INT8 Quantization with AMD Quark

Model Quantization section gives an overview of Versal AI Edge Gen 2 quantization using the AMD Quark tool.

This section demonstrates how to perform INT8 quantization using AMD Quark, and using the quantized INT8 model in AMD Vitis™ AI.

For BF16 quantization, Quark does not need to be invoked explicitly. The Vitis AI compiler automatically handles the FP32 to BF16 quantization process.

Basic Example: Quantizing ResNet50

The following steps outline the process for quantizing a ResNet50 ONNX model.

1. Download the ONNX Model

On your host machine, download the ResNet50 ONNX float model from the onnxmodelzoo model zoo:

cd <Working Path>
wget -P models https://huggingface.co/onnxmodelzoo/resnet50-v1-12/resolve/main/resnet50-v1-12.onnx

2. Prepare Calibration Data & Validation Data

Create a folder for calibration data and download sample images (PNG or JPG). For example, use images from microsoft/onnxruntime-inference-examples as a quick start.

mkdir calib_data
wget -O calib_data/daisy.jpg https://github.com/microsoft/onnxruntime-inference-examples/blob/main/quantization/image_classification/cpu/test_images/daisy.jpg?raw=true
chmod -R a+w <Working Path>

And prepare some pictures for validation:

mkdir val_data
wget -O val_data/daisy.jpg https://github.com/microsoft/onnxruntime-inference-examples/blob/main/quantization/image_classification/cpu/test_images/daisy.jpg?raw=true
wget -O val_data/rose.jpg https://github.com/microsoft/onnxruntime-inference-examples/blob/main/quantization/image_classification/cpu/test_images/rose.jpg?raw=true
wget -O val_data/tulip.jpg https://github.com/microsoft/onnxruntime-inference-examples/blob/main/quantization/image_classification/cpu/test_images/tulip.jpg?raw=true

3. Launch Vitis AI Docker

Refer to the Docker Setup section for steps to load the docker. Start the Docker container:

docker run -it --rm --network host -v </yourLicenseDir>:/usr/licenses -v <Working Path>:<Path in Docker>  <Docker REPOSITORY>:<Docker TAG> "bash"

4. Create the Quantization Script

Inside the docker, create a file quantize.py with the following contents:

import os
import cv2
import numpy as np
import onnx
from torchvision import transforms
from onnxruntime.quantization import CalibrationDataReader

def get_model_input_name(input_model_path: str) -> str:
        model = onnx.load(input_model_path)
        model_input_name = model.graph.input[0].name
        return model_input_name

class ImageDataReader(CalibrationDataReader):

        def __init__(self, calibration_image_folder: str, input_name: str):
                self.enum_data = None

                self.input_name = input_name

                self.data_list = self._preprocess_images(
                                calibration_image_folder)

        def _preprocess_images(self, image_folder: str):
                data_list = []
                img_names = [f for f in os.listdir(image_folder) if f.endswith('.png') or f.endswith('.jpg')]
                for name in img_names:
                        input_image = cv2.imread(os.path.join(image_folder, name))
                        # Resize the input image. Because the size of Resnet50 is 224.
                        input_image = cv2.resize(input_image, (224, 224))
                        input_data = np.array(input_image).astype(np.float32)
                        # Custom Pre-Process
                        input_data = input_data.transpose(2, 0, 1)
                        input_size = input_data.shape
                        if input_size[1] > input_size[2]:
                                input_data = input_data.transpose(0, 2, 1)
                        input_data = np.expand_dims(input_data, axis=0)
                        input_data = input_data / 255.0
                        data_list.append(input_data)

                return data_list

        def get_next(self):
                if self.enum_data is None:
                        self.enum_data = iter([{self.input_name: data} for data in self.data_list])
                return next(self.enum_data, None)

        def rewind(self):
                self.enum_data = None

from quark.onnx.quantization.config import Config, get_default_config
from quark.onnx import ModelQuantizer
from onnxruntime.quantization import QuantType

quant_config = get_default_config("VINT8")
quant_config.extra_options["Int32Bias"] = False
quant_config.enable_npu_cnn = True

