Skip to content

Commit

Permalink
Feat (ptq): support for minifloat in evaluation script (#717)
Browse files Browse the repository at this point in the history
  • Loading branch information
@fabianandresgrob
fabianandresgrob authored Oct 19, 2023
1 parent b051309 commit 1412174
Show file tree
Hide file tree
Showing 2 changed files with 122 additions and 68 deletions.
131 changes: 67 additions & 64 deletions src/brevitas_examples/imagenet_classification/ptq/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -61,40 +61,31 @@ This flow allows to specify which pre-trained torchvision model to quantize and
It also gives the possibility to export the model to either ONNX QCDQ format or in torch QCDQ format.
The quantization and export options to specify are:
```bash
usage: ptq_evaluate.py [-h] --calibration-dir CALIBRATION_DIR --validation-dir
VALIDATION_DIR [--workers WORKERS]
[--batch-size-calibration BATCH_SIZE_CALIBRATION]
[--batch-size-validation BATCH_SIZE_VALIDATION]
[--export-dir EXPORT_DIR] [--gpu GPU]
[--calibration-samples CALIBRATION_SAMPLES]
[--model-name ARCH]
[--target-backend {fx,layerwise,flexml}]
[--scale-factor-type {float32,po2}]
[--act-bit-width ACT_BIT_WIDTH]
[--weight-bit-width WEIGHT_BIT_WIDTH]
[--layerwise-first-last-bit-width LAYERWISE_FIRST_LAST_BIT_WIDTH]
[--bias-bit-width {int32,int16}]
[--act-quant-type {symmetric,asymmetric}]
[--act-equalization {fx,layerwise,None}]
[--act-quant-calibration-type {percentile,mse}]
[--graph-eq-iterations GRAPH_EQ_ITERATIONS]
[--learned-round-iters LEARNED_ROUND_ITERS]
[--learned-round-lr LEARNED_ROUND_LR]
[--act-quant-percentile ACT_QUANT_PERCENTILE]
[--export-onnx-qcdq] [--export-torch-qcdq]
[--scaling-per-output-channel | --no-scaling-per-output-channel]
[--bias-corr | --no-bias-corr]
[--graph-eq-merge-bias | --no-graph-eq-merge-bias]
[--weight-narrow-range | --no-weight-narrow-range]
[--gpfq-p GPFQ_P] [--gptq | --no-gptq]
[--gpfq | --no-gpfq]
[--gptq-act-order | --no-gptq-act-order]
[--learned-round | --no-learned-round]
usage: ptq_evaluate.py [-h] --calibration-dir CALIBRATION_DIR --validation-dir VALIDATION_DIR [--workers WORKERS]
[--batch-size-calibration BATCH_SIZE_CALIBRATION] [--batch-size-validation BATCH_SIZE_VALIDATION]
[--export-dir EXPORT_DIR] [--gpu GPU] [--calibration-samples CALIBRATION_SAMPLES] [--model-name ARCH]
[--target-backend {fx,layerwise,flexml}] [--scale-factor-type {float_scale,po2_scale}]
[--act-bit-width ACT_BIT_WIDTH] [--weight-bit-width WEIGHT_BIT_WIDTH]
[--layerwise-first-last-bit-width LAYERWISE_FIRST_LAST_BIT_WIDTH] [--bias-bit-width {32,16,None}]
[--act-quant-type {sym,asym}] [--weight-quant-type {sym,asym}]
[--weight-quant-granularity {per_tensor,per_channel}] [--weight-quant-calibration-type {stats,mse}]
[--act-equalization {fx,layerwise,None}] [--act-quant-calibration-type {stats,mse}]
[--graph-eq-iterations GRAPH_EQ_ITERATIONS] [--learned-round-iters LEARNED_ROUND_ITERS]
[--learned-round-lr LEARNED_ROUND_LR] [--act-quant-percentile ACT_QUANT_PERCENTILE] [--export-onnx-qcdq]
[--export-torch-qcdq] [--scaling-per-output-channel | --no-scaling-per-output-channel]
[--bias-corr | --no-bias-corr] [--graph-eq-merge-bias | --no-graph-eq-merge-bias]
[--weight-narrow-range | --no-weight-narrow-range] [--gpfq-p GPFQ_P] [--quant-format {int,float}]
