root@c2e4ebaf0adf:/home/zhutao/horizon_j6/model# hb_compile --config bev_fushion.yaml 2025-04-21 07:14:13,647 INFO log will be stored in /home/zhutao/horizon_j6/model/hb_compile.log 2025-04-21 07:14:13,647 INFO Start hb_compile... 2025-04-21 07:14:13,647 INFO hbdk version: 4.1.17 2025-04-21 07:14:13,647 INFO hmct version: 2.1.9 2025-04-21 07:14:13,647 INFO hb_compile version: 3.3.11 2025-04-21 07:14:13,647 INFO Start verifying yaml 2025-04-21 07:14:13,849 INFO Using onnx model file: /home/zhutao/horizon_j6/model/bev_fushion/onnx/bev_v1.onnx 2025-04-21 07:14:13,849 INFO Model has 3 inputs according to model file 2025-04-21 07:14:13,849 INFO Input shape [40000, 4] has length: 2, make sure it is a featuremap input 2025-04-21 07:14:13,849 WARNING Input shape is not four-dimensional, input_layout_train should be featuremap 2025-04-21 07:14:13,849 WARNING norm_type parameter is deprecated and will be determined by the configuration of mean/scale/std parameter 2025-04-21 07:14:13,850 WARNING Input node img's input_source not set, it will be set to None by default 2025-04-21 07:14:13,850 WARNING Input node feature's input_source not set, it will be set to None by default 2025-04-21 07:14:13,850 WARNING Input node onnx::Gather_3's input_source not set, it will be set to None by default 2025-04-21 07:14:13,850 INFO Parameter compile_mode is set to latency, balance_factor will set to 100. 2025-04-21 07:14:13,850 INFO End verifying yaml 2025-04-21 07:14:13,850 INFO Processing calibration set data for input[img] 2025-04-21 07:14:13,851 INFO ******************************************* 2025-04-21 07:14:13,851 INFO First calibration picture name: /home/zhutao/horizon_j6/code/bevfusion-swmotion/export/saved_inputs/img/frame_000000_img.npy 2025-04-21 07:14:13,872 INFO First calibration picture md5: 216191ef338679bc86a3427d9ea71855 2025-04-21 07:14:13,872 INFO ******************************************* 2025-04-21 07:14:14,618 INFO Finished. data num: 81 2025-04-21 07:14:14,618 INFO Processing calibration set data for input[feature] 2025-04-21 07:14:14,619 INFO ******************************************* 2025-04-21 07:14:14,619 INFO First calibration picture name: /home/zhutao/horizon_j6/code/bevfusion-swmotion/export/saved_inputs/feature/frame_000000_feature.npy 2025-04-21 07:14:14,678 INFO First calibration picture md5: 4cedbd9c7e1cec1def342d4611819d17 2025-04-21 07:14:14,678 INFO ******************************************* 2025-04-21 07:14:17,707 INFO Finished. data num: 81 2025-04-21 07:14:17,707 INFO Processing calibration set data for input[onnx::Gather_3] 2025-04-21 07:14:17,707 INFO ******************************************* 2025-04-21 07:14:17,707 INFO First calibration picture name: /home/zhutao/horizon_j6/code/bevfusion-swmotion/export/saved_inputs/coors/frame_000000_gather.npy 2025-04-21 07:14:17,709 INFO First calibration picture md5: ae7ccb6b5f7538e438f64d9fa07912f7 2025-04-21 07:14:17,709 INFO ******************************************* 2025-04-21 07:14:18,471 INFO Finished. data num: 81 2025-04-21 07:14:19,085 INFO Start to Horizon NN Model Convert. 2025-04-21 07:14:19,200 INFO Loading horizon_nn debug methods: set(). 2025-04-21 07:14:19,201 INFO The specified model compilation architecture: nash. 2025-04-21 07:14:19,201 INFO The specified model compilation optimization parameters: []. 2025-04-21 07:14:19,211 INFO Start to prepare the onnx model. 