# Set up quantization with a specified configuration
quantization_config = Config(global_quant_config=quant_config)
quantizer = ModelQuantizer(quantization_config)

float_model_path = "models/resnet50-v1-12.onnx"
quantized_model_path = "models/resnet50-v1-12_quantized.onnx"
calib_data_path = "calib_data"
model_input_name = get_model_input_name(float_model_path)
calib_data_reader = ImageDataReader(calib_data_path, model_input_name)

# Quantize the ONNX model and save to specified path
quantizer.quantize_model(float_model_path, quantized_model_path, calibration_data_reader=calib_data_reader)

5. Run Quark Quantizer

Use the quantize.py created above and run quantization:

python3 quantize.py

Example output:

+----------------------+--------------------------------------+
| Op Type              | Float Model                          |
+======================+======================================+
| Conv                 | 53                                   |
| BatchNormalization   | 53                                   |
| Relu                 | 49                                   |
| MaxPool              | 1                                    |
| Add                  | 16                                   |
| GlobalAveragePool    | 1                                    |
| Flatten              | 1                                    |
| Gemm                 | 1                                    |
+----------------------+--------------------------------------+
| Quantized model path | models/resnet50-v1-12_quantized.onnx |
+----------------------+--------------------------------------+

+-------------------+------------+----------+----------+
| Op Type           | Activation | Weights  | Bias     |
+===================+============+==========+==========+
| Conv              | INT8(53)   | INT8(53) | INT8(53) |
| Relu              | INT8(49)   |          |          |
| MaxPool           | INT8(1)    |          |          |
| Add               | INT8(16)   |          |          |
| GlobalAveragePool | INT8(1)    |          |          |
| Flatten           | INT8(1)    |          |          |
| Gemm              | INT8(1)    | INT8(1)  | INT8(1)  |
+-------------------+------------+----------+----------+

The output quantized model is saved to models/resnet50-v1-12_quantized.onnx. And then go next steps to do Vitis AI compilation.

Note

The ImageDataReader function in quantize.py is specific to ResNet50 model. For other models, it needs to be modified according to the model requirements.

6. Compile the Quantized Model in Vitis AI

Create the file compile.py with the following contents:

import onnxruntime

provider_options_dict = {
        "config_file": 'vitisai_config.json',
        "cache_dir":   'my_cache_dir',
        "cache_key":   'resnet50-v1-12_quantized',
        "log_level":   'info',
        "target": "VAIML"
}

session = onnxruntime.InferenceSession(
        'models/resnet50-v1-12_quantized.onnx',
        providers=["VitisAIExecutionProvider"],
        provider_options=[provider_options_dict]
)

Create configuration file vitisai_config.json with:

{
 "passes": [
         {
                 "name": "init",
                 "plugin": "vaip-pass_init"
         },
         {
                 "name": "vaiml_partition",
                 "plugin": "vaip-pass_vaiml_partition",
                 "vaiml_config": {
                        "device": "ve2-xc2ve3858",
                        "keep_outputs": true,
                        "optimize_level": 2,
                        "threshold_gops_percent": 20,
                        "logging_level": "info"
                }
         }
 ],
 "target": "VAIML",
 "targets": [
    {
        "name": "VAIML",
        "pass": [
            "init",
            "vaiml_partition"
        ]
    }
 ]
}

Run Vitis AI compilation with the following command:

python3 compile.py

It takes time to compile the model. After compilation you get the following output:

subpartition path = my_cache_dir/resnet50-v1-12_quantized/vaiml_par_0/0
INFO: [VAIP-VAIML-PASS] No. of Operators :
INFO:  VAIML     491
INFO: [VAIP-VAIML-PASS] No. of Subgraphs :
INFO:    NPU     1
INFO: [VAIP-VAIML-PASS] For detailed compilation results, please refer to my_cache_dir/resnet50-v1-12_quantized/final-vaiml-pass-summary.txt

You can get more details about the compilation results by displaying the content of`my_cache_dir/resnet50-v1-12_quantized/final-vaiml-pass-summary.txt` :

