Onnx model quantization error when converting to hef

General
1

Hello. I’m having a difficult problem. I’m getting an error when converting my onnx model to hef. The conversion code is:

import hailo_sdk_client
from hailo_sdk_client import ClientRunner
import numpy as np
import os
import logging
import traceback

# 设置日志级别为 DEBUG 以获取更多信息
logging.basicConfig(
    level=logging.DEBUG,
    format='%(asctime)s - %(levelname)s - %(message)s',
    handlers=[
        logging.FileHandler('hailo_quantization.log'),
        logging.StreamHandler()
    ]
)

# --- 1. Hailo 硬件架构 ---
chosen_hw_arch = "hailo8"

# --- 2. 模型路径和文件名 ---
onnx_model_name = "head_singleframe"
onnx_path = "/home/ubuntu/ycy/head_singleframe.onnx"
output_dir = "/home/ubuntu/ycy/onnx_models/hailo_outputs"
hailo_model_har_path = os.path.join(output_dir, f"{onnx_model_name}_hailo_model.har")
hailo_quantized_har_path = os.path.join(output_dir, f"{onnx_model_name}_hailo_quantized_model.har")
hailo_model_hef_path = os.path.join(output_dir, f"{onnx_model_name}.hef")

# --- 3. 确保输出目录存在 ---
os.makedirs(output_dir, exist_ok=True)

# --- 4. 初始化 Hailo ClientRunner ---
try:
    runner = ClientRunner(hw_arch=chosen_hw_arch)
    print(f"✅ Initialized ClientRunner with hardware architecture: {chosen_hw_arch}")
except Exception as e:
    print(f"❌ Failed to initialize ClientRunner: {e}")
    exit(1)

# --- 5. 验证 ONNX 模型文件存在 ---
if not os.path.exists(onnx_path):
    print(f"❌ ONNX model file not found: {onnx_path}")
    exit(1)
print(f"✅ ONNX model file found: {onnx_path}")

# --- 6. 指定 ONNX 模型的多个输入节点 ---
net_input_shapes = {
    "inst_dfg0": (1, 900, 512),
    "anchor": (1, 900, 11),
    "anchor_embed": (1, 900, 256),
    "time_interval": (1, 1),  # 与 ONNX 模型输入形状 [1, 1] 一致
}

# --- 7. 转换 ONNX 模型为 HAR ---
try:
    hn, npz = runner.translate_onnx_model(
        model=onnx_path,
        net_name=onnx_model_name,
        net_input_shapes=net_input_shapes,
        start_node_names=list(net_input_shapes.keys())
    )
    print(f"✅ Successfully translated ONNX model to HAR")
except Exception as e:
    print(f"❌ Failed to translate ONNX model to HAR: {e}")
    traceback.print_exc()
    exit(1)

# --- 8. 保存 HAR 模型 ---
try:
    runner.save_har(hailo_model_har_path)
    print(f"✅ HAR model saved to: {hailo_model_har_path}")
except Exception as e:
    print(f"❌ Failed to save HAR model: {e}")
    exit(1)

# --- 9. 校准数据集准备 ---
def generate_calibration_dataset(num_samples: int = 100):
    calib_data_path = "/home/ubuntu/ycy/calib_data/calib_data.npz"
    
    hailo_input_names_mapping = {
        'inst_dfg0': 'head_singleframe/input_layer1',
        'anchor': 'head_singleframe/input_layer2',
        'anchor_embed': 'head_singleframe/input_layer3',
        'time_interval': 'head_singleframe/input_layer4',
    }
    
    if os.path.exists(calib_data_path):
        print(f"✅ Loading real calibration data from {calib_data_path}")
        data = np.load(calib_data_path, allow_pickle=True)
        calib_data = data["calib_data"][:num_samples]

        inst_dfg0_array = np.stack([np.nan_to_num(s["inst_dfg0"], nan=0.0, posinf=0.0, neginf=0.0) / 10.0 for s in calib_data]).astype(np.float32)
        anchor_array = np.stack([np.nan_to_num(s["anchor"], nan=0.0, posinf=0.0, neginf=0.0) / 100.0 for s in calib_data]).astype(np.float32)
        anchor_embed_array = np.stack([np.nan_to_num(s["anchor_embed"], nan=0.0, posinf=0.0, neginf=0.0) / 10.0 for s in calib_data]).astype(np.float32)
        
        # 修正 time_interval 的形状为 (num_samples, 1, 1, 1)
        time_interval_array = np.stack([np.expand_dims(s["time_interval"], axis=-1) for s in calib_data]).astype(np.float32)
        
        print(f"✅ Loaded {num_samples} real calibration samples.")

