 Kria™ KV260 Vision AI Starter Kit
|
Customizing the AI Models used in the application |
Customizing the AI Models used in the application¶
Introduction¶
This document provides an overview of how to customize the default smartcam application to use other AI models.
Prerequisites¶
Other than the 3 models provided via command line option --aitask as documented here,
facedetect (densebox_320_320)
refinedet (refinedet_pruned_0_96)
ssd (ssd_adas_pruned_0_95)
customization can be made to use other Vitis AI models or retrained model by the users of the same class.
Model Preparation¶
Note The design currently only supports Vitis AI 2.5.0
Note As described in the Hardware Accelerator section, the DPU integrated in the platform uses the B3136 configuration.
The arch.json used to compile the xmodel for B3136 DPU can be obtained by build the accelerator, but if you won’t build all from the start, you can obtain the DPU fingerprint by using the following commands on target SOM:
xdputil query
You will get some output like this:
...
"kernels":[
{
...
"fingerprint":"0x101000016010406",
...
}
]
Save the fingerprint as arch.json.
{
"fingerprint":"0x101000016010406"
}
For detailed instructions on obtaining an alternative model from the Xilinx model zoo or training, pruning, quantizing, and compiling a new model, please refer to the Vitis AI documentation.
Configuration files¶
To integrate a different .xmodel into the smartcam application, the following configuration files must be updated accordingly:
AI Inference Config:
Take the refinedet aiinference.json
/opt/xilinx/kv260-smartcam/share/vvas/refinedet/aiinference.jsonas an example,
{
"vvas-library-repo": "/usr/lib/",
"element-mode":"inplace",
"kernels" :[
{
"library-name":"libvvas_xdpuinfer.so",
"config": {
"model-name" : "refinedet_pruned_0_96",
"model-class" : "REFINEDET",
"model-path" : "/opt/xilinx/kv260-smartcam/share/vitis_ai_library/models",
"run_time_model" : false,
"need_preprocess" : false,
"performance_test" : false,
"debug_level" : 0
}
}
]
}
You can change the “model-name” and “model-path” fields to use the customized xmdel file at ${model-path}/${model-name}/${model-name}.xmodel.
Pay attention to the field “need_preprocess”, which is now “false”, tells Vitis AI Library the input buffer is already the resized and quantized BGR image as required by the model. And the preprocess is done by the preprocess plugin with the proper configuration which will be detailed in next section.
When you set the “need_preprocess” here to “true” for some reason, you should also make change to the process configuration to ask the preprocess IP works just as colour conversion and resizing.
Preprocess Config:
"config": {
"debug_level" : 1,
"mean_r": 123,
"mean_g": 117,
"mean_b": 104,
"scale_r": 0.5,
"scale_g": 0.5,
"scale_b": 0.5
}
The configuration value of mean for r/g/b channels should be the same as the ones specified in Vitis AI Model prototxt file.
model {
name : "refinedet_480x360_5G"
kernel {
name: "refinedet_480x360_5G"
mean: 104.0
mean: 117.0
mean: 123.0
scale: 1.0
scale: 1.0
scale: 1.0
}
}
Notice the channels sequence in the Vitis AI model prototxt file is B, G, R, not R, G, B, as the above samples show.
The configuration value of scale is determined by both prototxt and input tensor’s fixpos info contained in compiled xmodel file.
First use the following command to get input tensor’s fixpos.
xdputil xmodel Path-to-xmodelfile -l
You will get some output like this:
{
"subgraphs":[
{
"index":0,
"name":"subgraph_data",
"device":"USER"
},
{
"index":0,
"name":"subgraph_Elt3",
"device":"DPU",
"fingerprint":"0x101000016010406",
"DPU Arch":"DPUCZDX8G_ISA1_B3136",
"input_tensor":[
{
"name":"data_fixed",
"shape":"[1, 360, 480, 3]",
"fixpos":-1
}
],
}
]
}
So the input tensor’s fixpos is “-1” here.
The configured scale = prototxt_scale * (2 ^ fixpos) = 1 * (2^-1) = 0.5.
Example¶
Take Vitis AI model yolov3_coco_416_tf2 as example, we provide the detailed steps to add an AI task for smartcam application.
