2.7、图调度功能

introduction

PPQ 的图调度功能
展示如何自由调度算子，并实现混合精度推理

code

import torch
import torchvision
from ppq import *
from ppq.api import *# ------------------------------------------------------------
# 在这个脚本中，我们将向你展示如何自由调度算子，并实现混合精度推理
# 在开始之前，我们首先设计一个支持混合精度的量化器
# ------------------------------------------------------------
class MyQuantizer(BaseQuantizer):# ------------------------------------------------------------# quant_operation_types 是一个类型枚举，在这里你需要写下所有该量化器所需要量化的算子# ------------------------------------------------------------@ propertydef quant_operation_types(self) -> set:return {'Conv'}# ------------------------------------------------------------# 一旦你确定了那些算子需要量化，则需要在 init_quantize_config 为他们初始化量化信息# 然而你需要注意的是，由于手动调度的存在，用户可以强制调度一个类型不在 quant_operation_types 中的算子来到量化平台# 我建议你针对这类情况进行回应。或者，在探测到算子类型并非可量化类型后进行报错# ------------------------------------------------------------def init_quantize_config(self, operation: Operation) -> OperationQuantizationConfig:# ------------------------------------------------------------# 为卷积算子初始化量化信息，只量化卷积算子的输入(input & weight)，bias 不做量化# ------------------------------------------------------------if operation.type == 'Conv':config = self.create_default_quant_config(operation_meta     = operation.meta_data, num_of_bits        = 4,quant_max          = 15, quant_min          = -16,observer_algorithm = 'percentile', policy             = QuantizationPolicy(QuantizationProperty.PER_TENSOR +QuantizationProperty.LINEAR +QuantizationProperty.SYMMETRICAL),rounding           = RoundingPolicy.ROUND_HALF_EVEN)# ------------------------------------------------------------# 关闭所有输出量化，状态设置为fp32# ------------------------------------------------------------for tensor_quant_config in config.output_quantization_config:tensor_quant_config.state = QuantizationStates.FP32# ------------------------------------------------------------# 关闭 bias 量化，状态设置为fp32# ------------------------------------------------------------if operation.num_of_input == 3:config.input_quantization_config[-1].state = QuantizationStates.FP32# ------------------------------------------------------------# 如果算子被调度到 INT8 平台上，执行 INT8 的量化# ------------------------------------------------------------if operation.platform == TargetPlatform.ACADEMIC_INT8:print(f'{operation.name} has been dispatched to INT8')config.input_quantization_config[0].num_of_bits = 8config.input_quantization_config[0].quant_max   = 127config.input_quantization_config[0].quant_min   = -128config.input_quantization_config[1].num_of_bits = 8config.input_quantization_config[1].quant_max   = 127config.input_quantization_config[1].quant_min   = -128return configelse:raise TypeError(f'Unsupported Op Type: {operation.type}')# ------------------------------------------------------------# 当前量化器进行量化的算子都将被发往一个指定的目标平台# 这里我们选择 TargetPlatform.ACADEMIC_INT4 作为目标平台# ------------------------------------------------------------@ propertydef target_platform(self) -> TargetPlatform:return TargetPlatform.ACADEMIC_INT4# 注册我们的量化器
register_network_quantizer(MyQuantizer, platform=TargetPlatform.ACADEMIC_INT4)# ------------------------------------------------------------
# 下面，我们向你展示 PPQ 的手动调度逻辑
# 我们仍然以 MobilenetV2 举例，向你展示如何完成混合精度调度
# ------------------------------------------------------------BATCHSIZE   = 32
INPUT_SHAPE = [BATCHSIZE, 3, 224, 224]
DEVICE      = 'cuda'
PLATFORM    = TargetPlatform.ACADEMIC_INT4
CALIBRATION = [torch.rand(size=INPUT_SHAPE) for _ in range(32)]
def collate_fn(batch: torch.Tensor) -> torch.Tensor: return batch.to(DEVICE)model = torchvision.models.mobilenet.mobilenet_v2(pretrained=True)
model = model.to(DEVICE)# ------------------------------------------------------------
# 为了执行手动调度，你必须首先创建 QuantizationSetting
# 使用 QuantizationSetting.dispatch_table 属性来传递调度方案
# 大多数预制量化器没有针对 INT4 写过量化方案，因此只支持 FP32 - INT8 的相互调度
# ------------------------------------------------------------
QS = QuantizationSettingFactory.default_setting()# 示例语句：下面的语句将把 Op1 调度到 FP32 平台
# QS.dispatching_table.append(operation='Op1', platform=TargetPlatform.FP32)# ------------------------------------------------------------
# 我如何知道调度那些层到高精度会得到最优的性能表现？
# layerwise_error_analyse 函数正是为此设计的，我们调用该方法
# 然后选择那些误差较高的层调度到高精度平台# 我们首先将所有层全部送往 INT4，然后执行误差分析
# ------------------------------------------------------------# ------------------------------------------------------------
# 如果你使用 ENABLE_CUDA_KERNEL 方法
# PPQ 将会尝试编译自定义的高性能量化算子，这一过程需要编译环境的支持
# 如果你在编译过程中发生错误，你可以删除此处对于 ENABLE_CUDA_KERNEL 方法的调用
# 这将显著降低 PPQ 的运算速度；但即使你无法编译这些算子，你仍然可以使用 pytorch 的 gpu 算子完成量化
# ------------------------------------------------------------
with ENABLE_CUDA_KERNEL():dump_torch_to_onnx(model=model, onnx_export_file='./model.onnx', input_shape=INPUT_SHAPE, input_dtype=torch.float32)graph = load_onnx_graph(onnx_import_file='./model.onnx')quantized = quantize_native_model(model=graph, calib_dataloader=CALIBRATION,calib_steps=32, input_shape=INPUT_SHAPE,collate_fn=collate_fn, platform=PLATFORM,device=DEVICE, verbose=0, setting=QS)# ------------------------------------------------------------# 使用 graphwise_analyse 衡量调度前的量化误差# ------------------------------------------------------------reports = graphwise_error_analyse(graph=quantized, running_device=DEVICE, collate_fn=collate_fn,dataloader=CALIBRATION)# ------------------------------------------------------------# 执行逐层分析，结果是一个字典，该字典内写入了所有算子的单层量化误差# ------------------------------------------------------------reports = layerwise_error_analyse(graph=quantized, running_device=DEVICE, collate_fn=collate_fn,dataloader=CALIBRATION, verbose=False)# 从大到小排序单层误差sensitivity = [(op_name, error) for op_name, error in reports.items()]sensitivity = sorted(sensitivity, key=lambda x: x[1], reverse=True)# ------------------------------------------------------------# 将前十个误差最大的层送上INT8，并重新量化# ------------------------------------------------------------for op_name, _ in sensitivity[: 10]:QS.dispatching_table.append(operation=op_name, platform=TargetPlatform.ACADEMIC_INT8)graph = load_onnx_graph(onnx_import_file='./model.onnx')quantized = quantize_native_model(model=graph, calib_dataloader=CALIBRATION,calib_steps=32, input_shape=INPUT_SHAPE,collate_fn=collate_fn, platform=PLATFORM,device=DEVICE, verbose=0, setting=QS)# ------------------------------------------------------------# 使用 graphwise_analyse 衡量最终的量化误差# ------------------------------------------------------------reports = graphwise_error_analyse(graph=quantized, running_device=DEVICE, collate_fn=collate_fn,dataloader=CALIBRATION)