        # 返回 NumPy 字典
        calib_dataset = {
            hailo_input_names_mapping['inst_dfg0']: inst_dfg0_array,
            hailo_input_names_mapping['anchor']: anchor_array,
            hailo_input_names_mapping['anchor_embed']: anchor_embed_array,
            hailo_input_names_mapping['time_interval']: time_interval_array,
        }
        
        print(f"✅ Generated calibration dataset shapes:")
        for name, data in calib_dataset.items():
            print(f"  - {name}: {data.shape}")
        
        return calib_dataset

    else:
        print(f"⚠️ Calibration data not found at {calib_data_path}, generating random data")
        
        random_data_dict = {
            hailo_input_names_mapping['inst_dfg0']: np.random.randn(num_samples, 1, 900, 512).astype(np.float32) / 10.0,
            hailo_input_names_mapping['anchor']: np.random.randn(num_samples, 1, 900, 11).astype(np.float32) / 100.0,
            hailo_input_names_mapping['anchor_embed']: np.random.randn(num_samples, 1, 900, 256).astype(np.float32) / 10.0,
            hailo_input_names_mapping['time_interval']: np.zeros((num_samples, 1, 1, 1), dtype=np.float32),
        }

        print(f"✅ Generated random dataset shapes:")
        for name, data in random_data_dict.items():
            print(f"  - {name}: {data.shape}")
        
        return random_data_dict

calib_dataset = generate_calibration_dataset(num_samples=100)
print(f"✅ Generated calibration dataset: {type(calib_dataset)}")

# --- 10. 量化 HAR 模型 ---
runner = ClientRunner(har=hailo_model_har_path)
alls_lines = [
    'model_optimization_flavor(optimization_level=0, compression_level=0)'
    # 使用最简优化配置，避免层规范化分解问题
]
print("--- Begin quantization ---")
runner.load_model_script('\n'.join(alls_lines))
try:
    runner.optimize(calib_dataset)
    runner.save_har(hailo_quantized_har_path)
    print(f"✅ Quantized HAR model saved to: {hailo_quantized_har_path}")
except Exception as e:
    print(f"❌ Failed to quantize HAR model: {e}")
    traceback.print_exc()
    exit(1)

# --- 11. 导出 HEF 模型 ---
runner = ClientRunner(har=hailo_quantized_har_path)
try:
    compiled_hef = runner.compile()
    with open(hailo_model_hef_path, "wb") as f:
        f.write(compiled_hef)
        print(f"✅ HEF model exported to: {hailo_model_hef_path}")
except Exception as e:
    print(f"❌ Failed to compile HEF model: {e}")
    exit(1)

The error message is as follows:

