diff --git a/src/brevitas_examples/imagenet_classification/ptq/README.md b/src/brevitas_examples/imagenet_classification/ptq/README.md index 29386659b..b9add94bb 100644 --- a/src/brevitas_examples/imagenet_classification/ptq/README.md +++ b/src/brevitas_examples/imagenet_classification/ptq/README.md @@ -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 @@ -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 @@ -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) diff --git a/src/brevitas_examples/imagenet_classification/ptq/ptq_evaluate.py b/src/brevitas_examples/imagenet_classification/ptq/ptq_evaluate.py index 319b429b7..dd49e8531 100644 --- a/src/brevitas_examples/imagenet_classification/ptq/ptq_evaluate.py +++ b/src/brevitas_examples/imagenet_classification/ptq/ptq_evaluate.py @@ -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( @@ -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( @@ -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}_" @@ -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)