result

      ____  ____  __   ____                    __              __/ __ \/ __ \/ /  / __ \__  ______ _____  / /_____  ____  / // /_/ / /_/ / /  / / / / / / / __ `/ __ \/ __/ __ \/ __ \/ // ____/ ____/ /__/ /_/ / /_/ / /_/ / / / / /_/ /_/ / /_/ / //_/   /_/   /_____\___\_\__,_/\__,_/_/ /_/\__/\____/\____/_/[31m[Warning] Compling Kernels... Please wait (It will take a few minutes).[0m
[05:38:56] PPQ Quantization Config Refine Pass Running ... Finished.
[05:38:56] PPQ Quantization Fusion Pass Running ...        [31m[Warning] Unexpected dispatching was found: Op Conv_0 and Clip_3 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_4 and Clip_7 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_9 and Clip_12 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_13 and Clip_16 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_18 and Clip_21 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_22 and Clip_25 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_28 and Clip_31 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_32 and Clip_35 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_37 and Clip_40 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_41 and Clip_44 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_47 and Clip_50 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_51 and Clip_54 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_57 and Clip_60 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_61 and Clip_64 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_66 and Clip_69 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_70 and Clip_73 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_76 and Clip_79 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_80 and Clip_83 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_86 and Clip_89 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_90 and Clip_93 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_96 and Clip_99 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_100 and Clip_103 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_105 and Clip_108 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_109 and Clip_112 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_115 and Clip_118 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_119 and Clip_122 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_125 and Clip_128 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_129 and Clip_132 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_134 and Clip_137 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_138 and Clip_141 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_144 and Clip_147 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_148 and Clip_151 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_154 and Clip_157 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_158 and Clip_161 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_163 and Clip_166 should both be quantized operation.[0m
Finished.
[05:38:56] PPQ Quantize Simplify Pass Running ...          Finished.
[05:38:56] PPQ Parameter Quantization Pass Running ...     Finished.
[05:38:56] PPQ Runtime Calibration Pass Running ...        
Calibration Progress(Phase 1):   0%|          | 0/32 [00:00<?, ?it/s]
Calibration Progress(Phase 1):   3%|▎         | 1/32 [00:00<00:03,  7.90it/s]
Calibration Progress(Phase 1):   9%|▉         | 3/32 [00:00<00:02, 10.00it/s]
Calibration Progress(Phase 1):  16%|█▌        | 5/32 [00:00<00:02, 10.62it/s]
Calibration Progress(Phase 1):  22%|██▏       | 7/32 [00:00<00:02, 11.00it/s]
Calibration Progress(Phase 1):  28%|██▊       | 9/32 [00:00<00:02, 11.09it/s]
Calibration Progress(Phase 1):  34%|███▍      | 11/32 [00:01<00:01, 11.19it/s]
Calibration Progress(Phase 1):  41%|████      | 13/32 [00:01<00:01, 11.25it/s]
Calibration Progress(Phase 1):  47%|████▋     | 15/32 [00:01<00:01, 11.35it/s]
Calibration Progress(Phase 1):  53%|█████▎    | 17/32 [00:01<00:01, 11.47it/s]
Calibration Progress(Phase 1):  59%|█████▉    | 19/32 [00:01<00:01, 11.51it/s]
Calibration Progress(Phase 1):  66%|██████▌   | 21/32 [00:01<00:00, 11.37it/s]
Calibration Progress(Phase 1):  72%|███████▏  | 23/32 [00:02<00:00, 11.38it/s]
Calibration Progress(Phase 1):  78%|███████▊  | 25/32 [00:02<00:00, 11.33it/s]
Calibration Progress(Phase 1):  84%|████████▍ | 27/32 [00:02<00:00, 11.28it/s]
Calibration Progress(Phase 1):  91%|█████████ | 29/32 [00:02<00:00, 11.24it/s]
Calibration Progress(Phase 1):  97%|█████████▋| 31/32 [00:02<00:00, 11.31it/s]
Calibration Progress(Phase 1): 100%|██████████| 32/32 [00:02<00:00, 11.19it/s]
Finished.
[05:38:59] PPQ Quantization Alignment Pass Running ...     Finished.
[05:38:59] PPQ Passive Parameter Quantization Running ...  Finished.