[warning] Cannot use graphviz, so no visualizations will be created
✅ Initialized ClientRunner with hardware architecture: hailo8
✅ ONNX model file found: /home/ubuntu/ycy/head_singleframe.onnx
[info] Translation started on ONNX model head_singleframe
[info] Restored ONNX model head_singleframe (completion time: 00:00:00.02)
[info] Extracted ONNXRuntime meta-data for Hailo model (completion time: 00:00:00.09)
[info] Start nodes mapped from original model: 'inst_dfg0': 'head_singleframe/input_layer1', 'anchor': 'head_singleframe/input_layer2', 'anchor_embed': 'head_singleframe/input_layer3', 'time_interval': 'head_singleframe/input_layer4'.
[info] End nodes mapped from original model: '/layers.3/cls_layers/cls_layers.6/Add', '/layers.3/Concat', '/layers.3/quality_layers/quality_layers.6/Add'.
[info] Translation completed on ONNX model head_singleframe (completion time: 00:00:00.24)
✅ Successfully translated ONNX model to HAR
[info] Saved HAR to: /home/ubuntu/ycy/onnx_models/hailo_outputs/head_singleframe_hailo_model.har
✅ HAR model saved to: /home/ubuntu/ycy/onnx_models/hailo_outputs/head_singleframe_hailo_model.har
✅ Loading real calibration data from /home/ubuntu/ycy/calib_data/calib_data.npz
✅ Loaded 100 real calibration samples.
✅ Generated calibration dataset shapes:
  - head_singleframe/input_layer1: (64, 1, 900, 512)
  - head_singleframe/input_layer2: (64, 1, 900, 11)
  - head_singleframe/input_layer3: (64, 1, 900, 256)
  - head_singleframe/input_layer4: (64, 1, 1, 1)
✅ Generated calibration dataset: <class 'dict'>
--- Begin quantization ---
[info] Loading model script commands to head_singleframe from string
[info] Starting Model Optimization
[warning] Running model optimization with zero level of optimization is not recommended for production use and might lead to suboptimal accuracy results
[info] Model received quantization params from the hn
[info] Starting Mixed Precision
[info] Mixed Precision is done (completion time is 00:00:00.08)
[info] Starting Layer Norm Decomposition
[info] Using dataset with 64 entries for calibration
Calibration: 100%|███████████████████████████████████████████████████████████████████████████████████████████████| 64/64 [00:36<00:00,  1.75entries/s]
[info] Layer Norm Decomposition is done (completion time is 00:00:39.98)
[info] Starting Stats Collector
[info] Using dataset with 64 entries for calibration
Calibration: 100%|███████████████████████████████████████████████████████████████████████████████████████████████| 64/64 [00:25<00:00,  2.53entries/s]
[info] Stats Collector is done (completion time is 00:00:26.45)
[info] Starting Fix zp_comp Encoding
[info] Fix zp_comp Encoding is done (completion time is 00:00:00.00)
[info] matmul_equalization skipped
[info] activation fitting started for head_singleframe/activation1/act_op
/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/pwlf/pwlf.py:1248: RuntimeWarning: invalid value encountered in divide
  self.slopes = np.divide(
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization1/act_op
[info] No shifts available for layer head_singleframe/conv3/conv_op, using max shift instead. delta=0.9112
[info] No shifts available for layer head_singleframe/conv3/conv_op, using max shift instead. delta=0.4556
[info] No shifts available for layer head_singleframe/conv3/conv_op, using max shift instead. delta=0.4556
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization2/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization3/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization4/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization5/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization6/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization7/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization8/act_op
[info] No shifts available for layer head_singleframe/conv14/conv_op, using max shift instead. delta=0.3263
[info] No shifts available for layer head_singleframe/conv14/conv_op, using max shift instead. delta=0.1632
❌ Failed to quantize HAR model: the exponent [[nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]]  is not in range [ 7 22]
Traceback (most recent call last):
  File "/home/ubuntu/ycy/projects/mmdet3d_plugin/tools/test/hailoe.py", line 1025, in <module>
    runner.optimize(calib_dataset)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_common/states/states.py", line 16, in wrapped_func
    return func(self, *args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_client/runner/client_runner.py", line 2093, in optimize
    self._optimize(calib_data, data_type=data_type, work_dir=work_dir)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_common/states/states.py", line 16, in wrapped_func
    return func(self, *args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_client/runner/client_runner.py", line 1935, in _optimize
    self._sdk_backend.full_quantization(calib_data, data_type=data_type, work_dir=work_dir)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_client/sdk_backend/sdk_backend.py", line 1045, in full_quantization
    self._full_acceleras_run(self.calibration_data, data_type)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_client/sdk_backend/sdk_backend.py", line 1229, in _full_acceleras_run
    optimization_flow.run()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/tools/orchestator.py", line 306, in wrapper
    return func(self, *args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/flows/optimization_flow.py", line 316, in run
    step_func()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/tools/orchestator.py", line 250, in wrapped
    result = method(*args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/tools/subprocess_wrapper.py", line 122, in parent_wrapper
    func(self, *args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/flows/optimization_flow.py", line 336, in step1
    self._update_quantize_data()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/tools/orchestator.py", line 250, in wrapped
    result = method(*args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/flows/optimization_flow.py", line 324, in _update_quantize_data
    self.set_quant_params(self._model.export_hw_params())
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/model/hailo_model/hailo_model.py", line 467, in export_hw_params
    return self._export_npz(mode=NpzExportMode.QNPZ)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/model/hailo_model/hailo_model.py", line 521, in _export_npz
    params_exported = acceleras_layer.export_hw_params()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/hailo_layers/base_hailo_layer.py", line 440, in export_hw_params
    return self.export_qnpz(convert=True)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/hailo_layers/base_hailo_layer.py", line 797, in export_qnpz
    export_params = self._export_quant_internal()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/hailo_layers/base_hailo_layer.py", line 710, in _export_quant_internal
    export_params.update(op.export_quant())
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/atomic_ops/base_atomic_op.py", line 493, in export_quant
    export_params.update(self.export_quant_weights())
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/atomic_ops/activation_op.py", line 1565, in export_quant_weights
    self.check_exp_range()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/atomic_ops/activation_op.py", line 1059, in check_exp_range
    raise AccelerasNegativeSlopesError(exponents.numpy(), valid_range)
hailo_model_optimization.acceleras.utils.acceleras_exceptions.AccelerasNegativeSlopesError: the exponent [[nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]]  is not in range [ 7 22]

Can you help me identify the problem?