[--layerwise-first-last-mantissa-bit-width LAYERWISE_FIRST_LAST_MANTISSA_BIT_WIDTH]
[--layerwise-first-last-exponent-bit-width LAYERWISE_FIRST_LAST_EXPONENT_BIT_WIDTH]
[--weight-mantissa-bit-width WEIGHT_MANTISSA_BIT_WIDTH]
[--weight-exponent-bit-width WEIGHT_EXPONENT_BIT_WIDTH] [--act-mantissa-bit-width ACT_MANTISSA_BIT_WIDTH]
[--act-exponent-bit-width ACT_EXPONENT_BIT_WIDTH] [--gptq | --no-gptq] [--gpfq | --no-gpfq]
[--gptq-act-order | --no-gptq-act-order] [--learned-round | --no-learned-round]
[--calibrate-bn | --no-calibrate-bn]

PyTorch ImageNet PTQ Validation

optional arguments:
options:
-h, --help show this help message and exit
--calibration-dir CALIBRATION_DIR
Path to folder containing Imagenet calibration folder
Expand All @@ -110,49 +101,47 @@ optional arguments:
--gpu GPU GPU id to use (default: None)
--calibration-samples CALIBRATION_SAMPLES
Calibration size (default: 1000)
--model-name ARCH model architecture: alexnet | convnext_base |
convnext_large | convnext_small | convnext_tiny |
densenet121 | densenet161 | densenet169 | densenet201
| efficientnet_b0 | efficientnet_b1 | efficientnet_b2
| efficientnet_b3 | efficientnet_b4 | efficientnet_b5
| efficientnet_b6 | efficientnet_b7 |
efficientnet_v2_l | efficientnet_v2_m |
efficientnet_v2_s | googlenet | inception_v3 |
list_models | maxvit_t | mnasnet0_5 | mnasnet0_75 |
mnasnet1_0 | mnasnet1_3 | mobilenet_v2 |
mobilenet_v3_large | mobilenet_v3_small |
regnet_x_16gf | regnet_x_1_6gf | regnet_x_32gf |
regnet_x_3_2gf | regnet_x_400mf | regnet_x_800mf |
regnet_x_8gf | regnet_y_128gf | regnet_y_16gf |
regnet_y_1_6gf | regnet_y_32gf | regnet_y_3_2gf |
regnet_y_400mf | regnet_y_800mf | regnet_y_8gf |
resnet101 | resnet152 | resnet18 | resnet34 | resnet50
| resnext101_32x8d | resnext101_64x4d |
resnext50_32x4d | shufflenet_v2_x0_5 |
shufflenet_v2_x1_0 | shufflenet_v2_x1_5 |
shufflenet_v2_x2_0 | squeezenet1_0 | squeezenet1_1 |
swin_b | swin_s | swin_t | swin_v2_b | swin_v2_s |
swin_v2_t | vgg11 | vgg11_bn | vgg13 | vgg13_bn |
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)
--model-name ARCH model architecture: alexnet | convnext_base | convnext_large | convnext_small | convnext_tiny |
densenet121 | densenet161 | densenet169 | densenet201 | efficientnet_b0 | efficientnet_b1 |
efficientnet_b2 | efficientnet_b3 | efficientnet_b4 | efficientnet_b5 | efficientnet_b6 | efficientnet_b7
| efficientnet_v2_l | efficientnet_v2_m | efficientnet_v2_s | googlenet | inception_v3 | list_models |
maxvit_t | mnasnet0_5 | mnasnet0_75 | mnasnet1_0 | mnasnet1_3 | mobilenet_v2 | mobilenet_v3_large |
mobilenet_v3_small | regnet_x_16gf | regnet_x_1_6gf | regnet_x_32gf | regnet_x_3_2gf | regnet_x_400mf |
regnet_x_800mf | regnet_x_8gf | regnet_y_128gf | regnet_y_16gf | regnet_y_1_6gf | regnet_y_32gf |
regnet_y_3_2gf | regnet_y_400mf | regnet_y_800mf | regnet_y_8gf | resnet101 | resnet152 | resnet18 |
resnet34 | resnet50 | resnext101_32x8d | resnext101_64x4d | resnext50_32x4d | shufflenet_v2_x0_5 |
shufflenet_v2_x1_0 | shufflenet_v2_x1_5 | shufflenet_v2_x2_0 | squeezenet1_0 | squeezenet1_1 | swin_b |
swin_s | swin_t | swin_v2_b | swin_v2_s | swin_v2_t | vgg11 | vgg11_bn | vgg13 | vgg13_bn | 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 {fx,layerwise,flexml}