Create folder
yolov3_cocounder/opt/xilinx/kv260-smartcam/share/vvas/, so thatyolov3_cococan be used as the value for argument –AItask.Download model file for
GPUandzcu102 & zcu104 & kv260from the link provided by https://github.com/Xilinx/Vitis-AI/blob/master/model_zoo/model-list/tf2_yolov3_coco_416_416_65.9G_2.5/model.yaml
After extraction, we get such file structure:
│ README.md
│ requirements.txt
│
├───code
│ └───test
│ │ cocoval.py
│ │ convert_data.sh
│ │ Dockerfile
│ │ download_data.sh
│ │ eval.py
│ │ LICENSE
│ │ README.md
│ │ requirements.txt
│ │ run_eval.sh
│ │ train.py
│ │ yolo.py
│ │
│ ├───cfg
│ │ csdarknet53-omega.cfg
│ │ darknet19_448_body.cfg
│ │ darknet53.cfg
│ │ yolo-fastest-xl.cfg
│ │ yolo-fastest.cfg
│ │ yolov2-tiny-voc.cfg
│ │ yolov2-tiny.cfg
│ │ yolov2-voc.cfg
│ │ yolov2.cfg
│ │ yolov3-spp.cfg
│ │ yolov3-tiny.cfg
│ │ yolov3.cfg
│ │ yolov4-csp.cfg
│ │ yolov4-csp_fixed.cfg
│ │ yolov4-sam-mish.cfg
│ │ yolov4-tiny.cfg
│ │ yolov4.cfg
│ │
│ ├───common
│ │ │ callbacks.py
│ │ │ data_utils.py
│ │ │ model_utils.py
│ │ │ utils.py
│ │ │ wbf_postprocess.py
│ │ │ yolo_postprocess_np.py
│ │ │
│ │ └───backbones
│ │ │ efficientnet.py
│ │ │ ghostnet.py
│ │ │ layers.py
│ │ │ mobilenet.py
│ │ │ mobilenet_v2.py
│ │ │ mobilenet_v3.py
│ │ │ mobilevit.py
│ │ │ peleenet.py
│ │ │ shufflenet.py
│ │ │ shufflenet_v2.py
│ │ │ squeezenet.py
│ │ │
│ │ └───imagenet_training
│ │ │ data_utils.py
│ │ │ heatmap_check.py
│ │ │ README.md
│ │ │ train_imagenet.py
│ │ │
│ │ └───imagenet_preprocess
│ │ imagenet_2012_label_map.txt
│ │ imagenet_2012_validation_synset_labels.txt
│ │ preprocess_imagenet_train_data.py
│ │ preprocess_imagenet_validation_data.py
│ │
│ ├───configs
│ │ coco2017_origin_classes.txt
│ │ coco_classes.txt
│ │ objects365_classes.txt
│ │ scaled-yolo4-csp_anchors.txt
│ │ tiny_yolo3_anchors.txt
│ │ VOC2012_person_test.txt
│ │ voc_classes.txt
│ │ yolo2-tiny-voc_anchors.txt
│ │ yolo2-tiny_anchors.txt
│ │ yolo2-voc_anchors.txt
│ │ yolo2_anchors.txt
│ │ yolo3_anchors.txt
│ │ yolo4_anchors.txt
│ │ yolo5_anchors.txt
│ │ yolo_fastest_anchors.txt
│ │
│ ├───inference
│ │ │ eval_inference.sh
│ │ │ README.md
│ │ │
│ │ ├───MNN
│ │ │ │ CMakeLists.txt
│ │ │ │ yoloDetection.cpp
│ │ │ │
│ │ │ └───configs
│ │ │ quantizeConfig.json
│ │ │
│ │ └───tflite
│ │ CMakeLists.txt
│ │ stb_image.h
│ │ stb_image_resize.h
│ │ yoloDetection.cpp
│ │ yoloDetection.h
│ │
│ ├───scaled_yolo4
│ │ └───models
│ │ layers.py
│ │ scaled_yolo4_csp_darknet.py
│ │
│ ├───tools
│ │ ├───dataset_converter
│ │ │ coco_annotation.py
│ │ │ coco_annotation_val.py
│ │ │ coco_dataset_to_voc.py
│ │ │ coco_to_pascal_voc.py
│ │ │ dataset_visualize.py
│ │ │ openimage_annotation.py
│ │ │ pascal_voc_to_coco.py
│ │ │ voc_annotation.py
│ │ │ voc_obj_similarity.py
│ │ │ voc_to_darknet.py
│ │ │ yolo_annotation.py
│ │ │
│ │ ├───evaluation
│ │ │ model_statistics.py
│ │ │ pycoco_eval.py
│ │ │ tide_eval.py
│ │ │ validate_yolo.py
│ │ │
│ │ ├───misc
│ │ │ augment_test.py