2

runner.optimize_full_precision(calib_dataset)

Try the above line first before runner.optimize(). IIRC the exponent fixing algorithm is run in optimize_full_precision().

4

Thank you for your reply. I tried the method you suggested, but still encountered the same error. The specific code and error message are below. Could you help me take a look?Thank you very much.
Code:

import hailo_sdk_client
from hailo_sdk_client import ClientRunner
import numpy as np
import os
import logging
import traceback

# 设置日志级别为 DEBUG 以获取更多信息
logging.basicConfig(
    level=logging.DEBUG,
    format='%(asctime)s - %(levelname)s - %(message)s',
    handlers=[
        logging.FileHandler('hailo_quantization.log'),
        logging.StreamHandler()
    ]
)

# --- 1. Hailo 硬件架构 ---
chosen_hw_arch = "hailo8"

# --- 2. 模型路径和文件名 ---
onnx_model_name = "head_singleframe"
onnx_path = "/home/ubuntu/ycy/head_singleframe.onnx"
output_dir = "/home/ubuntu/ycy/onnx_models/hailo_outputs"
hailo_model_har_path = os.path.join(output_dir, f"{onnx_model_name}_hailo_model.har")
hailo_quantized_har_path = os.path.join(output_dir, f"{onnx_model_name}_hailo_quantized_model.har")
hailo_model_hef_path = os.path.join(output_dir, f"{onnx_model_name}.hef")

# --- 3. 确保输出目录存在 ---
os.makedirs(output_dir, exist_ok=True)

# --- 4. 初始化 Hailo ClientRunner ---
try:
    runner = ClientRunner(hw_arch=chosen_hw_arch)
    print(f"✅ Initialized ClientRunner with hardware architecture: {chosen_hw_arch}")
except Exception as e:
    print(f"❌ Failed to initialize ClientRunner: {e}")
    exit(1)

# --- 5. 验证 ONNX 模型文件存在 ---
if not os.path.exists(onnx_path):
    print(f"❌ ONNX model file not found: {onnx_path}")
    exit(1)
print(f"✅ ONNX model file found: {onnx_path}")

# --- 6. 指定 ONNX 模型的多个输入节点 ---
net_input_shapes = {
    "inst_dfg0": (1, 900, 512),
    "anchor": (1, 900, 11),
    "anchor_embed": (1, 900, 256),
    "time_interval": (1, 1),  # 与 ONNX 模型输入形状 [1, 1] 一致
}

# --- 7. 转换 ONNX 模型为 HAR ---
try:
    hn, npz = runner.translate_onnx_model(
        model=onnx_path,
        net_name=onnx_model_name,
        net_input_shapes=net_input_shapes,
        start_node_names=list(net_input_shapes.keys())
    )
    print(f"✅ Successfully translated ONNX model to HAR")
except Exception as e:
    print(f"❌ Failed to translate ONNX model to HAR: {e}")
    traceback.print_exc()
    exit(1)

# --- 8. 保存 HAR 模型 ---
try:
    runner.save_har(hailo_model_har_path)
    print(f"✅ HAR model saved to: {hailo_model_har_path}")
except Exception as e:
    print(f"❌ Failed to save HAR model: {e}")
    exit(1)

# --- 9. 校准数据集准备 ---
def generate_calibration_dataset(num_samples: int = 100):
    calib_data_path = "/home/ubuntu/ycy/calib_data/calib_data1.npz"
    
    hailo_input_names_mapping = {
        'inst_dfg0': 'head_singleframe/input_layer1',
        'anchor': 'head_singleframe/input_layer2',
        'anchor_embed': 'head_singleframe/input_layer3',
        'time_interval': 'head_singleframe/input_layer4',
    }
    
    if os.path.exists(calib_data_path):
        print(f"✅ Loading real calibration data from {calib_data_path}")
        data = np.load(calib_data_path, allow_pickle=True)
        calib_data = data["calib_data"][:num_samples]

        inst_dfg0_array = np.stack([np.nan_to_num(s["inst_dfg0"], nan=0.0, posinf=0.0, neginf=0.0) / 10.0 for s in calib_data]).astype(np.float32)
        anchor_array = np.stack([np.nan_to_num(s["anchor"], nan=0.0, posinf=0.0, neginf=0.0) / 100.0 for s in calib_data]).astype(np.float32)
        anchor_embed_array = np.stack([np.nan_to_num(s["anchor_embed"], nan=0.0, posinf=0.0, neginf=0.0) / 10.0 for s in calib_data]).astype(np.float32)
        