[05:38:59] PPQ Parameter Baking Pass Running ...           Finished.
--------- Network Snapshot ---------
Num of Op:                    [100]
Num of Quantized Op:          [52]
Num of Variable:              [277]
Num of Quantized Var:         [201]
------- Quantization Snapshot ------
Num of Quant Config:          [208]
BAKED:                        [52]
ACTIVATED:                    [52]
FP32:                         [104]
Network Quantization Finished.Analysing Graphwise Quantization Error(Phrase 1)::   0%|          | 0/8 [00:00<?, ?it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  12%|█▎        | 1/8 [00:00<00:00,  9.44it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  25%|██▌       | 2/8 [00:00<00:00,  9.57it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  38%|███▊      | 3/8 [00:00<00:00,  9.38it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  50%|█████     | 4/8 [00:00<00:00,  9.30it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  62%|██████▎   | 5/8 [00:00<00:00,  9.25it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  75%|███████▌  | 6/8 [00:00<00:00,  9.22it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  88%|████████▊ | 7/8 [00:00<00:00,  9.21it/s]
Analysing Graphwise Quantization Error(Phrase 1):: 100%|██████████| 8/8 [00:00<00:00,  9.20it/s]
Analysing Graphwise Quantization Error(Phrase 1):: 100%|██████████| 8/8 [00:00<00:00,  8.22it/s]Analysing Graphwise Quantization Error(Phrase 2)::   0%|          | 0/8 [00:00<?, ?it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  12%|█▎        | 1/8 [00:00<00:00,  7.84it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  25%|██▌       | 2/8 [00:00<00:00,  8.24it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  38%|███▊      | 3/8 [00:00<00:00,  8.28it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  50%|█████     | 4/8 [00:00<00:00,  8.40it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  62%|██████▎   | 5/8 [00:00<00:00,  8.47it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  75%|███████▌  | 6/8 [00:00<00:00,  8.51it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  88%|████████▊ | 7/8 [00:00<00:00,  8.52it/s]
Analysing Graphwise Quantization Error(Phrase 2):: 100%|██████████| 8/8 [00:00<00:00,  8.54it/s]
Analysing Graphwise Quantization Error(Phrase 2):: 100%|██████████| 8/8 [00:01<00:00,  7.46it/s]
Layer     | NOISE:SIGNAL POWER RATIO 
Conv_162: | ████████████████████ | 4.469564
Conv_163: | █████████████████    | 3.796862
Conv_55:  | ███████████          | 2.358960
Conv_45:  | █████████            | 1.916194
Conv_104: | ████████             | 1.848709
Conv_65:  | ███████              | 1.608640
Conv_36:  | ███████              | 1.527322
Conv_113: | ███████              | 1.489844
Conv_125: | ██████               | 1.338494
Conv_26:  | ██████               | 1.292400
Conv_123: | ██████               | 1.233793
Conv_133: | █████                | 1.195669
Conv_148: | █████                | 1.163949
Conv_84:  | █████                | 1.156629
Conv_134: | █████                | 1.155597
Conv_142: | █████                | 1.123850
Conv_144: | █████                | 1.119863
Conv_138: | █████                | 1.115643
Conv_70:  | █████                | 1.106950
Conv_152: | █████                | 1.100367
Conv_90:  | █████                | 1.048857
Conv_94:  | █████                | 1.035837
Conv_51:  | █████                | 1.009898
Conv_115: | ████                 | 0.957407
Conv_57:  | ████                 | 0.947991
Conv_80:  | ████                 | 0.945680
Conv_154: | ████                 | 0.931150
Conv_105: | ████                 | 0.915809
Conv_109: | ████                 | 0.912738
Conv_41:  | ████                 | 0.876404
Conv_86:  | ████                 | 0.841202
Conv_17:  | ████                 | 0.796938
Conv_96:  | ████                 | 0.784382
Conv_100: | ███                  | 0.760816
Conv_129: | ███                  | 0.756461
Conv_119: | ███                  | 0.754041
Conv_74:  | ███                  | 0.744942
Conv_66:  | ███                  | 0.729332
Conv_32:  | ███                  | 0.726990
Conv_28:  | ███                  | 0.704079
Conv_47:  | ███                  | 0.653307
Conv_76:  | ███                  | 0.618381
Conv_8:   | ███                  | 0.603043
Conv_22:  | ███                  | 0.602080
Conv_37:  | ██                   | 0.509791
Conv_61:  | ██                   | 0.506709
Conv_13:  | ██                   | 0.495827
Conv_9:   | ██                   | 0.404825
Conv_158: | ██                   | 0.392166
Conv_18:  | █                    | 0.221014
Conv_4:   |                      | 0.082779
Conv_0:   |                      | 0.001353Analysing Layerwise quantization error::   0%|          | 0/52 [00:00<?, ?it/s]
