Yolo-v3-tiny-tf model with INT-8 precision gives bad inferences

[AishaSamaanKhan]

Hi We are working on integrating yolo-v3-tiny-tf int-8 IR model into the dlstreamer pipeline following the documentation provided for changing model We were able to integrate yolo-v3-tiny-tf IR model (non quantized) and test it. but we failed to get proper inference with the INT-8 models for the same. these converted INT-8 model were validated using open model zoo sample for object detection and it was giving a proper inference.

The steps follow for conversion of the yolo-v3-tiny-tf to int8 model are provided below:

This Quantization document is based on yolo_v3_tiny_tf model

Requirements

• Openvino-dev 2022.1
• Openvino 2022.1

Steps for Quantization

Step 1: Obtain the OMZ model (yolo_v3_tiny_tf)

 omz_downloader --name yolo_v3_tiny_tf
 omz_converter --name yolo_v3_tiny_tf

This step downloads the frozen model and converts them to it's appropriate IR representation.

Step 2 : Obtain the DataSet for Optimization

For this Model COCO 2017 Validation dataset was selected

wget http://images.cocodataset.org/zips/val2017.zip

unzip val2017.zip

Step 3: Create a Json (Optional instead of pot arguments)

Note: Use FP16 to convert to FP16-INT8

{

        "model": {
                "model_name": "yolo-v3-tiny-tf",
                "model": "<path to the model>/yolo-v3-tiny-tf/FP32/yolo-v3-tiny-tf.xml",
                "weights": "<path to the model>/yolo-v3-tiny-tf/FP32/yolo-v3-tiny-tf.bin"
        },

        "engine": {
            "type": "simplified",
                "data_source": "<path to the dataset where the images are stored>/val2017"
        },

        "compression": {
            "target_device": "CPU",
             "algorithms": [
                {
                "name": "DefaultQuantization",
                "params": {
                            "preset": "performance",
                            "stat_subset_size": 300,
                            "shuffle_data": false
                }
                }
        ]
        }
}

Step 4: Use post optimization tool of the openvino to finish the process

This step converts the FP32/FP16 models to its FP32-INT8/FP16-INT8 models

the INT8 models will be available in "yolov3_int8" directory

pot -c quantization_spec.json --output-dir yolov3_int8 -d

Step 5 : Validation

Test the converted model with open model zoo demo object detection sample.

 python3 object_detection_demo.py   -d CPU   -i <path to the input video>  -m <path to INT8 model xml>   -at yolo   --labels <OMZ_DIR>/data/dataset_classes/coco_80cl.txt

For integrating the pipeline server the steps followed are as per the document

copy the downloaded and converted model under

<pipeline-server>/models/object_detection/yolo-v3-tiny-tf

Directory structure looks something like this under yolo-v3-tiny-tf

coco-80cl.txt  FP16  FP32  FP32-INT8  yolo-v3-tiny-tf  yolo-v3-tiny-tf.json

Created new pipeline

cp -r pipelines/gstreamer/object_detection/person_vehicle_bike pipelines/gstreamer/object_detection/yolo-v3-tiny-tf

Edited the template of pipeline.json

sed -i -e s/\\[person_vehicle_bike\\]/\\[yolo-v3-tiny-tf\\]/g pipelines/gstreamer/object_detection/yolo-v3-tiny-tf/pipeline.json

Ran the pipeline server

./docker/run.sh -v /tmp:/tmp --models models --pipelines pipelines/gstreamer

with this we were able to the inferencing with FP16, FP32 but we were not able to do inference with FP32-INT8 IR model

Could you please let us know what steps we are missing to integrate Quantized model ?

Thanks

[whbruce]

If you change the precision you need to specify it in the pipeline as per Referencing Models in Pipeline Definitions. See updated pipeline definition entry for gvadetect.

gvadetect model={models[object_detection][yolo-v2-tiny-tf][FP32-INT8][network]} name=detection

[AishaSamaanKhan]

@whbruce Thank you for the quick response, I tried Updating the pipeline definition as suggested, Still the issue persist.

These are the logs on the pipeline server, I could see that it is using the INT-8 models

{"levelname": "DEBUG", "asctime": "2022-07-22 01:23:06,054", "message": "gst-launch-1.0 urisourcebin  uri=https://github.com/intel-iot-devkit/sample-videos/raw/master/bottle-detection.mp4?raw=true ! decodebin ! gvadetect  model=/home/pipeline-server/models/object_detection/yolo-v3-tiny-tf/FP32-INT8/yolo-v3-tiny-tf.xml model-instance-id=detection_d69badd6095c11ed91000242ac110002 model-proc=/home/pipeline-server/models/object_detection/yolo-v3-tiny-tf/yolo-v3-tiny-tf.json labels=/home/pipeline-server/models/object_detection/yolo-v3-tiny-tf/coco-80cl.txt ! gvametaconvert  source=https://github.com/intel-iot-devkit/sample-videos/raw/master/bottle-detection.mp4?raw=true ! gvametapublish  file-path=/tmp/results.jsonl file-format=json-lines ! appsink  sync=False emit-signals=True", "module": "gstreamer_pipeline"}

over the inferencing side keeps incorrectly predicting bench continuously for uri=https://github.com/intel-iot-devkit/sample-videos/raw/master/bottle-detection.mp4?raw=true

Timestamp 5832402234
- bench (0.55) [0.00, 0.00, 0.98, 1.00]

[whbruce]

Please run OMZ object detection demo to get a baseline for model accuracy.