        # 修正 time_interval 的形状为 (num_samples, 1, 1, 1)
        time_interval_array = np.stack([np.expand_dims(s["time_interval"], axis=-1) for s in calib_data]).astype(np.float32)
        
        print(f"✅ Loaded {num_samples} real calibration samples.")

        # 返回 NumPy 字典
        calib_dataset = {
            hailo_input_names_mapping['inst_dfg0']: inst_dfg0_array,
            hailo_input_names_mapping['anchor']: anchor_array,
            hailo_input_names_mapping['anchor_embed']: anchor_embed_array,
            hailo_input_names_mapping['time_interval']: time_interval_array,
        }
        
        print(f"✅ Generated calibration dataset shapes:")
        for name, data in calib_dataset.items():
            print(f"  - {name}: {data.shape}")
        
        return calib_dataset

    else:
        print(f"⚠️ Calibration data not found at {calib_data_path}, generating random data")
        
        random_data_dict = {
            hailo_input_names_mapping['inst_dfg0']: np.random.randn(num_samples, 1, 900, 512).astype(np.float32) / 10.0,
            hailo_input_names_mapping['anchor']: np.random.randn(num_samples, 1, 900, 11).astype(np.float32) / 100.0,
            hailo_input_names_mapping['anchor_embed']: np.random.randn(num_samples, 1, 900, 256).astype(np.float32) / 10.0,
            hailo_input_names_mapping['time_interval']: np.zeros((num_samples, 1, 1, 1), dtype=np.float32),
        }

        print(f"✅ Generated random dataset shapes:")
        for name, data in random_data_dict.items():
            print(f"  - {name}: {data.shape}")
        
        return random_data_dict

calib_dataset = generate_calibration_dataset(num_samples=100)
print(f"✅ Generated calibration dataset: {type(calib_dataset)}")

# --- 10. 量化 HAR 模型 ---
runner = ClientRunner(har=hailo_model_har_path)
alls_lines = [
    'model_optimization_flavor(optimization_level=0, compression_level=0)'
    # 使用最简优化配置，避免层规范化分解问题
]
print("--- Begin quantization ---")
runner.load_model_script('\n'.join(alls_lines))
try:
    runner.optimize_full_precision(calib_dataset)
    runner.optimize(calib_dataset)
    runner.save_har(hailo_quantized_har_path)
    print(f"✅ Quantized HAR model saved to: {hailo_quantized_har_path}")
except Exception as e:
    print(f"❌ Failed to quantize HAR model: {e}")
    traceback.print_exc()
    exit(1)

# --- 11. 导出 HEF 模型 ---
runner = ClientRunner(har=hailo_quantized_har_path)
try:
    compiled_hef = runner.compile()
    with open(hailo_model_hef_path, "wb") as f:
        f.write(compiled_hef)
        print(f"✅ HEF model exported to: {hailo_model_hef_path}")
except Exception as e:
    print(f"❌ Failed to compile HEF model: {e}")
    exit(1)