Analysing Layerwise quantization error::   2%|▏         | 1/52 [00:00<00:12,  4.01it/s]
Analysing Layerwise quantization error::   4%|▍         | 2/52 [00:00<00:12,  3.98it/s]
Analysing Layerwise quantization error::   6%|▌         | 3/52 [00:00<00:12,  3.99it/s]
Analysing Layerwise quantization error::   8%|▊         | 4/52 [00:01<00:11,  4.00it/s]
Analysing Layerwise quantization error::  10%|▉         | 5/52 [00:01<00:11,  3.99it/s]
Analysing Layerwise quantization error::  12%|█▏        | 6/52 [00:01<00:11,  4.00it/s]
Analysing Layerwise quantization error::  13%|█▎        | 7/52 [00:01<00:11,  3.99it/s]
Analysing Layerwise quantization error::  15%|█▌        | 8/52 [00:02<00:11,  4.00it/s]
Analysing Layerwise quantization error::  17%|█▋        | 9/52 [00:02<00:10,  4.00it/s]
Analysing Layerwise quantization error::  19%|█▉        | 10/52 [00:02<00:10,  4.01it/s]
Analysing Layerwise quantization error::  21%|██        | 11/52 [00:02<00:10,  4.01it/s]
Analysing Layerwise quantization error::  23%|██▎       | 12/52 [00:03<00:10,  4.00it/s]
Analysing Layerwise quantization error::  25%|██▌       | 13/52 [00:03<00:09,  4.00it/s]
Analysing Layerwise quantization error::  27%|██▋       | 14/52 [00:03<00:09,  4.00it/s]
Analysing Layerwise quantization error::  29%|██▉       | 15/52 [00:03<00:09,  4.01it/s]
Analysing Layerwise quantization error::  31%|███       | 16/52 [00:03<00:08,  4.01it/s]
Analysing Layerwise quantization error::  33%|███▎      | 17/52 [00:04<00:08,  3.99it/s]
Analysing Layerwise quantization error::  35%|███▍      | 18/52 [00:04<00:08,  4.00it/s]
Analysing Layerwise quantization error::  37%|███▋      | 19/52 [00:04<00:08,  4.01it/s]
Analysing Layerwise quantization error::  38%|███▊      | 20/52 [00:04<00:07,  4.01it/s]
Analysing Layerwise quantization error::  40%|████      | 21/52 [00:05<00:07,  4.00it/s]
Analysing Layerwise quantization error::  42%|████▏     | 22/52 [00:05<00:07,  4.00it/s]
Analysing Layerwise quantization error::  44%|████▍     | 23/52 [00:05<00:07,  4.01it/s]
Analysing Layerwise quantization error::  46%|████▌     | 24/52 [00:05<00:06,  4.01it/s]
Analysing Layerwise quantization error::  48%|████▊     | 25/52 [00:06<00:06,  4.00it/s]
Analysing Layerwise quantization error::  50%|█████     | 26/52 [00:06<00:06,  4.00it/s]
Analysing Layerwise quantization error::  52%|█████▏    | 27/52 [00:06<00:06,  4.00it/s]
Analysing Layerwise quantization error::  54%|█████▍    | 28/52 [00:07<00:06,  4.00it/s]
Analysing Layerwise quantization error::  56%|█████▌    | 29/52 [00:07<00:05,  4.01it/s]
Analysing Layerwise quantization error::  58%|█████▊    | 30/52 [00:07<00:05,  4.00it/s]
Analysing Layerwise quantization error::  60%|█████▉    | 31/52 [00:07<00:05,  3.98it/s]
Analysing Layerwise quantization error::  62%|██████▏   | 32/52 [00:08<00:05,  3.99it/s]
Analysing Layerwise quantization error::  63%|██████▎   | 33/52 [00:08<00:04,  3.99it/s]
Analysing Layerwise quantization error::  65%|██████▌   | 34/52 [00:08<00:04,  3.97it/s]
Analysing Layerwise quantization error::  67%|██████▋   | 35/52 [00:08<00:04,  3.99it/s]
Analysing Layerwise quantization error::  69%|██████▉   | 36/52 [00:09<00:04,  3.91it/s]
Analysing Layerwise quantization error::  71%|███████   | 37/52 [00:09<00:03,  3.93it/s]
Analysing Layerwise quantization error::  73%|███████▎  | 38/52 [00:09<00:03,  3.95it/s]
Analysing Layerwise quantization error::  75%|███████▌  | 39/52 [00:09<00:03,  3.94it/s]
Analysing Layerwise quantization error::  77%|███████▋  | 40/52 [00:10<00:03,  3.95it/s]
Analysing Layerwise quantization error::  79%|███████▉  | 41/52 [00:10<00:02,  3.89it/s]
Analysing Layerwise quantization error::  81%|████████  | 42/52 [00:10<00:02,  3.87it/s]
Analysing Layerwise quantization error::  83%|████████▎ | 43/52 [00:10<00:02,  3.87it/s]
Analysing Layerwise quantization error::  85%|████████▍ | 44/52 [00:11<00:02,  3.91it/s]
Analysing Layerwise quantization error::  87%|████████▋ | 45/52 [00:11<00:01,  3.91it/s]
Analysing Layerwise quantization error::  88%|████████▊ | 46/52 [00:11<00:01,  3.90it/s]
Analysing Layerwise quantization error::  90%|█████████ | 47/52 [00:11<00:01,  3.93it/s]
Analysing Layerwise quantization error::  92%|█████████▏| 48/52 [00:12<00:01,  3.83it/s]
Analysing Layerwise quantization error::  94%|█████████▍| 49/52 [00:12<00:00,  3.85it/s]
Analysing Layerwise quantization error::  96%|█████████▌| 50/52 [00:12<00:00,  3.88it/s]
Analysing Layerwise quantization error::  98%|█████████▊| 51/52 [00:12<00:00,  3.85it/s]
Analysing Layerwise quantization error:: 100%|██████████| 52/52 [00:13<00:00,  3.79it/s]
Analysing Layerwise quantization error:: 100%|██████████| 52/52 [00:13<00:00,  3.95it/s]
