Skip to content
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

Fix target KPI usage in tutorials #992

Merged
merged 1 commit into from
Mar 12, 2024
Merged

Fix target KPI usage in tutorials #992

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 11 additions & 0 deletions model_compression_toolkit/core/common/mixed_precision/mixed_precision_quantization_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,3 +66,14 @@ def __init__(self,
self.norm_scores = norm_scores

self.metric_normalization_threshold = metric_normalization_threshold

def set_target_kpi(self, target_kpi: KPI):
"""
Setting target KPI in mixed precision config.

Args:
target_kpi: A target KPI to set.

"""

self.target_kpi = target_kpi
4 changes: 2 additions & 2 deletions tutorials/notebooks/keras/gptq/example_keras_mobilenet_gptq_mixed_precision.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -137,6 +137,7 @@ def representative_data_gen() -> list:
# examples:
# weights_compression_ratio = 0.75 - About 0.75 of the model's weights memory size when quantized with 8 bits.
kpi = mct.core.KPI(kpi_data.weights_memory * args.weights_compression_ratio)
config.mixed_precision_config.set_target_kpi(kpi)

# Create a GPTQ quantization configuration and set the number of training iterations.
gptq_config = mct.gptq.get_keras_gptq_config(n_epochs=args.num_gptq_training_iterations,
Expand All @@ -146,8 +147,7 @@ def representative_data_gen() -> list:
representative_data_gen,
gptq_config=gptq_config,
core_config=config,
target_platform_capabilities=target_platform_cap,
target_kpi=kpi)
target_platform_capabilities=target_platform_cap)

# Export quantized model to TFLite and Keras.
# For more details please see: https://github.com/sony/model_optimization/blob/main/model_compression_toolkit/exporter/README.md
Expand Down
4 changes: 2 additions & 2 deletions tutorials/notebooks/keras/ptq/example_keras_effdet_lite0.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -387,12 +387,12 @@
"# set weights memory size, so the quantized model will fit the IMX500 memory\n",
"kpi = mct.core.KPI(weights_memory=2674291)\n",
"# set MixedPrecision configuration for compressing the weights\n",
"mp_config = mct.core.MixedPrecisionQuantizationConfig(use_hessian_based_scores=False)\n",
"mp_config = mct.core.MixedPrecisionQuantizationConfig(use_hessian_based_scores=False,\n",
" target_kpi=kpi)\n",
"core_config = mct.core.CoreConfig(mixed_precision_config=mp_config)\n",
"quant_model, _ = mct.ptq.keras_post_training_quantization(\n",
" model,\n",
" get_representative_dataset(20),\n",
" target_kpi=kpi,\n",
" core_config=core_config,\n",
" target_platform_capabilities=tpc)"
],
Expand Down
4 changes: 2 additions & 2 deletions tutorials/notebooks/keras/ptq/example_keras_mobilenet_mixed_precision.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -266,7 +266,8 @@
"# while the bias will not)\n",
"# examples:\n",
"weights_compression_ratio = 0.75 # About 0.75 of the model's weights memory size when quantized with 8 bits.\n",
"kpi = mct.core.KPI(kpi_data.weights_memory * weights_compression_ratio)"
"kpi = mct.core.KPI(kpi_data.weights_memory * weights_compression_ratio)\n",
"core_config.mixed_precision_config.set_target_kpi(kpi)"
],
"metadata": {
"collapsed": false
Expand Down Expand Up @@ -296,7 +297,6 @@
"quantized_model, quantization_info = mct.ptq.keras_post_training_quantization(\n",
" float_model,\n",
" representative_dataset_gen,\n",
" target_kpi=kpi,\n",
" core_config=core_config,\n",
" target_platform_capabilities=tpc)"
]
Expand Down
2 changes: 1 addition & 1 deletion tutorials/notebooks/keras/ptq/example_keras_mobilenet_mixed_precision.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -130,12 +130,12 @@ def representative_data_gen() -> list:
# examples:
# weights_compression_ratio = 0.75 - About 0.75 of the model's weights memory size when quantized with 8 bits.
kpi = mct.core.KPI(kpi_data.weights_memory * args.weights_compression_ratio)
configuration.mixed_precision_config.set_target_kpi(kpi)