Error message ：

[warning] Cannot use graphviz, so no visualizations will be created
✅ Initialized ClientRunner with hardware architecture: hailo8
✅ ONNX model file found: /home/ubuntu/ycy/head_singleframe.onnx
[info] Translation started on ONNX model head_singleframe
[info] Restored ONNX model head_singleframe (completion time: 00:00:00.02)
[info] Extracted ONNXRuntime meta-data for Hailo model (completion time: 00:00:00.11)
[info] Start nodes mapped from original model: 'inst_dfg0': 'head_singleframe/input_layer1', 'anchor': 'head_singleframe/input_layer2', 'anchor_embed': 'head_singleframe/input_layer3', 'time_interval': 'head_singleframe/input_layer4'.
[info] End nodes mapped from original model: '/layers.3/cls_layers/cls_layers.6/Add', '/layers.3/Concat', '/layers.3/quality_layers/quality_layers.6/Add'.
[info] Translation completed on ONNX model head_singleframe (completion time: 00:00:00.26)
✅ Successfully translated ONNX model to HAR
[info] Saved HAR to: /home/ubuntu/ycy/Sparse4D-3.2/onnx_models/hailo_outputs/head_singleframe_hailo_model.har
✅ HAR model saved to: /home/ubuntu/ycy/Sparse4D-3.2/onnx_models/hailo_outputs/head_singleframe_hailo_model.har
✅ Loading real calibration data from /home/ubuntu/ycy/Sparse4D-3.2/calib_data/calib_data1.npz
✅ Loaded 100 real calibration samples.
✅ Generated calibration dataset shapes:
  - head_singleframe/input_layer1: (64, 1, 900, 512)
  - head_singleframe/input_layer2: (64, 1, 900, 11)
  - head_singleframe/input_layer3: (64, 1, 900, 256)
  - head_singleframe/input_layer4: (64, 1, 1, 1)
✅ Generated calibration dataset: <class 'dict'>
--- Begin quantization ---
[info] Loading model script commands to head_singleframe from string
[info] Starting Model Optimization
[warning] Running model optimization with zero level of optimization is not recommended for production use and might lead to suboptimal accuracy results
[info] Model received quantization params from the hn
[info] Starting Mixed Precision
[info] Mixed Precision is done (completion time is 00:00:00.08)
[info] Starting Layer Norm Decomposition
[info] Using dataset with 64 entries for calibration
Calibration: 100%|██████████████████████████████████████████████████████████████████████████████████████████████| 64/64 [00:37<00:00,  1.72entries/s]
[info] Layer Norm Decomposition is done (completion time is 00:00:40.58)
[info] Starting Stats Collector
[info] Using dataset with 64 entries for calibration
Calibration: 100%|██████████████████████████████████████████████████████████████████████████████████████████████| 64/64 [00:25<00:00,  2.53entries/s]
[info] Stats Collector is done (completion time is 00:00:26.40)
[info] Starting Fix zp_comp Encoding
[info] Fix zp_comp Encoding is done (completion time is 00:00:00.00)
[info] matmul_equalization skipped
[info] activation fitting started for head_singleframe/activation1/act_op
/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/pwlf/pwlf.py:1248: RuntimeWarning: invalid value encountered in divide
  self.slopes = np.divide(
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization1/act_op
[info] No shifts available for layer head_singleframe/conv3/conv_op, using max shift instead. delta=0.7076
[info] No shifts available for layer head_singleframe/conv3/conv_op, using max shift instead. delta=0.3538
[info] No shifts available for layer head_singleframe/conv3/conv_op, using max shift instead. delta=0.3538
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization2/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization3/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization4/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization5/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization6/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization7/act_op
[info] activation fitting started for head_singleframe/conv_var_inv_layer_normalization8/act_op
[info] No shifts available for layer head_singleframe/conv14/conv_op, using max shift instead. delta=0.0355
[info] No shifts available for layer head_singleframe/conv14/conv_op, using max shift instead. delta=0.0178
❌ Failed to quantize HAR model: the exponent [[nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]]  is not in range [ 7 22]
Traceback (most recent call last):
  File "/home/ubuntu/ycy/Sparse4D-3.2/projects/mmdet3d_plugin/tools/test/hailoe.py", line 1305, in <module>
    runner.optimize(calib_dataset)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_common/states/states.py", line 16, in wrapped_func
    return func(self, *args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_client/runner/client_runner.py", line 2093, in optimize
    self._optimize(calib_data, data_type=data_type, work_dir=work_dir)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_common/states/states.py", line 16, in wrapped_func
    return func(self, *args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_client/runner/client_runner.py", line 1935, in _optimize
    self._sdk_backend.full_quantization(calib_data, data_type=data_type, work_dir=work_dir)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_client/sdk_backend/sdk_backend.py", line 1045, in full_quantization
    self._full_acceleras_run(self.calibration_data, data_type)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_sdk_client/sdk_backend/sdk_backend.py", line 1229, in _full_acceleras_run
    optimization_flow.run()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/tools/orchestator.py", line 306, in wrapper
    return func(self, *args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/flows/optimization_flow.py", line 316, in run
    step_func()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/tools/orchestator.py", line 250, in wrapped
    result = method(*args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/tools/subprocess_wrapper.py", line 122, in parent_wrapper
    func(self, *args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/flows/optimization_flow.py", line 336, in step1
    self._update_quantize_data()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/tools/orchestator.py", line 250, in wrapped
    result = method(*args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/flows/optimization_flow.py", line 324, in _update_quantize_data
    self.set_quant_params(self._model.export_hw_params())
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/model/hailo_model/hailo_model.py", line 467, in export_hw_params
    return self._export_npz(mode=NpzExportMode.QNPZ)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/model/hailo_model/hailo_model.py", line 521, in _export_npz
    params_exported = acceleras_layer.export_hw_params()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/hailo_layers/base_hailo_layer.py", line 440, in export_hw_params
    return self.export_qnpz(convert=True)
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/hailo_layers/base_hailo_layer.py", line 797, in export_qnpz
    export_params = self._export_quant_internal()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/hailo_layers/base_hailo_layer.py", line 710, in _export_quant_internal
    export_params.update(op.export_quant())
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/atomic_ops/base_atomic_op.py", line 493, in export_quant
    export_params.update(self.export_quant_weights())
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/atomic_ops/activation_op.py", line 1565, in export_quant_weights
    self.check_exp_range()
  File "/home/ubuntu/miniconda3/envs/hailo/lib/python3.10/site-packages/hailo_model_optimization/acceleras/atomic_ops/activation_op.py", line 1059, in check_exp_range
    raise AccelerasNegativeSlopesError(exponents.numpy(), valid_range)
hailo_model_optimization.acceleras.utils.acceleras_exceptions.AccelerasNegativeSlopesError: the exponent [[nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]
 [nan]]  is not in range [ 7 22]
5

