Path to save quantized model (Keras/Pytorch)"

**GPTQ**
Expand All @@ -81,8 +87,13 @@ def __init__(self):
:widths: 30, 30, 40

"sdsp_version", "'3.14'", "SDSP version for TPC"
"activation_error_method", "mct.core.QuantizationErrorMethod.MSE", "Activation quantization error method (low priority)"
"weights_bias_correction", "True", "Enable weights bias correction (low priority)"
"z_threshold", "float('inf')", "Z-threshold for quantization (low priority)"
"linear_collapsing", "True", "Enable linear layer collapsing (low priority)"
"residual_collapsing", "True", "Enable residual connection collapsing (low priority)"
"n_epochs", "5", "Number of training epochs for GPTQ"
"optimizer", "None", "Optimizer for GPTQ training"
"optimizer", "default of `get_keras_gptq_config <https://sonysemiconductorsolutions.github.io/mct-model-optimization/api/api_docs/methods/get_keras_gptq_config.html#model_compression_toolkit.gptq.get_keras_gptq_config>`_ or `get_pytorch_gptq_config <https://sonysemiconductorsolutions.github.io/mct-model-optimization/api/api_docs/methods/get_pytroch_gptq_config.html#model_compression_toolkit.gptq.get_pytorch_gptq_config>`_", "Optimizer for GPTQ training (low priority)"
"save_model_path", "'./qmodel.keras' / './qmodel.onnx'", "Path to save quantized model (Keras/Pytorch)"

**GPTQ, mixed_precision**
Expand All @@ -92,11 +103,17 @@ def __init__(self):
:widths: 30, 30, 40

"sdsp_version", "'3.14'", "SDSP version for TPC"
"activation_error_method", "mct.core.QuantizationErrorMethod.MSE", "Activation quantization error method (low priority)"
"weights_bias_correction", "True", "Enable weights bias correction (low priority)"
"z_threshold", "float('inf')", "Z-threshold for quantization (low priority)"
"linear_collapsing", "True", "Enable linear layer collapsing (low priority)"
"residual_collapsing", "True", "Enable residual connection collapsing (low priority)"
"weights_compression_ratio", "0.75", "Weights compression ratio for resource util (0.0～1.0)"
"n_epochs", "5", "Number of training epochs for GPTQ"
"optimizer", "None", "Optimizer for GPTQ training"
"optimizer", "default of `get_keras_gptq_config <https://sonysemiconductorsolutions.github.io/mct-model-optimization/api/api_docs/methods/get_keras_gptq_config.html#model_compression_toolkit.gptq.get_keras_gptq_config>`_ or `get_pytorch_gptq_config <https://sonysemiconductorsolutions.github.io/mct-model-optimization/api/api_docs/methods/get_pytroch_gptq_config.html#model_compression_toolkit.gptq.get_pytorch_gptq_config>`_", "Optimizer for GPTQ training (low priority)"
"distance_weighting_method", "See `MixedPrecisionQuantizationConfig <https://sonysemiconductorsolutions.github.io/mct-model-optimization/api/api_docs/classes/MixedPrecisionQuantizationConfig.html>`_", "Distance weighting method for mixed precision (low priority)"
"num_of_images", "5", "Number of images for mixed precision"
"use_hessian_based_scores", "False", "Use Hessian-based scores for mixed precision"
"weights_compression_ratio", "None", "Weights compression ratio for resource util"
"use_hessian_based_scores", "False", "Use Hessian-based scores for mixed precision (low priority)"
"save_model_path", "'./qmodel.keras' / './qmodel.onnx'", "Path to save quantized model (Keras/Pytorch)"

"""
Expand All @@ -112,16 +129,17 @@ def __init__(self):
LINEAR_COLLAPSING: True,
RESIDUAL_COLLAPSING: True,