Conv_4 has been dispatched to INT8
Conv_13 has been dispatched to INT8
Conv_22 has been dispatched to INT8
Conv_32 has been dispatched to INT8
Conv_41 has been dispatched to INT8
Conv_70 has been dispatched to INT8
Conv_80 has been dispatched to INT8
Conv_109 has been dispatched to INT8
Conv_148 has been dispatched to INT8
Conv_158 has been dispatched to INT8
[05:39:15] PPQ Quantization Config Refine Pass Running ... Finished.
[05:39:15] PPQ Quantization Fusion Pass Running ...        [31m[Warning] Unexpected dispatching was found: Op Conv_0 and Clip_3 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_4 and Clip_7 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_9 and Clip_12 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_13 and Clip_16 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_18 and Clip_21 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_22 and Clip_25 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_28 and Clip_31 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_32 and Clip_35 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_37 and Clip_40 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_41 and Clip_44 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_47 and Clip_50 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_51 and Clip_54 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_57 and Clip_60 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_61 and Clip_64 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_66 and Clip_69 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_70 and Clip_73 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_76 and Clip_79 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_80 and Clip_83 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_86 and Clip_89 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_90 and Clip_93 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_96 and Clip_99 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_100 and Clip_103 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_105 and Clip_108 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_109 and Clip_112 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_115 and Clip_118 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_119 and Clip_122 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_125 and Clip_128 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_129 and Clip_132 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_134 and Clip_137 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_138 and Clip_141 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_144 and Clip_147 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_148 and Clip_151 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_154 and Clip_157 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_158 and Clip_161 should both be quantized operation.[0m
[31m[Warning] Unexpected dispatching was found: Op Conv_163 and Clip_166 should both be quantized operation.[0m
Finished.
[05:39:15] PPQ Quantize Simplify Pass Running ...          Finished.
[05:39:15] PPQ Parameter Quantization Pass Running ...     Finished.
[05:39:15] PPQ Runtime Calibration Pass Running ...        
Calibration Progress(Phase 1):   0%|          | 0/32 [00:00<?, ?it/s]
Calibration Progress(Phase 1):   6%|▋         | 2/32 [00:00<00:02, 12.12it/s]
Calibration Progress(Phase 1):  12%|█▎        | 4/32 [00:00<00:02, 13.05it/s]
Calibration Progress(Phase 1):  19%|█▉        | 6/32 [00:00<00:01, 13.25it/s]
Calibration Progress(Phase 1):  25%|██▌       | 8/32 [00:00<00:01, 13.44it/s]
Calibration Progress(Phase 1):  31%|███▏      | 10/32 [00:00<00:01, 13.48it/s]
Calibration Progress(Phase 1):  38%|███▊      | 12/32 [00:00<00:01, 13.28it/s]
Calibration Progress(Phase 1):  44%|████▍     | 14/32 [00:01<00:01, 13.10it/s]
Calibration Progress(Phase 1):  50%|█████     | 16/32 [00:01<00:01, 12.92it/s]
Calibration Progress(Phase 1):  56%|█████▋    | 18/32 [00:01<00:01, 12.99it/s]
Calibration Progress(Phase 1):  62%|██████▎   | 20/32 [00:01<00:00, 13.03it/s]
Calibration Progress(Phase 1):  69%|██████▉   | 22/32 [00:01<00:00, 12.95it/s]
Calibration Progress(Phase 1):  75%|███████▌  | 24/32 [00:01<00:00, 13.02it/s]
Calibration Progress(Phase 1):  81%|████████▏ | 26/32 [00:01<00:00, 13.10it/s]
Calibration Progress(Phase 1):  88%|████████▊ | 28/32 [00:02<00:00, 13.05it/s]
Calibration Progress(Phase 1):  94%|█████████▍| 30/32 [00:02<00:00, 13.02it/s]
Calibration Progress(Phase 1): 100%|██████████| 32/32 [00:02<00:00, 13.00it/s]
Calibration Progress(Phase 1): 100%|██████████| 32/32 [00:02<00:00, 13.06it/s]
Finished.
[05:39:17] PPQ Quantization Alignment Pass Running ...     Finished.
