Skip to content

Commit

Permalink
Docs (examples/ptq): update README
Browse files Browse the repository at this point in the history
  • Loading branch information
@Giuseppe5
Giuseppe5 committed Feb 19, 2024
1 parent 40bce51 commit 2cde5ed
Showing 1 changed file with 141 additions and 70 deletions.
211 changes: 141 additions & 70 deletions src/brevitas_examples/imagenet_classification/ptq/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -24,10 +24,12 @@ The implementation for programmatic quantization is still experimental and might

For both these flows, the following options are exposed:
- Bit-width of weight and activations.
- In case of minifloat quantization, the exponent and mantissa bit-width of weights and activations.
- Scales can be either float32 or power-of-two (po2) numbers.
- Weights' scale factors can be either per-tensor or per-channel.
- Biases can be int16 or int32.
- Biases can be floating point, int16, or int32.
- Activation quantization can be symmetric or asymmetric.
- Possibility to use statistics or MSE for scale factor computations for weights and activations.
- Percentiles used for the activations' statistics computation during calibration.

Furthermore, Brevitas additional PTQ techniques can be enabled:
Expand All @@ -37,6 +39,8 @@ Furthermore, Brevitas additional PTQ techniques can be enabled:
- GPTQ [<sup>4 </sup>].
- Learned Round [<sup>5 </sup>].
- GPFQ [<sup>6 </sup>].
- Channel splitting [<sup>7 </sup>].
- Activation Equalization [<sup>8 </sup>].


Internally, when defining a quantized model programmatically, Brevitas leverages `torch.fx` and its `symbolic_trace` functionality, meaning that an input model is required to pass symbolic tracing for it to work.
Expand All @@ -61,27 +65,51 @@ 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 {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}]
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] [--dtype {float,bfloat16}]
[--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]
[--weight-exponent-bit-width WEIGHT_EXPONENT_BIT_WIDTH]
[--act-mantissa-bit-width ACT_MANTISSA_BIT_WIDTH]
[--act-exponent-bit-width ACT_EXPONENT_BIT_WIDTH]
[--accumulator-bit-width ACCUMULATOR_BIT_WIDTH]
[--onnx-opset-version ONNX_OPSET_VERSION]
[--channel-splitting-ratio CHANNEL_SPLITTING_RATIO]
[--gptq | --no-gptq] [--gpfq | --no-gpfq]
[--gpfa2q | --no-gpfa2q]
[--gpxq-act-order | --no-gpxq-act-order]
[--learned-round | --no-learned-round]
[--calibrate-bn | --no-calibrate-bn]
[--channel-splitting-split-input | --no-channel-splitting-split-input]
[--merge-bn | --no-merge-bn]