2025-04-21 07:14:19,231 INFO Input ONNX Model Information: ONNX IR version: 8 Opset version: ['ai.onnx v16', 'horizon v1'] Producer: pytorch v1.12.1 Domain: None Version: None Graph input: img: shape=[5, 3, 256, 704], dtype=FLOAT32 feature: shape=[40000, 10, 1, 20], dtype=FLOAT32 ...Gather_3: shape=[40000, 4], dtype=INT32 Graph output: reg_0: shape=[1, 2, 120, 180], dtype=FLOAT32 height_0: shape=[1, 1, 120, 180], dtype=FLOAT32 dim_0: shape=[1, 3, 120, 180], dtype=FLOAT32 rot_0: shape=[1, 2, 120, 180], dtype=FLOAT32 vel_0: shape=[1, 2, 120, 180], dtype=FLOAT32 hm_0: shape=[1, 1, 120, 180], dtype=FLOAT32 reg_1: shape=[1, 2, 120, 180], dtype=FLOAT32 height_1: shape=[1, 1, 120, 180], dtype=FLOAT32 dim_1: shape=[1, 3, 120, 180], dtype=FLOAT32 rot_1: shape=[1, 2, 120, 180], dtype=FLOAT32 vel_1: shape=[1, 2, 120, 180], dtype=FLOAT32 hm_1: shape=[1, 1, 120, 180], dtype=FLOAT32 reg_2: shape=[1, 2, 120, 180], dtype=FLOAT32 height_2: shape=[1, 1, 120, 180], dtype=FLOAT32 dim_2: shape=[1, 3, 120, 180], dtype=FLOAT32 rot_2: shape=[1, 2, 120, 180], dtype=FLOAT32 vel_2: shape=[1, 2, 120, 180], dtype=FLOAT32 hm_2: shape=[1, 1, 120, 180], dtype=FLOAT32 reg_3: shape=[1, 2, 120, 180], dtype=FLOAT32 height_3: shape=[1, 1, 120, 180], dtype=FLOAT32 dim_3: shape=[1, 3, 120, 180], dtype=FLOAT32 rot_3: shape=[1, 2, 120, 180], dtype=FLOAT32 vel_3: shape=[1, 2, 120, 180], dtype=FLOAT32 hm_3: shape=[1, 2, 120, 180], dtype=FLOAT32 2025-04-21 07:14:19,714 INFO The original model's opset version is 16, try converting to opset 19. 2025-04-21 07:14:20,211 INFO End to prepare the onnx model. 2025-04-21 07:14:20,458 INFO Saving model to: /home/zhutao/horizon_j6/model/bev_fushion/hbm_31/bev_fusion_j6_original_float_model.onnx. 2025-04-21 07:14:20,458 INFO Start to optimize the onnx model. 2025-04-21 07:14:23,008 INFO End to optimize the onnx model. 2025-04-21 07:14:23,230 INFO Saving model to: /home/zhutao/horizon_j6/model/bev_fushion/hbm_31/bev_fusion_j6_optimized_float_model.onnx. 2025-04-21 07:14:23,230 INFO The weight calibration parameters: bias_correction: {'num_sample': 1, 'metric': 'cosine-similarity'} The layerwise search parameters: metric: cosine-similarity The input of all ScatterND nodes are set to : float16 2025-04-21 07:14:23,230 INFO Start to calibrate the model. Warning: The input0 of Node(name:ScatterND_19, type:ScatterND) does not support data type: float16 Warning: The input2 of Node(name:ScatterND_19, type:ScatterND) does not support data type: float16 Warning: The input2 of Node(name:ScatterND_19, type:ScatterND) does not support data type: int16 Warning: The input2 of Node(name:ScatterND_19, type:ScatterND) does not support data type: int8 2025-04-21 07:14:23,691 INFO There are 81 samples in the dataset. 2025-04-21 07:14:27,669 INFO Provided calibration data md5: 742407a3e0b097ab1ca540c028169194 2025-04-21 07:14:27,670 INFO Run calibration model with layerwise search method. 2025-04-21 07:14:28,021 INFO Calibration using batch 1 calibration in progress: 100%|██████████████████████████████████████████████████████████████████████| 81/81 [09:27<00:00, 7.01s/it] 2025-04-21 07:24:11,018 INFO Calculate all sensitivity nodes with max calibration method using cosine-similarity metric. 2025-04-21 07:24:11,093 INFO Start calculating activation sensitivity... 2025-04-21 07:24:11,822 INFO There are 1 samples in the dataset. Progress: 0%| | 0/96 [00:00