Backend to target for quantization (default: fx)
--scale-factor-type {float32,po2}
Type for scale factors (default: float32)
--scale-factor-type {float_scale,po2_scale}
Type for scale factors (default: float_scale)
--act-bit-width ACT_BIT_WIDTH
Activations bit width (default: 8)
--weight-bit-width WEIGHT_BIT_WIDTH
Weights bit width (default: 8)
--bias-bit-width {int32,int16}
Bias bit width (default: int32)
--act-quant-type {symmetric,asymmetric}
Activation quantization type (default: symmetric)
--layerwise-first-last-bit-width LAYERWISE_FIRST_LAST_BIT_WIDTH
Input and weights bit width for first and last layer w/ layerwise backend (default: 8)
--bias-bit-width {32,16,None}
Bias bit width (default: 32)
--act-quant-type {sym,asym}
Activation quantization type (default: sym)
--weight-quant-type {sym,asym}
Weight quantization type (default: sym)
--weight-quant-granularity {per_tensor,per_channel}
Activation quantization type (default: per_tensor)
--weight-quant-calibration-type {stats,mse}
Weight quantization calibration type (default: stats)
--act-equalization {fx,layerwise,None}
Activation equalization type (default: None)
--act-quant-calibration-type {stats,mse}
Activation quantization calibration type (default: stats)
--graph-eq-iterations GRAPH_EQ_ITERATIONS
Numbers of iterations for graph equalization (default: 20)
--learned-round-iters LEARNED_ROUND_ITERS
Numbers of iterations for learned round for each layer
(default: 1000)
Numbers of iterations for learned round for each layer (default: 1000)
--learned-round-lr LEARNED_ROUND_LR
Learning rate for learned round (default: 1e-3)
--act-quant-percentile ACT_QUANT_PERCENTILE
Expand All @@ -174,6 +163,20 @@ optional arguments:
--no-weight-narrow-range
Disable Narrow range for weight quantization (default: enabled)
--gpfq-p GPFQ_P P parameter for GPFQ (default: 0.25)
--quant-format {int,float}
Quantization format to use for weights and activations (default: int)
--layerwise-first-last-mantissa-bit-width LAYERWISE_FIRST_LAST_MANTISSA_BIT_WIDTH
Mantissa bit width used with float layerwise quantization for first and last layer (default: 4)
--layerwise-first-last-exponent-bit-width LAYERWISE_FIRST_LAST_EXPONENT_BIT_WIDTH
Exponent bit width used with float layerwise quantization for first and last layer (default: 3)
--weight-mantissa-bit-width WEIGHT_MANTISSA_BIT_WIDTH
Mantissa bit width used with float quantization for weights (default: 4)
--weight-exponent-bit-width WEIGHT_EXPONENT_BIT_WIDTH
Exponent bit width used with float quantization for weights (default: 3)
--act-mantissa-bit-width ACT_MANTISSA_BIT_WIDTH
Mantissa bit width used with float quantization for activations (default: 4)
--act-exponent-bit-width ACT_EXPONENT_BIT_WIDTH
Exponent bit width used with float quantization for activations (default: 3)
--gptq Enable GPTQ (default: enabled)
--no-gptq Disable GPTQ (default: enabled)
--gpfq Enable GPFQ (default: disabled)
Expand Down
59 changes: 55 additions & 4 deletions src/brevitas_examples/imagenet_classification/ptq/ptq_evaluate.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -79,9 +79,9 @@
help='Backend to target for quantization (default: fx)')
parser.add_argument(
'--scale-factor-type',
default='float',