--------- Final Summary of VAIML Pass ----------
OS: Linux X64
VAIP commit: bd5c863c3084e7534f5cd59173c631bb8fa07491
Model: /Resnet/models/resnet50-v1-12_quantized.onnx
Model signature: 5633833a66fc76d4758bcb34d9f75ee3
Device: ve2
Model data type: int8 quantized
Device data type: int8
Number of operators in the model: 493
GOPs of the model: 8.09703
Number of operators supported by VAIML: 491 (99.594%)
GOPs supported by VAIML: 8.096 (99.993%)
Number of subgraphs supported by VAIML: 1
Number of operators offloaded by VAIML: 491 (99.594%)
GOPs offloaded by VAIML: 8.096 (99.993%)
Number of subgraphs offloaded by VAIML: 1
Number of subgraphs with compilation errors (fall back to CPU): 0
Number of subgraphs below 20% GOPs threshold (fall back to CPU): 0
Number of subgraphs above max number of subgraphs allowed(7): 0 (fall back to CPU)
Stats for offloaded subgraphs
Subgraph vaiml_par_0 stats:
    Type: npu
    Operators: 491 (99.594%)
    GOPs : 8.096 (99.993%)  OPs: 8,096,419,216

7. Run Inference

Create the file runmodel_pre_cpu.py with the following contents:

import numpy as np
import onnxruntime as ort
import os
import cv2
from torchvision import transforms

onnx_model='models/resnet50-v1-12.onnx'
onnx_model_vint8 ='models/resnet50-v1-12_quantized.onnx'

# CPU session
cpu_session = ort.InferenceSession(
    onnx_model,
    providers=["CPUExecutionProvider"]
)

# CPU session with VINT8 quantized model
cpu_session_vint8 = ort.InferenceSession(
    onnx_model_vint8,
    providers=["CPUExecutionProvider"]
)

# You can define your preprocess method
def preprocess_image(image_path):
    transform = transforms.Compose([
        transforms.ToPILImage(),
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    ])

    image = cv2.imread(image_path)
    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    image = transform(image)
    image = image.unsqueeze(0)
    image = image.numpy()
    return image

input_folder="val_data"
def run_inference(session, output_dir="output_cpu"):
    files = sorted([f for f in os.listdir(input_folder) if f.endswith(".jpg")])
    input_name = session.get_inputs()[0].name
    runs=0
    for i,f in enumerate(files):
        runs+=1
        fp = os.path.join(input_folder, f)
        image = preprocess_image(fp)
        os.makedirs("input", exist_ok=True)
        np.save(f"input/input_{i}.npy", image)
        outputs = session.run(None, {input_name:image})
        # Create outpu directory if it doesn't exist
        os.makedirs(output_dir, exist_ok=True)
        for idx, out in enumerate(outputs):
            np.save(f"{output_dir}/output_{i}_{idx}.npy", out)

run_inference(cpu_session,"output_cpu")
run_inference(cpu_session_vint8,"output_cpu_vint8")

Run following command inside the docker:

python3 runmodel_pre_cpu.py

It runs inferences on CPU and save input numpy vectors for NPU processing. Next, create file runmodel.py with the following contents:

import numpy as np
import onnxruntime as ort
import os

onnx_model_vint8 ='models/resnet50-v1-12_quantized.onnx'

provider_options_dict = {
    "config_file": 'vitisai_config.json',
    "cache_dir":   'my_cache_dir',
    "cache_key":   'resnet50-v1-12_quantized',
    "log_level":   'info',
        "target": "VAIML"
}

# NPU session
npu_session = ort.InferenceSession(
    onnx_model_vint8,
    providers=["VitisAIExecutionProvider"],
    provider_options=[provider_options_dict]
)

input_folder="input"
output_folder="output_vek385"
files = sorted([f for f in os.listdir(input_folder) if f.endswith(".npy")])
input_name = npu_session.get_inputs()[0].name
for i,f in enumerate(files):
    fp = os.path.join(input_folder, f)
    image = np.load(fp)
    outputs = npu_session.run(None, {input_name:image})
    # Create outpu directory if it doesn't exist
    os.makedirs(output_folder, exist_ok=True)
    for idx, out in enumerate(outputs):
        np.save(f"{output_folder}/output_{i}_{idx}.npy", out)
print("Inference done")