[AishaSamaanKhan]

@whbruce These are the same model that i used for the pipeline server I have tried this the open model zoo sample openvino 2022.1.0 version, Below are the logs and attachment for the same

(setup) intel@intel-WL10:~/workspace/open_model_zoo/demos/object_detection_demo/python$ python3 object_detection_demo.py -d CPU -i bottle.mp4 -m /home/intel/workspace/pipeline-server/models/object_detection/yolo-v3-tiny-tf/FP32-INT8/yolo-v3-tiny-tf.xml -at yolo --labels /home/intel/workspace/pipeline-server/models/object_detection/yolo-v3-tiny-tf/coco-80cl.txt [ INFO ] OpenVINO Runtime [ INFO ] build: 2022.1.0-7019-cdb9bec7210-releases/2022/1 [ INFO ] Reading model /home/intel/workspace/pipeline-server/models/object_detection/yolo-v3-tiny-tf/FP32-INT8/yolo-v3-tiny-tf.xml [ WARNING ] The parameter "input_size" not found in YOLO wrapper, will be omitted [ WARNING ] The parameter "num_classes" not found in YOLO wrapper, will be omitted [ INFO ] Input layer: image_input, shape: [1, 416, 416, 3], precision: f32, layout: NHWC [ INFO ] Output layer: conv2d_12/Conv2D/YoloRegion, shape: [1, 255, 26, 26], precision: f32, layout: [ INFO ] Output layer: conv2d_9/Conv2D/YoloRegion, shape: [1, 255, 13, 13], precision: f32, layout: [ INFO ] The model /home/intel/workspace/pipeline-server/models/object_detection/yolo-v3-tiny-tf/FP32-INT8/yolo-v3-tiny-tf.xml is loaded to CPU [ INFO ] Device: CPU [ INFO ] Number of streams: 4 [ INFO ] Number of threads: AUTO [ INFO ] Number of model infer requests: 5 [ INFO ] Metrics report: [ INFO ] Latency: 84.1 ms [ INFO ] FPS: 30.5 [ INFO ] Decoding: 0.4 ms [ INFO ] Preprocessing: 0.6 ms [ INFO ] Inference: 81.0 ms [ INFO ] Postprocessing: 1.9 ms [ INFO ] Rendering: 0.2 ms

[brmarkus]

Using the default-algorithm for INT8 quantization (instead of using the "accuracy-aware" algorithm) can come with a slightly reduced accuracy...

The sample "object_detection_demo.py" has a default-value '-t', '--prob_threshold', default=0.5, which you might need to increase to filter-out those with lower confidence-level.

gvadetect has the same default-value:

  threshold           : Threshold for detection results. Only regions of interest with confidence values above the threshold will be added to the frame
                        flags: readable, writable
                        Float. Range: 0 - 1 Default: 0.5

[AishaSamaanKhan]

Hi, I tried using the Accuracy Aware Quantization to create the Quantized model, And Followed the steps to integrate the this model into the pipeline server i still see same behavior for the inference below is the config for pot quantization

{
    "model": {
        "model_name": "yolo-v3-tiny-tf",
        "model": "yolo_v3_tiny/FP32/yolo-v3-tiny-tf.xml",
        "weights": "yolo_v3_tiny/FP32/yolo-v3-tiny-tf.bin"
    },
    "compression": {
        "dump_intermediate_model": true,
        "target_device": "ANY",
        "algorithms": [
            {
                "name": "AccuracyAwareQuantization",
                "params": {
                    "num_samples_for_tuning": 2000,
                    "preset": "performance",
                    "stat_subset_size": 300,
                    "use_layerwise_tuning": false
                }
            }
        ]
    },
    "engine": {
        "launchers": [
            {
                "framework": "openvino",
                "adapter": {
                    "type": "yolo_v3",
                    "outputs": [
                        "conv2d_9/Conv2D/YoloRegion",
                        "conv2d_12/Conv2D/YoloRegion"
                    ]
                },
                "device": "CPU"
            }
        ],
        "datasets": [
            {
                "name": "ms_coco_detection_80_class_without_background",
                "data_source": "val2017/val2017/",
                "annotation_conversion": {
                    "converter": "mscoco_detection",
                    "annotation_file": "./annotations/instances_val2017.json",
                    "has_background": false,
                    "sort_annotations": true,
                    "use_full_label_map": false
                },
                "annotation": "./dir_conv/mscoco_det_80.pickle",
                "dataset_meta": "./dir_conv/mscoco_det_80.json",
                "preprocessing": [
                    {
                        "type": "resize",
                        "size": 416

                    }
                ],
                "postprocessing": [
                    {
                        "type": "resize_prediction_boxes"
                    },
                    {
                        "type": "clip_boxes",
                        "apply_to": "prediction"
                    }
                ],
                "metrics": [
                    {
                            "type": "coco_precision",
                            "reference": 0.244
                    }
                ]
            }
        ]
    }
}

I am able to use it with the omz object detection sample, However it seems to fail on the pipeline server, could you please suggest any solution for this issue

[brmarkus]

Can you provide more details about the problem you are seeing, please? After INT8-quantization no objects were detected at all anymore, compared to the "original models" (F16, FP32)?

[AishaSamaanKhan]

yes, @brmarkus
With the downloaded models from the open model zoo (i.e.,The Original models), the pipeline server is able to give the inference, but when it comes the INT-8 Quantized model based on Accuracy Aware Algo the inference is failing with the pipeline server These same Quantized model were validated with the object detection sample from the OMZ repo.