choices=['float', 'po2'],
help='Type for scale factors (default: float)')
default='float_scale',
choices=['float_scale', 'po2_scale'],
help='Type for scale factors (default: float_scale)')
parser.add_argument(
'--act-bit-width', default=8, type=int, help='Activations bit width (default: 8)')
parser.add_argument(
Expand Down Expand Up @@ -168,6 +168,45 @@
help='Narrow range for weight quantization (default: enabled)')
parser.add_argument(
'--gpfq-p', default=0.25, type=float, help='P parameter for GPFQ (default: 0.25)')
parser.add_argument(
'--quant-format',
default='int',
choices=['int', 'float'],
help='Quantization format to use for weights and activations (default: int)')
parser.add_argument(
'--layerwise-first-last-mantissa-bit-width',
default=4,
type=int,
help=
'Mantissa bit width used with float layerwise quantization for first and last layer (default: 4)'
)
parser.add_argument(
'--layerwise-first-last-exponent-bit-width',
default=3,
type=int,
help=
'Exponent bit width used with float layerwise quantization for first and last layer (default: 3)'
)
parser.add_argument(
'--weight-mantissa-bit-width',
default=4,
type=int,
help='Mantissa bit width used with float quantization for weights (default: 4)')
parser.add_argument(
'--weight-exponent-bit-width',
default=3,
type=int,
help='Exponent bit width used with float quantization for weights (default: 3)')
parser.add_argument(
'--act-mantissa-bit-width',
default=4,
type=int,
help='Mantissa bit width used with float quantization for activations (default: 4)')
parser.add_argument(
'--act-exponent-bit-width',
default=3,
type=int,
help='Exponent bit width used with float quantization for activations (default: 3)')
add_bool_arg(parser, 'gptq', default=True, help='GPTQ (default: enabled)')
add_bool_arg(parser, 'gpfq', default=False, help='GPFQ (default: disabled)')
add_bool_arg(
Expand All @@ -191,6 +230,11 @@ def main():
config = (
f"{args.model_name}_"
f"{args.target_backend}_"
f"{args.quant_format}_"
f"{str(args.weight_mantissa_bit_width) + '_' if args.quant_format == 'float' else ''}"
f"{str(args.weight_exponent_bit_width) + '_' if args.quant_format == 'float' else ''}"
f"{str(args.act_mantissa_bit_width) + '_' if args.quant_format == 'float' else ''}"
f"{str(args.act_exponent_bit_width) + '_' if args.quant_format == 'float' else ''}"
f"{args.scale_factor_type}_"
f"a{args.act_bit_width}"
f"w{args.weight_bit_width}_"
Expand Down Expand Up @@ -295,7 +339,14 @@ def main():
act_bit_width=args.act_bit_width,
act_param_method=args.act_quant_calibration_type,
act_quant_percentile=args.act_quant_percentile,
act_quant_type=args.act_quant_type)
act_quant_type=args.act_quant_type,
quant_format=args.quant_format,
layerwise_first_last_mantissa_bit_width=args.layerwise_first_last_mantissa_bit_width,
layerwise_first_last_exponent_bit_width=args.layerwise_first_last_exponent_bit_width,
weight_mantissa_bit_width=args.weight_mantissa_bit_width,
weight_exponent_bit_width=args.weight_exponent_bit_width,
act_mantissa_bit_width=args.act_mantissa_bit_width,
act_exponent_bit_width=args.act_exponent_bit_width)
# If available, use the selected GPU
if args.gpu is not None:
torch.cuda.set_device(args.gpu)
Expand Down

0 comments on commit 1412174

Please sign in to comment.