Refer to Hardware Setup to setup the VEK385 board. On the board, run inference with the following command:

python3 runmodel.py

The script runs three inferences of the model and displays messages similar to the following:

Inference done

If you want to see detailed NPU execution logs, set the environment variable DEBUG_LOG_LEVEL=info before running the script:

export DEBUG_LOG_LEVEL=info
python runmodel.py

The output includes the number of VAIML operators and the number of NPU-executed subgraphs:

I20380419 04:55:16.083313  2046 stat.cpp:193] [Vitis AI EP] No. of Operators :
I20380419 04:55:16.083366  2046 stat.cpp:204]  VAIML   491
I20380419 04:55:16.083384  2046 stat.cpp:204] VITIS_EP_CPU     2
I20380419 04:55:16.083395  2046 stat.cpp:213]
I20380419 04:55:16.083404  2046 stat.cpp:218] [Vitis AI EP] No. of Subgraphs :
I20380419 04:55:16.083410  2046 stat.cpp:226]    NPU     1
I20380419 04:55:16.083419  2046 stat.cpp:229] Actually running on NPU      1
I20380419 04:55:16.088616  2046 vitisai_compile_model.cpp:1477] AVG CPU Usage 93.3333%
I20380419 04:55:16.088671  2046 vitisai_compile_model.cpp:1478] Peak Working Set size 129.445 MB
2038-04-19 04:55:16.093139780 [W:onnxruntime:, session_state.cc:1316 VerifyEachNodeIsAssignedToAnEp] Some nodes were not assigned to the preferred execution providers which may or may not have an negative impact on performance. e.g. ORT explicitly assigns shape related ops to CPU to improve perf.
2038-04-19 04:55:16.093185510 [W:onnxruntime:, session_state.cc:1318 VerifyEachNodeIsAssignedToAnEp] Rerunning with verbose output on a non-minimal build will show node assignments.
[2038-04-19 04:55:16.146] [console] [info] [FLEXMLRT] /proj/aiebuilds/IPU/ONNX-EP/IPU/Vitis-AI-IPU-ONNX_EP/RAI_1_7_VAI_RT_BUILD/2242/arm/flexml_rt/src/FlexMLClient.cpp:1269 FlexMLRT Git Hash: 512d4e65
Inference done

8. Evaluation

Quark provides a tool to compare the differences between different inference results using L2 Loss and other metrics. For example, run following command inside the docker:

python3 -m quark.onnx.tools.evaluate --baseline_results_folder output_cpu_vint8 --quantized_results_folder output_vek385

Example output:

Mean cos similarity: 0.9987082481384277
Min cos similarity: 0.9984074831008911
Mean l2 distance: 3.657916784286499
Max l2 distance: 4.2315778732299805
Mean psnr: 38.722272316614784
Min psnr: 36.79283380508423
Mean ssim: 0.9994073510169983
Min ssim: 0.9993097186088562

In addition to the built-in evaluation capabilities, you can also develop and integrate domain specific post-processing scripts. These scripts can be tailored to specific use cases, enabling more detailed comparison of results or the computation of custom accuracy metrics.

Quantization Using Random Data

To use random data rather than calibration images, replace quantize.py with:

import os
import cv2
import numpy as np
from torchvision import transforms

from quark.onnx.quantization.config import Config, get_default_config
from quark.onnx import ModelQuantizer
from onnxruntime.quantization import QuantType

quant_config = get_default_config("VINT8")
quant_config.extra_options["Int32Bias"] = False
quant_config.enable_npu_cnn = True
quant_config.extra_options['UseRandomData'] = True

# Set up quantization with a specified configuration
quantization_config = Config(global_quant_config=quant_config)
quantizer = ModelQuantizer(quantization_config)
float_model_path = "models/resnet50-v1-12.onnx"
quantized_model_path = "models/resnet50-v1-12_quantized.onnx"
# Quantize the ONNX model and save to specified path
quantizer.quantize_model(float_model_path, quantized_model_path, calibration_data_reader=None)

Execute:

python3 quantize.py

Other steps remain same for quantization with calibration data or random data.

For more information about Quark quantization, refer to https://quark.docs.amd.com/latest/.