PyTorch ImageNet PTQ Validation

Expand All @@ -101,19 +129,34 @@ options:
--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)
--dtype {float,bfloat16}
Data type to use
--target-backend {fx,layerwise,flexml}
Backend to target for quantization (default: fx)
--scale-factor-type {float_scale,po2_scale}
Expand All @@ -123,7 +166,8 @@ options:
--weight-bit-width WEIGHT_BIT_WIDTH
Weights bit width (default: 8)
--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)
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}
Expand All @@ -137,56 +181,95 @@ options:
--act-equalization {fx,layerwise,None}
Activation equalization type (default: None)
--act-quant-calibration-type {stats,mse}
Activation quantization calibration type (default: stats)
Activation quantization calibration type (default:
stats)
--graph-eq-iterations GRAPH_EQ_ITERATIONS
Numbers of iterations for graph equalization (default: 20)
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
Percentile to use for stats of activation quantization (default: 99.999)
Percentile to use for stats of activation quantization
(default: 99.999)
--export-onnx-qcdq If true, export the model in onnx qcdq format
--export-torch-qcdq If true, export the model in torch qcdq format
--scaling-per-output-channel
Enable Weight scaling per output channel (default: enabled)
Enable Weight scaling per output channel (default:
enabled)
--no-scaling-per-output-channel
Disable Weight scaling per output channel (default: enabled)
--bias-corr Enable Bias correction after calibration (default: enabled)
--no-bias-corr Disable Bias correction after calibration (default: enabled)
Disable Weight scaling per output channel (default:
enabled)
--bias-corr Enable Bias correction after calibration (default:
enabled)
--no-bias-corr Disable Bias correction after calibration (default:
enabled)
--graph-eq-merge-bias
Enable Merge bias when performing graph equalization (default: enabled)
Enable Merge bias when performing graph equalization
(default: enabled)
--no-graph-eq-merge-bias
Disable Merge bias when performing graph equalization (default: enabled)
Disable Merge bias when performing graph equalization
(default: enabled)
--weight-narrow-range
Enable Narrow range for weight quantization (default: enabled)
Enable Narrow range for weight quantization (default:
disabled)
--no-weight-narrow-range
Disable Narrow range for weight quantization (default: enabled)
--gpfq-p GPFQ_P P parameter for GPFQ (default: 0.25)
Disable Narrow range for weight quantization (default:
disabled)
--gpfq-p GPFQ_P P parameter for GPFQ (default: 1.0)
--quant-format {int,float}
Quantization format to use for weights and activations (default: int)
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)
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)
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)
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)
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)
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)
Exponent bit width used with float quantization for
activations (default: 3)
--accumulator-bit-width ACCUMULATOR_BIT_WIDTH
Accumulator Bit Width for GPFA2Q (default: None)
--onnx-opset-version ONNX_OPSET_VERSION
ONNX opset version
--channel-splitting-ratio CHANNEL_SPLITTING_RATIO
Split Ratio for Channel Splitting. When set to 0.0,
Channel Splitting will not be applied. (default: 0.0)
--gptq Enable GPTQ (default: disabled)
--no-gptq Disable GPTQ (default: disabled)
--gpfq Enable GPFQ (default: disabled)
--no-gpfq Disable GPFQ (default: disabled)
--gptq-act-order Enable GPTQ Act order heuristic (default: disabled)
--no-gptq-act-order Disable GPTQ Act order heuristic (default: disabled)
--gpfa2q Enable GPFA2Q (default: disabled)
--no-gpfa2q Disable GPFA2Q (default: disabled)
--gpxq-act-order Enable GPxQ Act order heuristic (default: disabled)
--no-gpxq-act-order Disable GPxQ Act order heuristic (default: disabled)
--learned-round Enable Learned round (default: disabled)
--no-learned-round Disable Learned round (default: disabled)
--calibrate-bn Enable Calibrate BN (default: disabled)
--no-calibrate-bn Disable Calibrate BN (default: disabled)
--channel-splitting-split-input
Enable Input Channels Splitting for channel splitting
(default: disabled)
--no-channel-splitting-split-input
Disable Input Channels Splitting for channel splitting
(default: disabled)
--merge-bn Enable Merge BN layers before quantizing the model
(default: enabled)
--no-merge-bn Disable Merge BN layers before quantizing the model
(default: enabled)

```

The script requires to specify the calibration folder (`--calibration-dir`), from which the calibration samples will be taken (configurable with the `--calibration-samples` argument), and a validation folder (`--validation-dir`).
Expand All @@ -197,23 +280,11 @@ brevitas_ptq_imagenet_val --calibration-dir /path/to/imagenet/calibration/folder
```


## Benchmark flow

This scripts evaluate a variety of quantization configurations on different models.

For example, to run the script on the GPU 0:
```bash
brevitas_ptq_imagenet_benchmark --calibration-dir /path/to/imagenet/calibration/folder --validation-dir /path/to/imagenet/validation/folder --gpu 0
```

After launching the script, a `RESULTS_TORCHVISION.csv` markdown file will be generated with the results on the torchvision models,
and a `RESULTS_IMGCLSMOB.csv` with the results on manually quantized models starting from floating point weights.



[<sup>1 </sup>]: https://arxiv.org/abs/1906.04721
[<sup>2 </sup>]: https://github.com/Xilinx/Vitis-AI/blob/50da04ddae396d10a1545823aca30b3abb24a276/src/vai_quantizer/vai_q_pytorch/nndct_shared/optimization/commander.py#L450
[<sup>3 </sup>]: https://github.com/openppl-public/ppq/blob/master/ppq/quantization/algorithm/equalization.py
[<sup>4 </sup>]: https://arxiv.org/abs/2210.17323
[<sup>5 </sup>]: https://arxiv.org/abs/2004.10568
[<sup>6 </sup>]: https://arxiv.org/abs/2201.11113
[<sup>7 </sup>]: https://arxiv.org/abs/1901.09504
[<sup>8 </sup>]: https://arxiv.org/abs/2211.10438

0 comments on commit 2cde5ed

Please sign in to comment.