# It is also possible to constraint only part of the KPI metric, e.g., by providing only weights_memory target
# in the past KPI object, e.g., kpi = mct.core.KPI(kpi_data.weights_memory * 0.75)
quantized_model, quantization_info = mct.ptq.keras_post_training_quantization(model,
representative_data_gen,
target_kpi=kpi,
core_config=configuration,
target_platform_capabilities=target_platform_cap)

Expand Down
2 changes: 1 addition & 1 deletion tutorials/notebooks/keras/ptq/example_keras_mobilenet_mixed_precision_lut.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -136,10 +136,10 @@ def representative_data_gen() -> list:
# weights_compression_ratio = 0.4 - About 0.4 of the model's weights memory size when quantized with 8 bits.
kpi = mct.core.KPI(kpi_data.weights_memory * args.weights_compression_ratio)
# Note that in this example, activations are quantized with fixed bit-width (non mixed-precision) of 8-bit.
configuration.mixed_precision_config.set_target_kpi(kpi)

quantized_model, quantization_info = mct.ptq.keras_post_training_quantization(model,
representative_data_gen,
target_kpi=kpi,
core_config=configuration,
target_platform_capabilities=target_platform_cap)

Expand Down
2 changes: 1 addition & 1 deletion tutorials/notebooks/keras/ptq/example_keras_yolov8n.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -297,11 +297,11 @@
" config,\n",
" target_platform_capabilities=tpc)\n",
"kpi = mct.core.KPI(kpi_data.weights_memory * 0.75)\n",
"config.mixed_precision_config.set_target_kpi(kpi)\n",
"\n",
"# Perform post training quantization\n",
"quant_model, _ = mct.ptq.keras_post_training_quantization(model,\n",
" representative_dataset_gen,\n",
" target_kpi=kpi,\n",
" core_config=config,\n",
" target_platform_capabilities=tpc)\n",
"print('Quantized model is ready')"
Expand Down
2 changes: 1 addition & 1 deletion tutorials/notebooks/keras/ptq/keras_yolov8n_for_imx500.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -193,11 +193,11 @@
" config,\n",
" target_platform_capabilities=tpc)\n",
"kpi = mct.core.KPI(kpi_data.weights_memory * 0.75)\n",
"config.mixed_precision_config.set_target_kpi(kpi)\n",
"\n",
"# Perform post training quantization\n",
"quant_model, _ = mct.ptq.keras_post_training_quantization(model,\n",
" representative_dataset_gen,\n",
" target_kpi=kpi,\n",
" core_config=config,\n",
" target_platform_capabilities=tpc)\n",
"print('Quantized model is ready')"
Expand Down
2 changes: 1 addition & 1 deletion tutorials/notebooks/pytorch/ptq/example_pytorch_mobilenet_mixed_precision.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,13 +123,13 @@ def representative_data_gen() -> list:
# examples:
# weights_compression_ratio = 0.75 - About 0.75 of the model's weights memory size when quantized with 8 bits.
kpi = mct.core.KPI(kpi_data.weights_memory * args.weights_compression_ratio)
configuration.mixed_precision_config.set_target_kpi(kpi)

# It is also possible to constraint only part of the KPI metric, e.g., by providing only weights_memory target
# in the past KPI object, e.g., kpi = mct.core.KPI(kpi_data.weights_memory * 0.75)

quantized_model, quantization_info = mct.ptq.pytorch_post_training_quantization(model,
representative_data_gen,
target_kpi=kpi,
core_config=configuration,
target_platform_capabilities=target_platform_cap)

Expand Down
2 changes: 1 addition & 1 deletion tutorials/notebooks/pytorch/ptq/example_pytorch_mobilenet_mixed_precision_lut.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -128,10 +128,10 @@ def representative_data_gen() -> list:
# weights_compression_ratio = 0.4 - About 0.4 of the model's weights memory size when quantized with 8 bits.
kpi = mct.core.KPI(kpi_data.weights_memory * args.weights_compression_ratio)
# Note that in this example, activations are quantized with fixed bit-width (non mixed-precision) of 8-bit.
configuration.mixed_precision_config.set_target_kpi(kpi)

quantized_model, quantization_info = mct.ptq.pytorch_post_training_quantization(model,
representative_data_gen,
target_kpi=kpi,
core_config=configuration,
target_platform_capabilities=target_platform_cap)