[05:39:17] PPQ Passive Parameter Quantization Running ...  Finished.
[05:39:17] PPQ Parameter Baking Pass Running ...           Finished.
--------- Network Snapshot ---------
Num of Op:                    [100]
Num of Quantized Op:          [52]
Num of Variable:              [277]
Num of Quantized Var:         [201]
------- Quantization Snapshot ------
Num of Quant Config:          [208]
BAKED:                        [52]
ACTIVATED:                    [52]
FP32:                         [104]
Network Quantization Finished.Analysing Graphwise Quantization Error(Phrase 1)::   0%|          | 0/8 [00:00<?, ?it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  12%|█▎        | 1/8 [00:00<00:01,  6.86it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  25%|██▌       | 2/8 [00:00<00:00,  7.97it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  38%|███▊      | 3/8 [00:00<00:00,  8.36it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  50%|█████     | 4/8 [00:00<00:00,  8.61it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  62%|██████▎   | 5/8 [00:00<00:00,  8.65it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  75%|███████▌  | 6/8 [00:00<00:00,  8.65it/s]
Analysing Graphwise Quantization Error(Phrase 1)::  88%|████████▊ | 7/8 [00:00<00:00,  8.55it/s]
Analysing Graphwise Quantization Error(Phrase 1):: 100%|██████████| 8/8 [00:00<00:00,  8.70it/s]
Analysing Graphwise Quantization Error(Phrase 1):: 100%|██████████| 8/8 [00:01<00:00,  7.60it/s]Analysing Graphwise Quantization Error(Phrase 2)::   0%|          | 0/8 [00:00<?, ?it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  12%|█▎        | 1/8 [00:00<00:00,  7.75it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  25%|██▌       | 2/8 [00:00<00:00,  8.14it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  38%|███▊      | 3/8 [00:00<00:00,  8.18it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  50%|█████     | 4/8 [00:00<00:00,  8.25it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  62%|██████▎   | 5/8 [00:00<00:00,  7.03it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  75%|███████▌  | 6/8 [00:00<00:00,  6.16it/s]
Analysing Graphwise Quantization Error(Phrase 2)::  88%|████████▊ | 7/8 [00:01<00:00,  5.92it/s]
Analysing Graphwise Quantization Error(Phrase 2):: 100%|██████████| 8/8 [00:01<00:00,  5.76it/s]
Analysing Graphwise Quantization Error(Phrase 2):: 100%|██████████| 8/8 [00:01<00:00,  5.64it/s]
Layer     | NOISE:SIGNAL POWER RATIO 
Conv_162: | ████████████████████ | 2.602241
Conv_163: | █████████████████    | 2.160837
Conv_142: | ████████             | 0.991362
Conv_152: | ███████              | 0.943710
Conv_133: | ███████              | 0.885206
Conv_154: | ██████               | 0.840136
Conv_144: | ██████               | 0.808373
Conv_113: | ██████               | 0.771283
Conv_123: | ██████               | 0.761448
Conv_55:  | ██████               | 0.751858
Conv_148: | ██████               | 0.734521
Conv_134: | █████                | 0.700948
Conv_138: | █████                | 0.688839
Conv_104: | █████                | 0.653958
Conv_125: | █████                | 0.606137
Conv_129: | ████                 | 0.512940
Conv_94:  | ███                  | 0.420290
Conv_26:  | ███                  | 0.380846
Conv_65:  | ███                  | 0.348712
Conv_45:  | ███                  | 0.326942
Conv_115: | ██                   | 0.313652
Conv_51:  | ██                   | 0.309173
Conv_90:  | ██                   | 0.308499
Conv_109: | ██                   | 0.307272
Conv_119: | ██                   | 0.305434
Conv_84:  | ██                   | 0.300523
Conv_105: | ██                   | 0.291879
Conv_100: | ██                   | 0.289319
Conv_36:  | ██                   | 0.285215
Conv_158: | ██                   | 0.242913
Conv_57:  | ██                   | 0.236399
Conv_96:  | ██                   | 0.228828
Conv_28:  | ██                   | 0.224474
Conv_41:  | ██                   | 0.221757
Conv_70:  | ██                   | 0.210314
Conv_86:  | ██                   | 0.198304
Conv_80:  | █                    | 0.193799
Conv_22:  | █                    | 0.188963
Conv_17:  | █                    | 0.183890
Conv_47:  | █                    | 0.162605
Conv_32:  | █                    | 0.143125
Conv_74:  | █                    | 0.141152
Conv_66:  | █                    | 0.135066
Conv_8:   | █                    | 0.122442
Conv_76:  | █                    | 0.115116
Conv_37:  | █                    | 0.103618
Conv_13:  | █                    | 0.101174
Conv_9:   | █                    | 0.094143
Conv_61:  | █                    | 0.093169
Conv_18:  | █                    | 0.069118
Conv_4:   |                      | 0.007497
Conv_0:   |                      | 0.001353