# GradientPTQConfig
N_EPOCHS: 5,
OPTIMIZER: None,

# MixedPrecisionQuantizationConfig
DISTANCE_WEIGHTING_METHOD: None,
NUM_OF_IMAGES: 5,
USE_HESSIAN_BASED_SCORES: False,

# ResourceUtilization
WEIGHTS_COMPRESSION_RATIO: None,
WEIGHTS_COMPRESSION_RATIO: DEFAULT_COMPRESSION_RATIO,

# GradientPTQConfig
N_EPOCHS: 5,
OPTIMIZER: None,

# low_bit_quantizer_ptq
LEARNING_RATE: 0.001,
Expand Down Expand Up @@ -172,16 +190,21 @@ def _initialize_and_validate(self, float_model: Any,
Z_THRESHOLD, LINEAR_COLLAPSING, RESIDUAL_COLLAPSING,
SAVE_MODEL_PATH]
else:
allowed_keys = [FW_NAME, SDSP_VERSION, NUM_OF_IMAGES, USE_HESSIAN_BASED_SCORES,
allowed_keys = [FW_NAME, SDSP_VERSION, ACTIVATION_ERROR_METHOD, WEIGHTS_BIAS_CORRECTION,
Z_THRESHOLD, LINEAR_COLLAPSING, RESIDUAL_COLLAPSING,
DISTANCE_WEIGHTING_METHOD, NUM_OF_IMAGES, USE_HESSIAN_BASED_SCORES,
WEIGHTS_COMPRESSION_RATIO, SAVE_MODEL_PATH]
else:
if not use_mixed_precision:
allowed_keys = [FW_NAME, SDSP_VERSION, N_EPOCHS, OPTIMIZER,
SAVE_MODEL_PATH]
allowed_keys = [FW_NAME, SDSP_VERSION, ACTIVATION_ERROR_METHOD, WEIGHTS_BIAS_CORRECTION,
Z_THRESHOLD, LINEAR_COLLAPSING, RESIDUAL_COLLAPSING,
N_EPOCHS, OPTIMIZER, SAVE_MODEL_PATH]
else:
allowed_keys = [FW_NAME, SDSP_VERSION, N_EPOCHS, OPTIMIZER,
allowed_keys = [FW_NAME, SDSP_VERSION, ACTIVATION_ERROR_METHOD, WEIGHTS_BIAS_CORRECTION,
Z_THRESHOLD, LINEAR_COLLAPSING, RESIDUAL_COLLAPSING,
WEIGHTS_COMPRESSION_RATIO, N_EPOCHS, OPTIMIZER, DISTANCE_WEIGHTING_METHOD,
NUM_OF_IMAGES, USE_HESSIAN_BASED_SCORES,
WEIGHTS_COMPRESSION_RATIO, SAVE_MODEL_PATH]
SAVE_MODEL_PATH]

self.params = { k: v for k, v in self.params.items() if k in allowed_keys }

Expand Down Expand Up @@ -320,22 +343,33 @@ def _setting_PTQ_mixed_precision(self) -> Dict[str, Any]:
Returns:
dict: Parameter dictionary for PTQ.
"""
params_QCfg = {
ACTIVATION_ERROR_METHOD: self.params[ACTIVATION_ERROR_METHOD],
WEIGHTS_BIAS_CORRECTION: self.params[WEIGHTS_BIAS_CORRECTION],
Z_THRESHOLD: self.params[Z_THRESHOLD],
LINEAR_COLLAPSING: self.params[LINEAR_COLLAPSING],
RESIDUAL_COLLAPSING: self.params[RESIDUAL_COLLAPSING]
}
q_config = mct.core.QuantizationConfig(**params_QCfg)

params_MPCfg = {
DISTANCE_WEIGHTING_METHOD: self.params[DISTANCE_WEIGHTING_METHOD],
NUM_OF_IMAGES: self.params[NUM_OF_IMAGES],
USE_HESSIAN_BASED_SCORES: self.params[USE_HESSIAN_BASED_SCORES]
}
mixed_precision_config = mct.core.MixedPrecisionQuantizationConfig(**params_MPCfg)
core_config = mct.core.CoreConfig(mixed_precision_config=mixed_precision_config)

core_config = mct.core.CoreConfig(quantization_config=q_config,
mixed_precision_config=mixed_precision_config)

params_RUDCfg = {
IN_MODEL: self.float_model,
REPRESENTATIVE_DATA_GEN: self.representative_dataset,
CORE_CONFIG: core_config,
TARGET_PLATFORM_CAPABILITIES: self.tpc
}
ru_data = self.resource_utilization_data(**params_RUDCfg)
weights_compression_ratio = (
0.75 if self.params[WEIGHTS_COMPRESSION_RATIO] is None
else self.params[WEIGHTS_COMPRESSION_RATIO])
weights_compression_ratio = self.params[WEIGHTS_COMPRESSION_RATIO]
resource_utilization = mct.core.ResourceUtilization(
ru_data.weights_memory * weights_compression_ratio)

