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

Jump to bottom

Refactor of ptq_common #649

Merged
merged 3 commits into from
Jul 17, 2023
Merged

Refactor of ptq_common #649

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
61a630a
Refactor of ptq_common
Giuseppe5 Jul 2, 2023
685b9e1
Review
Giuseppe5 Jul 13, 2023
f7064d1
Fix out proj input bit width
Giuseppe5 Jul 17, 2023
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
8 changes: 4 additions & 4 deletions src/brevitas_examples/imagenet_classification/ptq/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@ We provide two workflows:
- A benchmark suite that tests several quantization configurations on a few selected models.

Three types of target backend are exposed for programmatic quantization. Different backends dictate different structural policies for how a network should be quantized:
- *generic*:
- *fx*:
- The number of re-quantization ops is minimized by re-quantizing only when necessary, avoiding consecutive quantization ops if possible.
- Adds are quantized to have the same scale at the input, but allows for different signs.
- Concats are quantized to have the same scale, zero-point, sign and bit-width.
Expand Down Expand Up @@ -67,7 +67,7 @@ usage: ptq_evaluate.py [-h] --calibration-dir CALIBRATION_DIR --validation-dir
[--export-dir EXPORT_DIR] [--gpu GPU]
[--calibration-samples CALIBRATION_SAMPLES]
[--model-name ARCH]
[--target-backend {generic,layerwise,flexml}]
[--target-backend {fx,layerwise,flexml}]
[--scale-factor-type {float32,po2}]
[--act-bit-width ACT_BIT_WIDTH]
[--weight-bit-width WEIGHT_BIT_WIDTH]
Expand Down Expand Up @@ -134,8 +134,8 @@ optional arguments:
vgg16 | vgg16_bn | vgg19 | vgg19_bn | vit_b_16 |
vit_b_32 | vit_h_14 | vit_l_16 | vit_l_32 |
wide_resnet101_2 | wide_resnet50_2 (default: resnet18)
--target-backend {generic,layerwise,flexml}
Backend to target for quantization (default: generic)
--target-backend {fx,layerwise,flexml}
Backend to target for quantization (default: fx)
--scale-factor-type {float32,po2}
Type for scale factors (default: float32)
--act-bit-width ACT_BIT_WIDTH
Expand Down
2 changes: 1 addition & 1 deletion src/brevitas_examples/imagenet_classification/ptq/benchmark/post_processing.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,7 +27,7 @@ def main():
'Weight bit width',
'Act bit width',
'Bias bit width',
'Scaling per output channel',
'Weight quant granularity',
'Act quant type'])
idx = grouped_df['Top 1% quant accuracy'].transform(max) == df['Top 1% quant accuracy']
best_config_df = df[idx]
Expand Down
104 changes: 65 additions & 39 deletions src/brevitas_examples/imagenet_classification/ptq/benchmark/ptq_benchmark_torchvision.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,13 +45,14 @@

OPTIONS = {
'model_name': TORCHVISION_TOP1_MAP.keys(),
'target_backend': ['generic', 'layerwise', 'flexml'], # Target backend
'scale_factor_type': ['float32', 'po2'], # Scale factor type
'weight_bit_width': [8, 6, 4], # Weight Bit Width
'act_bit_width': [8], # Act bit width
'bias_bit_width': ['int32', 'int16'], # Bias Bit-Width for Po2 scale
'scaling_per_output_channel': [False, True], # Scaling Per Output Channel
'act_quant_type': ['asymmetric', 'symmetric'], # Act Quant Type
'target_backend': ['fx', 'layerwise', 'flexml'], # Target backend
'scale_factor_type': ['float', 'po2'], # Scale factor type
'weight_bit_width': [8, 4], # Weight Bit Width
'act_bit_width': [8, 4], # Act bit width
'bias_bit_width': [32, 16], # Bias Bit-Width for Po2 scale
'weight_quant_granularity': ['per_tensor', 'per_channel'], # Scaling Per Output Channel
'act_quant_type': ['asym', 'sym'], # Act Quant Type
'act_param_method': ['stats', 'mse'], # Act Param Method
'bias_corr': [True], # Bias Correction
'graph_eq_iterations': [0, 20], # Graph Equalization
'graph_eq_merge_bias': [False, True], # Merge bias for Graph Equalization
Expand All @@ -63,17 +64,18 @@
}

OPTIONS_DEFAULT = {
'target_backend': ['generic'], # Target backend
'scale_factor_type': ['float32'], # Scale factor type
'target_backend': ['fx'], # Target backend
'scale_factor_type': ['float'], # Scale factor type
'weight_bit_width': [8], # Weight Bit Width
'act_bit_width': [8], # Act bit width
'bias_bit_width': ['int32'], # Bias Bit-Width for Po2 scale
'scaling_per_output_channel': [True], # Scaling Per Output Channel
'act_quant_type': ['symmetric'], # Act Quant Type
'bias_bit_width': [32], # Bias Bit-Width for Po2 scale
'weight_quant_granularity': ['per_channel'], # Scaling Per Output Channel
'act_quant_type': ['sym'], # Act Quant Type
'act_param_method': ['stats'], # Act Param Method
'bias_corr': [True], # Bias Correction
'graph_eq_iterations': [20], # Graph Equalization
'graph_eq_merge_bias': [True], # Merge bias for Graph Equalization
'act_equalization': ['fx'], # Perform Activation Equalization (Smoothquant)
'act_equalization': [None], # Perform Activation Equalization (Smoothquant)
'learned_round': [False], # Enable/Disable Learned Round
'gptq': [True], # Enable/Disable GPTQ
'gptq_act_order': [False], # Use act_order euristics for GPTQ
Expand Down Expand Up @@ -131,32 +133,29 @@ def main():