│ │ │ gif_create.py
│ │ │ kmeans.py
│ │ │ rotate_gridmask_test.py
│ │ │ tensor_compare.py
│ │ │ yuv_convert.py
│ │ │
│ │ └───model_converter
│ │ │ convert.py
│ │ │ custom_tflite_convert.py
│ │ │ keras_to_onnx.py
│ │ │ keras_to_tensorflow.py
│ │ │ keras_to_tensorflow_bk.py
│ │ │ post_train_quant_convert.py
│ │ │ tensorflow_to_coreml.py
│ │ │ tensorflow_to_rknn.py
│ │ │
│ │ └───ghostnet_convertor
│ │ convertor.py
│ │ ghostnet_pytorch.py
│ │
│ ├───tracking
│ │ │ mot_tracker.py
│ │ │ README.md
│ │ │ requirements.txt
│ │ │
│ │ ├───cpp_inference
│ │ │ │ README.md
│ │ │ │
│ │ │ └───yoloSort
│ │ │ │ CMakeLists.txt
│ │ │ │ Hungarian.cpp
│ │ │ │ Hungarian.h
│ │ │ │ KalmanTracker.cpp
│ │ │ │ KalmanTracker.h
│ │ │ │ Sort.cpp
│ │ │ │ Sort.h
│ │ │ │ yoloSort.cpp
│ │ │ │
│ │ │ └───kalman_demo
│ │ │ CMakeLists.txt
│ │ │ kalman_demo.cpp
│ │ │ kalman_demo.py
│ │ │ README.md
│ │ │
│ │ ├───eval
│ │ │ │ formatchecker.py
│ │ │ │ importers.py
│ │ │ │ pymot.py
│ │ │ │ README.md
│ │ │ │ rect.py
│ │ │ │ utilities.py
│ │ │ │
│ │ │ ├───example
│ │ │ │ groundtruth.json
│ │ │ │ hypotheses.json
│ │ │ │
│ │ │ └───tools
│ │ │ mot16_annotation.py
│ │ │
│ │ └───model
│ │ ├───deep_sort
│ │ │ detection.py
│ │ │ generate_detections.py
│ │ │ iou_matching.py
│ │ │ kalman_filter.py
│ │ │ linear_assignment.py
│ │ │ nn_matching.py
│ │ │ preprocessing.py
│ │ │ track.py
│ │ │ tracker.py
│ │ │ __init__.py
│ │ │
│ │ └───sort
│ │ sort.py
│ │
│ ├───weights
│ │ .gitignore
│ │
│ ├───yolo2
│ │ │ data.py
│ │ │ loss.py
│ │ │ model.py
│ │ │ postprocess.py
│ │ │ postprocess_np.py
│ │ │ __init__.py
│ │ │
│ │ └───models
│ │ layers.py
│ │ yolo2_darknet.py
│ │ yolo2_efficientnet.py
│ │ yolo2_mobilenet.py
│ │ yolo2_mobilenetv2.py
│ │ yolo2_mobilenetv3_large.py
│ │ yolo2_mobilenetv3_small.py
│ │ yolo2_xception.py
│ │
│ ├───yolo3
│ │ │ data.py
│ │ │ loss.py
│ │ │ model.py
│ │ │ postprocess.py
│ │ │ postprocess_np.py
│ │ │ __init__.py
│ │ │
│ │ └───models
│ │ layers.py
│ │ ultralite_layers.py
│ │ yolo3_darknet.py
│ │ yolo3_efficientnet.py
│ │ yolo3_ghostnet.py
│ │ yolo3_mobilenet.py
│ │ yolo3_mobilenetv2.py
│ │ yolo3_mobilenetv3_large.py
│ │ yolo3_mobilenetv3_small.py
│ │ yolo3_nano.py
│ │ yolo3_peleenet.py
│ │ yolo3_resnet50.py
│ │ yolo3_resnet50v2.py
│ │ yolo3_shufflenetv2.py
│ │ yolo3_vgg16.py
│ │ yolo3_xception.py
│ │
│ ├───yolo4
│ │ └───models
│ │ layers.py
│ │ yolo4_darknet.py
│ │ yolo4_efficientnet.py
│ │ yolo4_mobilenet.py
│ │ yolo4_mobilenetv2.py
│ │ yolo4_mobilenetv3_large.py
│ │ yolo4_mobilenetv3_small.py
│ │ yolo4_resnet50.py
│ │ yolo4_resnet50v2.py
│ │
│ └───yolo5
│ │ data.py
│ │ loss.py
│ │ model.py
│ │ postprocess.py
│ │ postprocess_np.py
│ │
│ └───models
│ layers.py
│ yolo5_darknet.py
│ yolo5_mobilenet.py
│ yolo5_mobilenetv2.py
│
├───data
│ .gitignore
│ demo_list.txt
├───float
│ yolov3.h5
└───quantized
quantized.h5
Prepare
yolov3_coco_416_tf2.xmodelfile for DPU 3136.
The file quantized.h5 is the quantized model generated by quantizer. It can be used to generate xmodel file.