This is part of my calibration data.

Sample 62:
  - inst_dfg0: (1, 900, 512), dtype: float32
    Example data: [[[-1.6497484  -0.9341989  -0.6281898  ...  0.          0.
    0.        ]
  [-0.7132959  -0.39532393 -0.35687146 ...  0.          0.
    0.        ]
  [ 0.22623117 -0.48939723  0.31539616 ...  0.          0.
    0.        ]
  ...
  [ 0.35018504 -0.51488274  0.33371377 ...  0.          0.
    0.        ]
  [-0.10681286 -0.5627486   0.10305168 ...  0.          0.
    0.        ]
  [-0.01737372 -0.73980147 -0.08361288 ...  0.          0.
    0.        ]]]
  - anchor: (1, 900, 11), dtype: float32
    Example data: [[[ -3.0179338    5.499397    -0.8637523  ...   0.           0.
     0.        ]
  [  3.8772693  -27.766438    -2.0446386  ...   0.           0.
     0.        ]
  [-23.310377    28.47417      0.11812632 ...   0.           0.
     0.        ]
  ...
  [ -9.503412    47.198254     1.0450996  ...   0.           0.
     0.        ]
  [ 17.598536   -37.61403     -1.8509827  ...   0.           0.
     0.        ]
  [-11.910927    29.049755     0.20627801 ...   0.           0.
     0.        ]]]
  - anchor_embed: (1, 900, 256), dtype: float32
    Example data: [[[-0.7687792   0.32845938 -0.0335262  ... -0.71368104  1.6489092
    0.15629242]
  [-0.25253317 -0.74378735 -0.74378735 ... -0.71368104  1.6489092
    0.15629242]
  [-0.758829   -0.758829    1.1654099  ... -0.71368104  1.6489092
    0.15629242]
  ...
  [-0.7419545   1.0029572  -0.72872317 ... -0.71368104  1.6489092
    0.15629242]
  [-0.23271966 -0.77823156 -0.77823156 ... -0.71368086  1.6489089
    0.15629196]
  [-0.7472431   0.17190394 -0.7472431  ... -0.71368086  1.6489089
    0.15629196]]]
  - time_interval: (1, 1), dtype: float32
    Example data: [[0.5]]