Expand All @@ -357,7 +391,6 @@ def _setting_PTQ(self) -> Dict[str, Any]:
"""
params_QCfg = {
ACTIVATION_ERROR_METHOD: self.params[ACTIVATION_ERROR_METHOD],
WEIGHTS_ERROR_METHOD: mct.core.QuantizationErrorMethod.MSE,
WEIGHTS_BIAS_CORRECTION: self.params[WEIGHTS_BIAS_CORRECTION],
Z_THRESHOLD: self.params[Z_THRESHOLD],
LINEAR_COLLAPSING: self.params[LINEAR_COLLAPSING],
Expand All @@ -383,36 +416,42 @@ def _setting_GPTQ_mixed_precision(self) -> Dict[str, Any]:
Returns:
dict: Parameter dictionary for GPTQ.
"""
params_QCfg = {
ACTIVATION_ERROR_METHOD: self.params[ACTIVATION_ERROR_METHOD],
WEIGHTS_BIAS_CORRECTION: self.params[WEIGHTS_BIAS_CORRECTION],
Z_THRESHOLD: self.params[Z_THRESHOLD],
LINEAR_COLLAPSING: self.params[LINEAR_COLLAPSING],
RESIDUAL_COLLAPSING: self.params[RESIDUAL_COLLAPSING]
}
q_config = mct.core.QuantizationConfig(**params_QCfg)

params_GPTQCfg = {
N_EPOCHS: self.params[N_EPOCHS],
OPTIMIZER: self.params[OPTIMIZER]
}
gptq_config = self.get_gptq_config(**params_GPTQCfg)

params_MPCfg = {
DISTANCE_WEIGHTING_METHOD: self.params[DISTANCE_WEIGHTING_METHOD],
NUM_OF_IMAGES: self.params[NUM_OF_IMAGES],
USE_HESSIAN_BASED_SCORES: self.params[USE_HESSIAN_BASED_SCORES],
}
mixed_precision_config = mct.core.MixedPrecisionQuantizationConfig(**params_MPCfg)
core_config = mct.core.CoreConfig(mixed_precision_config=mixed_precision_config)

core_config = mct.core.CoreConfig(quantization_config=q_config,
mixed_precision_config=mixed_precision_config)

params_RUDCfg = {
IN_MODEL: self.float_model,
REPRESENTATIVE_DATA_GEN: self.representative_dataset,
CORE_CONFIG: core_config,
TARGET_PLATFORM_CAPABILITIES: self.tpc
}
ru_data = self.resource_utilization_data(**params_RUDCfg)
weights_compression_ratio = (
0.75 if self.params[WEIGHTS_COMPRESSION_RATIO] is None
else self.params[WEIGHTS_COMPRESSION_RATIO])
weights_compression_ratio = self.params[WEIGHTS_COMPRESSION_RATIO]
resource_utilization = mct.core.ResourceUtilization(
ru_data.weights_memory * weights_compression_ratio)

core_config = mct.core.CoreConfig(
mixed_precision_config = mixed_precision_config,
quantization_config = mct.core.QuantizationConfig()
)

params_GPTQ = {
self.argname_model: self.float_model,
REPRESENTATIVE_DATA_GEN: self.representative_dataset,
Expand All @@ -430,6 +469,16 @@ def _setting_GPTQ(self) -> Dict[str, Any]:
Returns:
dict: Parameter dictionary for GPTQ.
"""
params_QCfg = {
ACTIVATION_ERROR_METHOD: self.params[ACTIVATION_ERROR_METHOD],
WEIGHTS_BIAS_CORRECTION: self.params[WEIGHTS_BIAS_CORRECTION],
Z_THRESHOLD: self.params[Z_THRESHOLD],
LINEAR_COLLAPSING: self.params[LINEAR_COLLAPSING],
RESIDUAL_COLLAPSING: self.params[RESIDUAL_COLLAPSING]
}
q_config = mct.core.QuantizationConfig(**params_QCfg)
core_config = mct.core.CoreConfig(quantization_config=q_config)

params_GPTQCfg = {
N_EPOCHS: self.params[N_EPOCHS],
OPTIMIZER: self.params[OPTIMIZER]
Expand All @@ -440,6 +489,7 @@ def _setting_GPTQ(self) -> Dict[str, Any]:
self.argname_model: self.float_model,
REPRESENTATIVE_DATA_GEN: self.representative_dataset,
GPTQ_CONFIG: gptq_config,
CORE_CONFIG: core_config,
TARGET_PLATFORM_CAPABILITIES: self.tpc
}
return params_GPTQ
Expand Down
Loading