本文来自互联网用户投稿，该文观点仅代表作者本人，不代表本站立场。本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如若转载，请注明出处：http://www.mzph.cn/news/123680.shtml

如若内容造成侵权/违法违规/事实不符，请联系多彩编程网进行投诉反馈email:809451989@qq.com，一经查实，立即删除！

相关文章

10_4阻塞和非阻塞跟poll和等待队列

10_4阻塞和非阻塞跟poll和等待队列

阻塞和非阻塞 广泛上的区别就是 应用程序如果非阻塞那读取不到数据就应该马上有返回值 阻塞的话就是在应用程序去read数据,但是设备驱动没有数据,就一直卡住,直到有数据再继续往下 补充阻塞知识,应用层大部分都是阻塞 如果要非阻塞 ,应用程序在打开设备节点的时候填写int fd …
阅读更多...
辅助驾驶功能开发-功能规范篇(22)-4-L2级辅助驾驶方案功能规范

辅助驾驶功能开发-功能规范篇(22)-4-L2级辅助驾驶方案功能规范

1.3.4 LDW系统功能定义 1.3.4.1 状态机 1.3.4.2 功能定义 1.3.4.2.1 信号需求列表 1.3.4.2.2 系统开启关闭 1）初始化 车辆上电后，车道偏离预警系统（LDW）进行初始化，控制器需要在上电后 220ms 内发出第一帧报文，并在 3s 内 完成内部自检，同时上电 3s 内不进行关联系统…
阅读更多...
面经——java后端第一天

面经——java后端第一天

1.java和c区别是&#xff1f; 性能⾼&#xff1a;C 是⼀种编译型语⾔&#xff0c;可以⽣成⾼效的机器代码&#xff0c;所以在处理⼤量数据和对计算效率有要求的应⽤场景下具有优势&#xff1b; C 和 Java 是两种编程语言&#xff0c;它们之间有很多区别&#xff0c;主要包括以…
阅读更多...
计算机网络常识

计算机网络常识

文章目录 1、HTTP2、HTTP状态码1xx&#xff08;信息性状态码&#xff09;&#xff1a;2xx&#xff08;成功状态码&#xff09;&#xff1a;3xx&#xff08;重定向状态码&#xff09;&#xff1a;4xx&#xff08;客户端错误状态码&#xff09;&#xff1a;5xx&#xff08;服务器…
阅读更多...
ActiveMQ

ActiveMQ

ActiveMQ 安装 下载网址&#xff1a;ActiveMQ 一定要和自己安装的jdk版本匹配&#xff0c;不然会报错 下载到本地之后解压缩 有可能端口号被占用 解除端口号占用&#xff0c;参考&#xff1a;Windows_端口被占用 打开cmd 查询所有的端口号 netstat -nao查询指定端口号 n…
阅读更多...
数组元素逆序交换

数组元素逆序交换

数组元素逆序交换 例如&#xff1a; 输入&#xff1a; 1,2,3,4,5 输出&#xff1a; 5,4,3,2,1 代码&#xff1a; #include<stdio.h> #define N 5 int main() {int a[N], i, temp;for (i 0; i < N; i) {scanf("%d", &a[i]);}for (i 0; i < N…
阅读更多...
C语言选择排序

C语言选择排序

1.选择排序(Selection sort) 选择排序是简单直观的排序算法。 基本思想&#xff1a;从首元素开始&#xff0c;首元素与它后面的所有元素进行比较&#xff0c;找到数列中最小的元素&#xff0c;与首元素值交换。然后下一个元素与它后面的元素比较&#xff0c;得到第二小的元素…
阅读更多...
对Linux线程的理解（什么是线程，线程的创建终止等待分离，线程互斥，Linux常见的锁，线程同步），两万字总结，有这一篇文章就够了！

对Linux线程的理解（什么是线程，线程的创建终止等待分离，线程互斥，Linux常见的锁，线程同步），两万字总结，有这一篇文章就够了！

文章目录 一、什么是线程1.线程是怎样描述的2.线程与进程的区别3.线程的优缺点4.理解Linux的轻量级进程 二、Linux线程控制1.线程创建:pthread_create()2.线程终止:pthread_exit()3.线程等待:pthread_join()4.分离线程:pthread_detach() 三、Linux线程互斥1.互斥量2.线程安全与…
阅读更多...
移动端自动化-Appium元素定位

移动端自动化-Appium元素定位

文章目录 Appium元素定位第一类&#xff1a;属性定位第二类&#xff1a;路径定位 常见问题理解appium server 和 appium inspector 以及 appium-python-client的关系 appium是跨平台的&#xff0c;支持OSX&#xff0c;Windows以及Linux系统。它允许测试人员在不同的平台&#x…
阅读更多...
我在Vscode学OpenCV 初步接触

