hb_config_generator --full-yaml --model model.onnx --march nash-e

#单输入
hb_compile --march ${march} \                      --proto ${caffe_proto} \           --model ${caffe_model/onnx_model} \           --input-shape ${input_node_name} ${input_shape} #多输入hb_compile --march ${march} \           --proto ${caffe_proto} \           --model ${caffe_model/onnx_model} \           --input-shape input.0 1x1x224x224           --input-shape input.1 1x1x224x224           --input-shape input.2 1x1x224x224        

hb_compile --config ${config_file} \           --model ${model.bc} 

hb_compile --fast-perf --model ${caffe_model/onnx_model} \           --proto ${caffe_proto} \           --march ${march} \            --input-shape ${input_node_name} ${input_shape} 

hb_compile --config ${config_file}  

hb_verifier -m googlenet_optimized_float_model.onnx,googlenet_calibrated_model.onnx -i input.npy 

hb_verifier -m googlenet_optimized_float_model.onnx,googlenet_quantized_model.bc -i input.npy

hb_verifier -m googlenet_quantized_model.bc,googlenet.hbm -i runtime_input.npy

#输出bc模型/hbm的输入输出信息hb_model_info model.bc/model.hbm#输出bc模型/hbm的输入输出信息并在隐藏文件夹生成onnx/prototxthb_model_info model.bc/model.hbm -v

import onnxfrom hbdk4.compiler.onnx import exportfrom hbdk4.compiler import convert,save#加载onnx模型ptq_onnx = onnx.load("xx_ptq_model.onnx")#export为bc模型bc_model = export(ptq_onnx)#保存bc模型save(bc_model,"model.bc")

import torchimport torchvisionfrom hbdk4.compiler.torch import statistics as torch_statisticsfrom hbdk4.compiler.torch import exportfrom hbdk4.compiler import save# 载入浮点resnetmodule = torchvision.models.resnet18(pretrained=True)example_input = torch.rand(1, 3, 224, 224)module = torch.jit.trace(module, example_input)
torch_statistics(module, example_input) # 打印torch op列表和数量# 将torchscript导出为bcexported_module = export(module, example_input, name="TorchModel", input_names=["image"], output_names=["pred"]) save(exported_modul,"model.bc")

horizon_plugin_pytorch.quantization.hbdk4.export(model: Module, example_inputs: Any, *, name: str = 'forward', input_names: Any | None = None, output_names: Any | None = None, input_descs: Any | None = None, output_descs: Any | None = None)Export nn.Module to hbir model.Parameters:--model – Input model.#输入的伪量化模型，需要是eval()后的--example_inputs – Example input for tracing.#给定的作为trace的输入--name – The name of func in exported module. Users can get the func by getattr(hbir_module, name).#导出的bc的名称--input_names – Set hbir inputs with given names, should have the same structure with example_inputs.#导出的bc的输入名称--output_names – Set hbir outputs with given names, should have the same structure with model output.#导出的bc的输出名称--input_descs – Set hbir inputs with given descriptions, should have the same structure with example_inputs.#导出的bc的输入描述信息，用户可自定义，板端部署时使用--output_descs – Set hbir outputs with given descriptions, should have the same structure with model output.##导出的bc的输出描述信息，用户可自定义，板端部署时使用，例如anchor的score信息等Returns: Hbir model wrapped with Module.