Start the vitis-ai model compile docker and copy quantized.h5 and arch.json into it. Using the following command to compile .xmodel file.
conda activate vitis-ai-tensorflow2
vai_c_tensorflow2 -m /PATH/TO/quantized.h5 -a /PATH/TO/arch.json -o /OUTPUTPATH -n yolov3_coco_416_tf2
The yolov3_coco_416_tf2.prototxt can be found in existing file from zcu102 & zcu104 & kv260.
The YOLOv3 object detection model provides prediction of multiple classes in the label file.
person
bicycle
car
motorbike
aeroplane
bus
...
We need to convert the above label info to json file as requested by VVAS framework as bellow and named it ‘label.json’.
{
"model-name": "yolov3_coco_416_tf2",
"num-labels": 80,
"labels": [
{
"label": 0,
"name": "person",
"display_name": "person"
},
{
"label": 1,
"name": "bicycle",
"display_name": "bicycle"
},
{
"label": 2,
"name": "car",
"display_name": "car"
},
{
"label": 3,
"name": "motorbike",
"display_name": "motorbike"
},
{
"label": 4,
"name": "aeroplane",
"display_name": "aeroplane"
},
{
"label": 5,
"name": "bus",
"display_name": "bus"
},
{
"label": 6,
"name": "train",
"display_name": "train"
},
{
"label": 7,
"name": "truck",
"display_name": "truck"
},
{
"label": 8,
"name": "boat",
"display_name": "boat"
},
{
"label": 9,
"name": "traffic light",
"display_name": "traffic light"
},
{
"label": 10,
"name": "fire hydrant",
"display_name": "fire hydrant"
},
{
"label": 11,
"name": "stop sign",
"display_name": "stop sign"
},
{
"label": 12,
"name": "parking meter",
"display_name": "parking meter"
},
{
"label": 13,
"name": "bench",
"display_name": "bench"
},
{
"label": 14,
"name": "bird",
"display_name": "bird"
},
{
"label": 15,
"name": "cat",
"display_name": "cat"
},
{
"label": 16,
"name": "dog",
"display_name": "dog"
},
{
"label": 17,
"name": "horse",
"display_name": "horse"
},
{
"label": 18,
"name": "sheep",
"display_name": "sheep"
},
{
"label": 19,
"name": "cow",
"display_name": "cow"
},
{
"label": 20,
"name": "elephant",
"display_name": "elephant"
},
{
"label": 21,
"name": "bear",
"display_name": "bear"
},
{
"label": 22,
"name": "zebra",
"display_name": "zebra"
},
{
"label": 23,
"name": "giraffe",
"display_name": "giraffe"
},
{
"label": 24,
"name": "backpack",
"display_name": "backpack"
},
{
"label": 25,
"name": "umbrella",
"display_name": "umbrella"
},
{
"label": 26,
"name": "handbag",
"display_name": "handbag"
},
{
"label": 27,
"name": "tie",
"display_name": "tie"
},
{
"label": 28,
"name": "suitcase",
"display_name": "suitcase"
},
{
"label": 29,
"name": "frisbee",
"display_name": "frisbee"
},
{
"label": 30,
"name": "skis",
"display_name": "skis"
},
{
"label": 31,
"name": "snowboard",
"display_name": "snowboard"
},
{
"label": 32,
"name": "sports ball",
"display_name": "sports ball"
},
{
"label": 33,
"name": "kite",
"display_name": "kite"
},
{
"label": 34,
"name": "baseball bat",
"display_name": "baseball bat"
},
{
"label": 35,
"name": "baseball glove",