我在Vscode学OpenCV 初步接触

OpenCV是一个开源的计算机视觉库&#xff0c;可以处理图像和视频数据。它包含了超过2500个优化过的算法&#xff0c;用于对图像和视频进行处理&#xff0c;包括目标识别、面部识别、运动跟踪、立体视觉等。OpenCV支持多种编程语言&#xff0c;包括C、Python、Java等&#xff0c…
阅读更多...
在HTML当中引入Vue控件，以element-ui为例

在HTML当中引入Vue控件，以element-ui为例

前情&#xff1a;需要实现一个同时满足按天、按周、按月选择的时间选择器&#xff0c;但是以HTML为基础写的都不太满足我的要求&#xff0c;要么只能按天选择&#xff0c;要么就是想选择久远的时间得点很久&#xff0c;除非自己写捷径&#xff0c;所以就看上了element-ui的这个…
阅读更多...
动静分离技术

动静分离技术

一、HAproxy 动静分离 1、概念&#xff1a; HAproxy 动静分离技术是一种用于优化 Web 服务器性能和提高用户体验的策略&#xff0c;它通过将动态内容和静态内容分别路由到不同的后端服务器来实现&#xff0c;减轻服务器负载&#xff0c;提高网站的响应速度。 动态内容包括由…
阅读更多...
MySQL篇---第二篇

MySQL篇---第二篇

系列文章目录 文章目录 系列文章目录一、数据库的事务二、索引是什么三、SQL优化手段有哪些一、数据库的事务 什么是事务？： 多条sql语句，要么全部成功，要么全部失败。 事务的特性： 数据库事务特性：原子性(Atomic)、一致性(Consistency)、隔离性(Isolation)、持久性 (Du…
阅读更多...
SylixOS BSP开发(七)

SylixOS BSP开发(七)

实现系统调试信息打印接口 当系统出错时或者使用内核日志时会输出一些打印信息&#xff0c;这最终都是调用到bspLib.c中的bspDebugMsg 这个接口来实现的&#xff0c;所以我们在开发BSP时&#xff0c;第一个要做的工作就是实现这个接口。 一般的调试信息都是通过串口来输出的&am…
阅读更多...
linux-tools-$(uname -r) linux-headers-$(uname -r)工具安装:

linux-tools-$(uname -r) linux-headers-$(uname -r)工具安装:

linux-tools-$(uname -r) linux-headers-$(uname -r)工具安装: ebpfebpf-virtual-machine:~$ sudo apt-get install linux-tools-$(uname -r) [sudo] ebpf 的密码&#xff1a; 正在读取软件包列表... 完成 正在分析软件包的依赖关系树... 完成 正在读取状态信息... 完成 linux…
阅读更多...
经典目标检测神经网络 - RCNN、SSD、YOLO

经典目标检测神经网络 - RCNN、SSD、YOLO

文章目录 1. 目标检测算法分类2. 区域卷积神经网络2.1 R-CNN2.2 Fast R-CNN2.3 Faster R-CNN2.4 Mask R-CNN2.5 速度和精度比较 3. 单发多框检测&#xff08;SSD&#xff09;4. YOLO 1. 目标检测算法分类 目标检测算法主要分两类&#xff1a;One-Stage与Two-Stage。One-Stage与…
阅读更多...
虚拟机Ubuntu下运行vue-element-admin项目

虚拟机Ubuntu下运行vue-element-admin项目

一.环境搭建 1.安装nodejs sudo apt install nodejs 安装完成后&#xff0c;查看对应的版本号 nodejs -v没有问题&#xff0c;会输出对应版本号&#xff0c;我这里是10.19.0 v10.19.0 2.安装npm sudo apt install npm安装完成查看对应的版本号&#xff0c;确认OK npm -…
阅读更多...
linux中nginx配置https

linux中nginx配置https

一、版本适配 版本一定要适配&#xff0c;否则会报错各种参数定位不到不识别的错误&#xff0c;以下是版本适配信息&#xff0c;各位观客自行按照以下信息匹配版本。 Nginx 1.11.5及以上版本与OpenSSL 1.0.2及以上版本兼容。Nginx 1.15.2及以上版本与OpenSSL 1.1.1及以上版本兼…
阅读更多...
代码随想录图论 第四天| 827.最大人工岛 127. 单词接龙

代码随想录图论 第四天| 827.最大人工岛 127. 单词接龙

代码随想录图论 第四天 | 827.最大人工岛 127. 单词接龙 一、827.最大人工岛 题目链接&#xff1a;https://leetcode.cn/problems/making-a-large-island/ 思路&#xff1a; class Solution {int[][] position {{-1, 0}, {1, 0}, {0, -1}, {0, 1}};int dfs(int[][] grid, i…
阅读更多...
JavaScript_Pig Game保存当前分数

JavaScript_Pig Game保存当前分数

上个文章我们基本上完成了摇色子和切换当前玩家的功能。 现在我们开始写用户选择不再摇骰子的话&#xff0c;我们将用户的当前分数存入到持有分数中&#xff01; ● 首先我们应该利用一个数组去存储两个用户的分数 const scores [0, 0];● 接着我们利用数组来对分数进行累…
阅读更多...
最新文章

Copyright @ 2022~2023