Sample 63:
  - inst_dfg0: (1, 900, 512), dtype: float32
    Example data: [[[-0.48714238 -0.48403412 -0.01391043 ...  0.          0.
    0.        ]
  [-0.5837286  -0.89062554 -0.48944795 ...  0.          0.
    0.        ]
  [-0.00468572 -0.48425508  0.26780668 ...  0.          0.
    0.        ]
  ...
  [-0.0675825  -0.9088189  -0.34612507 ...  0.          0.
    0.        ]
  [-0.15163456 -0.5922837   0.07407173 ...  0.          0.
    0.        ]
  [-0.24981606 -1.0280172  -0.5579036  ...  0.          0.
    0.        ]]]
  - anchor: (1, 900, 11), dtype: float32
    Example data: [[[ -3.0179338    5.499397    -0.8637523  ...   0.           0.
     0.        ]
  [  3.8772693  -27.766438    -2.0446386  ...   0.           0.
     0.        ]
  [-23.310377    28.47417      0.11812632 ...   0.           0.
     0.        ]
  ...
  [ -9.503412    47.198254     1.0450996  ...   0.           0.
     0.        ]
  [ 17.598536   -37.61403     -1.8509827  ...   0.           0.
     0.        ]
  [-11.910927    29.049755     0.20627801 ...   0.           0.
     0.        ]]]
  - anchor_embed: (1, 900, 256), dtype: float32
    Example data: [[[ 1.625965    1.5313278   2.081998   ... -0.78937036 -0.27210197
   -0.78937036]
  [-0.82382953 -0.60970265 -0.6268218  ... -0.78937036 -0.27210197
   -0.78937036]
  [ 1.0799159   1.3293153   3.4304981  ... -0.78937036 -0.27210197
   -0.78937036]
  ...
  [ 0.63368434  1.7816951   2.0409667  ... -0.78937036 -0.27210197
   -0.78937036]
  [-0.746233    0.13802315 -0.52640516 ... -0.7893705  -0.27210188
   -0.7893705 ]
  [ 1.0722997   1.3785032   2.4316168  ... -0.7893705  -0.27210188
   -0.7893705 ]]]
  - time_interval: (1, 1), dtype: float32
    Example data: [[0.5]]

Sample 64:
  - inst_dfg0: (1, 900, 512), dtype: float32
    Example data: [[[-0.65854096 -0.81289583 -0.17044853 ...  0.          0.
    0.        ]
  [-0.7126457  -0.8206468  -0.62967294 ...  0.          0.
    0.        ]
  [-0.12946074 -0.59914047 -0.08469803 ...  0.          0.
    0.        ]
  ...
  [-0.702889   -1.6441876  -0.95275044 ...  0.          0.
    0.        ]
  [-0.16227457 -0.5545236  -0.01061052 ...  0.          0.
    0.        ]
  [ 0.17467183 -1.1306826  -0.30235663 ...  0.          0.
    0.        ]]]
  - anchor: (1, 900, 11), dtype: float32
    Example data: [[[ -3.0179338    5.499397    -0.8637523  ...   0.           0.
     0.        ]
  [  3.8772693  -27.766438    -2.0446386  ...   0.           0.
     0.        ]
  [-23.310377    28.47417      0.11812632 ...   0.           0.
     0.        ]
  ...
  [ -9.503412    47.198254     1.0450996  ...   0.           0.
     0.        ]
  [ 17.598536   -37.61403     -1.8509827  ...   0.           0.
     0.        ]
  [-11.910927    29.049755     0.20627801 ...   0.           0.
     0.        ]]]
  - anchor_embed: (1, 900, 256), dtype: float32
    Example data: [[[ 1.4262058  -0.37804246 -0.6569587  ... -0.43395245 -0.5874483
   -0.5874483 ]
  [ 1.1445364  -0.63541424 -0.63541424 ... -0.43395245 -0.5874483
   -0.5874483 ]
  [ 2.1624691  -0.6484863  -0.6484863  ... -0.43395245 -0.5874483
   -0.5874483 ]
  ...
  [ 2.4468293  -0.1902553  -0.60778666 ... -0.43395245 -0.5874483
   -0.5874483 ]
  [-0.3910775  -0.7021856  -0.7021856  ... -0.4339523  -0.5874483
   -0.5874483 ]
  [ 2.3260374  -0.5261029  -0.6358418  ... -0.4339523  -0.5874483
   -0.5874483 ]]]
  - time_interval: (1, 1), dtype: float32
    Example data: [[0.5]]

### Calibration Data Stats ###

inst_dfg0:
  - Min: -7.339239120483398, Max: 4.43999719619751, Mean: -0.013176496140658855, Std: 0.4585071802139282

anchor:
  - Min: -53.18491744995117, Max: 53.5064697265625, Mean: 0.2622547149658203, Std: 10.470222473144531

anchor_embed:
  - Min: -0.762471079826355, Max: 5.299157619476318, Mean: 1.3010058808049507e-08, Std: 0.9999400973320007

time_interval:
  - Min: 0.5, Max: 0.5, Mean: 0.5, Std: 0.0