"display_name": "baseball glove"
},
{
"label": 36,
"name": "skateboard",
"display_name": "skateboard"
},
{
"label": 37,
"name": "surfboard",
"display_name": "surfboard"
},
{
"label": 38,
"name": "tennis racket",
"display_name": "tennis racket"
},
{
"label": 39,
"name": "bottle",
"display_name": "bottle"
},
{
"label": 40,
"name": "wine glass",
"display_name": "wine glass"
},
{
"label": 41,
"name": "cup",
"display_name": "cup"
},
{
"label": 42,
"name": "fork",
"display_name": "fork"
},
{
"label": 43,
"name": "knife",
"display_name": "knife"
},
{
"label": 44,
"name": "spoon",
"display_name": "spoon"
},
{
"label": 45,
"name": "bowl",
"display_name": "bowl"
},
{
"label": 46,
"name": "banana",
"display_name": "banana"
},
{
"label": 47,
"name": "apple",
"display_name": "apple"
},
{
"label": 48,
"name": "sandwich",
"display_name": "sandwich"
},
{
"label": 49,
"name": "orange",
"display_name": "orange"
},
{
"label": 50,
"name": "broccoli",
"display_name": "broccoli"
},
{
"label": 51,
"name": "carrot",
"display_name": "carrot"
},
{
"label": 52,
"name": "hot dog",
"display_name": "hot dog"
},
{
"label": 53,
"name": "pizza",
"display_name": "pizza"
},
{
"label": 54,
"name": "donut",
"display_name": "donut"
},
{
"label": 55,
"name": "cake",
"display_name": "cake"
},
{
"label": 56,
"name": "chair",
"display_name": "chair"
},
{
"label": 57,
"name": "sofa",
"display_name": "sofa"
},
{
"label": 58,
"name": "pottedplant",
"display_name": "pottedplant"
},
{
"label": 59,
"name": "bed",
"display_name": "bed"
},
{
"label": 60,
"name": "diningtable",
"display_name": "diningtable"
},
{
"label": 61,
"name": "toilet",
"display_name": "toilet"
},
{
"label": 62,
"name": "tvmonitor",
"display_name": "tvmonitor"
},
{
"label": 63,
"name": "laptop",
"display_name": "laptop"
},
{
"label": 64,
"name": "mouse",
"display_name": "mouse"
},
{
"label": 65,
"name": "remote",
"display_name": "remote"
},
{
"label": 66,
"name": "keyboard",
"display_name": "keyboard"
},
{
"label": 67,
"name": "cell phone",
"display_name": "cell phone"
},
{
"label": 68,
"name": "microwave",
"display_name": "microwave"
},
{
"label": 69,
"name": "oven",
"display_name": "oven"
},
{
"label": 70,
"name": "toaster",
"display_name": "toaster"
},
{
"label": 71,
"name": "sink",
"display_name": "sink"
},
{
"label": 72,
"name": "refrigerator",
"display_name": "refrigerator"
},
{
"label": 73,
"name": "book",
"display_name": "book"
},
{
"label": 74,
"name": "clock",
"display_name": "clock"
},
{
"label": 75,
"name": "vase",
"display_name": "vase"
},
{
"label": 76,
"name": "scissors",
"display_name": "scissors"
},
{
"label": 77,
"name": "teddy bear",
"display_name": "teddy bear"
},
{
"label": 78,
"name": "hair drier",
"display_name": "hair drier"
},
{
"label": 79,
"name": "toothbrush",
"display_name": "toothbrush"
}
]
}
Put the prototxt file together with xmodel and label.json file to /opt/xilinx/kv260-smartcam/share/vitis_ai_library/models/yolov3_coco_416_tf2/.