def ptq_torchvision_models(df, args):

combinations = list(product(*OPTIONS.values()))
if args.idx > len(combinations):
# Generate all possible combinations, including invalid ones
# Split stats and mse due to the act_quant_percentile value
percentile_options = OPTIONS.copy()
percentile_options['act_param_method'] = ['stats']
mse_options = OPTIONS.copy()
mse_options['act_param_method'] = ['mse']
mse_options['act_quant_percentile'] = [None]
# Combine the two sets of combinations
combinations = list(product(*percentile_options.values())) + list(
product(*mse_options.values()))
# Generate Namespace for each configuration
configs = [
SimpleNamespace(**{k: v
for k, v in zip(OPTIONS.keys(), combination)})
for combination in combinations]
# Define which configurations are not valid
configs = list(map(validate_config, configs))
# Drop invalid configurations
configs = list(config for config in configs if config.is_valid)
if args.idx > len(configs):
return
combination = combinations[args.idx]

config_namespace = SimpleNamespace()
for key, value in zip(OPTIONS.keys(), combination):
setattr(config_namespace, key, value)

# Flexml supports only per-tensor scale factors, power of two scale factors
if config_namespace.target_backend == 'flexml' and (
config_namespace.scaling_per_output_channel or
config_namespace.scale_factor_type == 'float32'):
return
# Merge bias can be enabled only when graph equalization is enabled
if config_namespace.graph_eq_iterations == 0 and config_namespace.graph_eq_merge_bias:
return
# For generic and layerwise backend, we only test for int32 bias bit width
if (config_namespace.target_backend == 'generic' or config_namespace.target_backend
== 'layerwise') and config_namespace.bias_bit_width == 'int16':
return

# If GPTQ is disabled, we do not care about the act_order heuristic
if not config_namespace.gptq and config_namespace.gptq_act_order:
return
config_namespace = configs[args.idx]

fp_accuracy = TORCHVISION_TOP1_MAP[config_namespace.model_name]
# Get model-specific configurations about input shapes and normalization
Expand Down Expand Up @@ -194,7 +193,7 @@ def ptq_torchvision_models(df, args):
torch.ones(1, 3, img_shape, img_shape),
equalize_iters=config_namespace.graph_eq_iterations,
equalize_merge_bias=config_namespace.graph_eq_merge_bias)
elif config_namespace.target_backend == 'generic' or config_namespace.target_backend == 'layerwise':
elif config_namespace.target_backend == 'fx' or config_namespace.target_backend == 'layerwise':
model = preprocess_for_quantize(
model,
equalize_iters=config_namespace.graph_eq_iterations,
Expand All @@ -214,7 +213,7 @@ def ptq_torchvision_models(df, args):
act_bit_width=config_namespace.act_bit_width,
weight_bit_width=config_namespace.weight_bit_width,
bias_bit_width=config_namespace.bias_bit_width,
scaling_per_output_channel=config_namespace.scaling_per_output_channel,
weight_quant_granularity=config_namespace.weight_quant_granularity,
act_quant_percentile=config_namespace.act_quant_percentile,
act_quant_type=config_namespace.act_quant_type,
scale_factor_type=config_namespace.scale_factor_type)
Expand Down Expand Up @@ -275,5 +274,32 @@ def ptq_torchvision_models(df, args):
torchvision_df.to_csv(os.path.join(folder, 'RESULTS_TORCHVISION.csv'), index=False)


def validate_config(config_namespace):
is_valid = True
# Flexml supports only per-tensor scale factors, power of two scale factors
if config_namespace.target_backend == 'flexml' and (
config_namespace.weight_quant_granularity == 'per_channel' or
config_namespace.scale_factor_type == 'float32'):
is_valid = False
# Merge bias can be enabled only when graph equalization is enabled
if config_namespace.graph_eq_iterations == 0 and config_namespace.graph_eq_merge_bias:
is_valid = False
# For fx and layerwise backend, we only test for bias with bit width equals to 32
if (config_namespace.target_backend == 'fx' or config_namespace.target_backend
== 'layerwise') and config_namespace.bias_bit_width == 16:
is_valid = False
# If GPTQ is disabled, we do not care about the act_order heuristic
if not config_namespace.gptq and config_namespace.gptq_act_order:
is_valid = False

if config_namespace.act_equalization == 'layerwise' and config_namespace.target_backend == 'fx':
is_valid = False
if config_namespace.act_bit_width < config_namespace.weight_bit_width:
is_valid = False

config_namespace.is_valid = is_valid
return config_namespace


if __name__ == '__main__':
main()
Loading