from hbdk4.compiler.torch import exportfrom hbdk4.compiler import statistics, save, load,visualize,compilefrom hbdk4.compiler.march import Marchfrom hbdk4.compiler import convert,hbm_perf #使用load加载伪量化bcmodel=load("qat.bc")#使用visualize生成onnx可视化bcvisualize(model, "qat_ori.onnx") func = model.functions[0]#batch拆分，此过程为batch nv12输入的必须操作batch_input = ["_input_0"] for input in func.inputs[::-1]:    for name in batch_input[::-1]:        if name in input.name:            input.insert_split(dim=0)#可视化已做完batch拆分的bcvisualize(model, "qat_split_batch.onnx")#插入预处理节点func = model.functions[0]#pyramid_input为模型中NV12输入的name,可以通过可视化qat_split_batch.onnx获取#ddr_input为模型中ddr输入的name,可以通过可视化qat_split_batch.onnx获取pyramid_input = ['_input_0_0','_input_0_1','_input_0_2','_input_0_3','_input_0_4','_input_0_5'] # 部署时数据来源于pyramid的输入节点名称列表ddr_input = "_input_1"     # 部署时数据来源于ddr的输入节点名称列表#插入nv12节点for input in func.inputs[::-1]:    print(input.name)    if input.name in pyramid_input:        #pyramid&resizer 只支持 NHWC 的 input layout        input.insert_transpose(permutes=[0, 3, 1, 2])        # 插入前处理节点，这里模型训练是YUV444图像，所以mode配置为None        input.insert_image_preprocess(mode=None, divisor=1, mean=[128, 128, 128], std=[128, 128, 128])        input.insert_image_convert("nv12")        print("-----insert nv12 success-----")#插入resizer节点#for input in func.inputs[::-1]:    #if input.name in resizer_input:        # pyramid&resizer 只支持 NHWC 的 input layout        #node = input.insert_transpose(permutes=[0, 3, 1, 2])        # 插入前处理节点，具体可参考下一节的说明        #node = input.insert_image_preprocess(mode=None, divisor=1, mean=[128, 128, 128], std=[128, 128, 128])        #node.insert_roi_resize("nv12")#插入transpose节点for input in func.inputs[::1]:    if input.name == ddr_input:        #layerout变换：NCHW->NHWC        input.insert_transpose(permutes=[0, 2, 3, 1])#可视化插入预处理节点后的模型visualize(model, "qat_preprocess.onnx") #将插入预处理节点后hbir保存为bcsave(model,"qat_preprocess.bc")#将伪量化bc convert为定点bc#配置advice参数显示算子相关信息quantized_model=convert(model,'nash-e',advice=True,advice_path='./')#可视化定点bc visualize(quantized_model, "quantized_ori.onnx")#删除量化/反量化节点# convert后的bc的首尾部默认包含量化反量化节点，可以进行手工删除node_type_mapping = {    "qnt.quantize": "Quantize",    "qnt.dequantize": "Dequantize",    "hbir.transpose": "Transpose",    "hbtl.call::quant::qcast": "Quantize",    "hbtl.call::quant::dcast": "Dequantize",    "hbtl.call::native::Transpose": "Transpose",    "hbir.cast_type": "Cast",    "hbir.reshape": "Reshape",    "hbtl.call::native::Cast": "Cast",    "hbtl.call::native::Reshape": "Reshape",}def get_type_for_hbtl_call(attached_op):    schema = attached_op.schema    node_type = attached_op.type + "::" + \        schema.namespace + "::" + schema.signature    return node_typedef remove_op(func, op_type=None, op_name=None):    for loc in func.inputs + func.outputs:        if not loc.is_removable[0]:            continue        attached_op = loc.get_attached_op[0]        removed = None        # 目前hbir模型中的op name格式还未完全确定，暂建议使用op type来删除节点        attached_op_name = attached_op.name        if op_name and attached_op.name in op_name:            removed, diagnostic = loc.remove_attached_op()        elif op_type and attached_op.type in node_type_mapping.keys() \                and node_type_mapping[attached_op.type] in op_type:            removed, diagnostic = loc.remove_attached_op()        elif attached_op.type == "hbtl.call":            # 由于同一type的op在后端可能对应多种实现，因此采用“签名”的方式确认具体类型            node_type = get_type_for_hbtl_call(attached_op)            if op_type and node_type in node_type_mapping.keys() \                    and node_type_mapping[node_type] in op_type:                removed, diagnostic = loc.remove_attached_op()        if removed is True:            print(f'Remove node', op_type, "successfully")        if removed is False:            raise ValueError(f'Remove node type', op_type,                f"Failed when deleting {attached_op.name} operator,"                f"error: {diagnostic}")func = quantized_model[0]   # 删除reshape节点#remove_op(func, op_type="Reshape")#remove_op(func, op_type="Cast")# 删除量化反量化节点remove_op(func, op_type="Dequantize")remove_op(func, op_type="Quantize")# 删除max后的reshape节点#remove_op(func, op_type="Reshape")# 删除Transpose节点#remove_op(func, op_type="Transpose")print("-----remove_quant_dequant OK-----")save(quantized_model,"quantized_modified.bc")visualize(quantized_model, "quantized_remove_dequa.onnx")
#使用compile编译定点bc为hbmprint("-----start to compile model-----")#params = {'jobs': 48, 'balance': 100, 'progress_bar': True,          'opt': 2,'debug':True}compile(  quantized_bc,   march="nash-e",  path="model.hbm",  **params)print("-----end to compile model-----")#模型性能预估print("-----start to perf model-----")save_path="./perf"hbm_perf('model.hbm',save_path)