opt
└── xilinx
└── kv260-smartcam
└── share
└── vitis_ai_library
└── models
└── yolov3_coco_416_tf2
├── yolov3_coco_416_tf2.prototxt
├── yolov3_coco_416_tf2.xmodel
└── label.json
Create configuration files for yolov3_coco_416_tf2.
preprocess.json
The mean and scale of B, G, R channel is taken from the prototxt of the model.
model { kernel { mean: 0.0 mean: 0.0 mean: 0.0 scale: 0.00390625 scale: 0.00390625 scale: 0.00390625 } model_type : YOLOv3 yolo_v3_param { num_classes: 80 anchorCnt: 3 layer_name: "58" layer_name: "66" layer_name: "74" conf_threshold: 0.3 nms_threshold: 0.45 biases: 10 biases: 13 biases: 16 biases: 30 biases: 33 biases: 23 biases: 30 biases: 61 biases: 62 biases: 45 biases: 59 biases: 119 biases: 116 biases: 90 biases: 156 biases: 198 biases: 373 biases: 326 test_mAP: false } }Using the following command to get the fixpos.
xdputil xmodel ./yolov3_coco_416_tf2.xmodel -l ... "input_tensor":[ { "index":0, "name":"quant_image_input", "shape":[ 1, 416, 416, 3 ], "fixpos":6 } ], ...
The configured scale = prototxt_scale * (2 ^ fixpos) = 0.00390625 * (2^6) = 0.25.
{ "xclbin-location":"/lib/firmware/xilinx/kv260-smartcam/kv260-smartcam.xclbin", "vvas-library-repo": "/opt/xilinx/kv260-smartcam/lib", "element-mode": "transform", "kernels": [ { "kernel-name": "pp_pipeline_accel:{pp_pipeline_accel_1}", "library-name": "libvvas_xpp.so", "config": { "debug_level" : 1, "mean_r": 0, "mean_g": 0, "mean_b": 0, "scale_r": 0.25, "scale_g": 0.25, "scale_b": 0.25 } } ] }
aiinference.json:
{ "xclbin-location":"/lib/firmware/xilinx/kv260-smartcam/kv260-smartcam.xclbin", "vvas-library-repo": "/usr/lib/aarch64-linux-gnu/", "element-mode":"inplace", "kernels" :[ { "library-name":"libvvas_xdpuinfer.so", "config": { "model-name" : "yolov3_coco_416_tf2", "model-class" : "YOLOV3", "model-path" : "/opt/xilinx/kv260-smartcam/share/vitis_ai_library/models", "run_time_model" : false, "need_preprocess" : false, "performance_test" : false, "debug_level" : 0 } } ] }
drawresult.json
Here we pick up 4 classes to be shown: car, person, truck, bicycle; and customize the color for each class as bellow.
{ "xclbin-location":"/usr/lib/dpu.xclbin", "vvas-library-repo": "/opt/xilinx/kv260-smartcam/lib", "element-mode":"inplace", "kernels" :[ { "library-name":"libvvas_airender.so", "config": { "fps_interval" : 10, "font_size" : 2, "font" : 3, "thickness" : 2, "debug_level" : 0, "label_color" : { "blue" : 0, "green" : 0, "red" : 255 }, "label_filter" : [ "class", "probability" ], "classes" : [ { "name" : "car", "blue" : 255, "green" : 0, "red" : 0 }, { "name" : "person", "blue" : 0, "green" : 255, "red" : 0 }, { "name" : "truck", "blue" : 128, "green" : 255, "red" : 0 }, { "name" : "bicycle", "blue" : 0, "green" : 0, "red" : 255 }] } } ] }
Next Steps¶
Go back to the KV260 SOM Smart camera design start page
References¶
Vitis AI User Guide UG1414
License¶
Licensed under the Apache License, Version 2.0 (the “License”); you may not use this file except in compliance with the License.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
Copyright© 2022 Xilinx