4 changes: 2 additions & 2 deletions tutorials/notebooks/pytorch/ptq/example_pytorch_mobilenetv2_cifar100_mixed_precision.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -515,7 +515,8 @@
"# while the bias will not)\n",
"# examples:\n",
"# weights_compression_ratio = 0.75 - About 0.75 of the model's weights memory size when quantized with 8 bits.\n",
"kpi = mct.core.KPI(kpi_data.weights_memory * 0.75)"
"kpi = mct.core.KPI(kpi_data.weights_memory * 0.75)\n",
"configuration.mixed_precision_config.set_target_kpi(kpi)"
]
},
{
Expand All @@ -537,7 +538,6 @@
"source": [
"quantized_model, quantization_info = mct.ptq.pytorch_post_training_quantization(model,\n",
" representative_data_gen,\n",
" target_kpi=kpi,\n",
" core_config=configuration,\n",
" target_platform_capabilities=target_platform_cap)\n",
" "
Expand Down
2 changes: 1 addition & 1 deletion tutorials/notebooks/pytorch/ptq/example_pytorch_mobilenetv2_cifar100_mixed_precision.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -241,12 +241,12 @@ def representative_data_gen() -> list:
# examples:
# weights_compression_ratio = 0.75 - About 0.75 of the model's weights memory size when quantized with 8 bits.
kpi = mct.core.KPI(kpi_data.weights_memory * args.weights_compression_ratio)
configuration.mixed_precision_config.set_target_kpi(kpi)

# It is also possible to constraint only part of the KPI metric, e.g., by providing only weights_memory target
# in the past KPI object, e.g., kpi = mct.core.KPI(kpi_data.weights_memory * 0.75)
quantized_model, quantization_info = mct.ptq.pytorch_post_training_quantization(model,
representative_data_gen,
target_kpi=kpi,
core_config=configuration,
target_platform_capabilities=target_platform_cap)
# Finally, we evaluate the quantized model:
Expand Down
4 changes: 1 addition & 3 deletions tutorials/quick_start/keras_fw/quant.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -100,10 +100,10 @@ def quantize(model: tf.keras.Model,
shift_negative_activation_correction=True),
mixed_precision_config=mp_conf)
target_kpi = get_target_kpi(model, mp_wcr, representative_data_gen, core_conf, tpc)
core_conf.mixed_precision_config.set_target_kpi(target_kpi)
else:
core_conf = CoreConfig(quantization_config=mct.core.QuantizationConfig(
shift_negative_activation_correction=True))
target_kpi = None

# Quantize model
if args.get('gptq', False):
Expand All @@ -118,7 +118,6 @@ def quantize(model: tf.keras.Model,
quantized_model, quantization_info = \
mct.gptq.keras_gradient_post_training_quantization(model,
representative_data_gen=representative_data_gen,
target_kpi=target_kpi,
core_config=core_conf,
gptq_config=gptq_conf,
gptq_representative_data_gen=representative_data_gen,
Expand All @@ -130,7 +129,6 @@ def quantize(model: tf.keras.Model,
quantized_model, quantization_info = \
mct.ptq.keras_post_training_quantization(model,
representative_data_gen=representative_data_gen,
target_kpi=target_kpi,
core_config=core_conf,
target_platform_capabilities=tpc)

Expand Down
4 changes: 1 addition & 3 deletions tutorials/quick_start/pytorch_fw/quant.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,10 +101,10 @@ def quantize(model: nn.Module,
shift_negative_activation_correction=True),
mixed_precision_config=mp_conf)
target_kpi = get_target_kpi(model, mp_wcr, representative_data_gen, core_conf, tpc)
core_conf.mixed_precision_config.set_target_kpi(target_kpi)
else:
core_conf = CoreConfig(quantization_config=mct.core.QuantizationConfig(
shift_negative_activation_correction=True))
target_kpi = None

# Quantize model
if args.get('gptq', False):
Expand All @@ -119,7 +119,6 @@ def quantize(model: nn.Module,
quantized_model, quantization_info = \
mct.gptq.pytorch_gradient_post_training_quantization(model,
representative_data_gen=representative_data_gen,
target_kpi=target_kpi,
core_config=core_conf,
gptq_config=gptq_conf,
gptq_representative_data_gen=representative_data_gen,
Expand All @@ -131,7 +130,6 @@ def quantize(model: nn.Module,
quantized_model, quantization_info = \
mct.ptq.pytorch_post_training_quantization(model,
representative_data_gen=representative_data_gen,
target_kpi=target_kpi,
core_config=core_conf,
target_platform_capabilities=tpc)

Expand Down
Loading