打开微信,使用扫一扫进入页面后,点击右上角菜单,
点击“发送给朋友”或“分享到朋友圈”完成分享
环境: cambricon_pytorch_u16.04_v1.7.0_pt0.15.0.tar
背景:
UNet模型用于分割任务, 在量化时, 发现最后结果有一定差异, 根据官方pytorch文档使用dump_utils逐层dump结果, 得到以下数据.
问题:
从逐层输出的差异来看, 是否可以确认是BN层引起的差异?
这属于使用问题还是算子量化的问题?
_00000_ 1.0.conv1_in errRate = 0.0037525531370192766 _00000_ 1.0.conv1_out errRate = 0.013088855892419815 _00001_ 1.0.bn1_in errRate = 0.013088855892419815 _00001_ 1.0.bn1_out errRate = 0.017539318650960922 _00002_ 1.0.relu_in errRate = 0.017539318650960922 _00002_ 1.0.relu_out errRate = 0.018641946837306023 _00003_ 1.0.conv2_in errRate = 0.018641946837306023 _00003_ 1.0.conv2_out errRate = 0.026660917326807976 _00004_ 1.0.bn2_in errRate = 0.026660917326807976 _00004_ 1.0.bn2_out errRate = 0.0341462679207325 _00005_ 1.0.downsample.0_in errRate = 0.0037525531370192766 _00005_ 1.0.downsample.0_out errRate = 0.013248566538095474 _00006_ 1.0.downsample.1_in errRate = 0.013248566538095474 _00006_ 1.0.downsample.1_out errRate = 0.01710827648639679 _00007_ 1.0.relu_in errRate = 0.02975812368094921 _00007_ 1.0.relu_out errRate = 0.031128061935305595 _00008_ 1.1.conv1_in errRate = 0.031128061935305595 _00008_ 1.1.conv1_out errRate = 0.05052230879664421 _00009_ 1.1.bn1_in errRate = 0.05052230879664421 _00009_ 1.1.bn1_out errRate = 0.06432941555976868 _00010_ 1.1.relu_in errRate = 0.06432941555976868 _00010_ 1.1.relu_out errRate = 0.06157730147242546 _00011_ 1.1.conv2_in errRate = 0.06157730147242546 _00011_ 1.1.conv2_out errRate = 0.07610540091991425 _00012_ 1.1.bn2_in errRate = 0.07610540091991425 _00012_ 1.1.bn2_out errRate = 0.088807612657547 _00013_ 1.1.relu_in errRate = 0.06592924147844315 _00013_ 1.1.relu_out errRate = 0.05876987800002098 _00014_ 2.0.conv1_in errRate = 0.05876987800002098 _00014_ 2.0.conv1_out errRate = 0.06823940575122833 _00015_ 2.0.bn1_in errRate = 0.06823940575122833 _00015_ 2.0.bn1_out errRate = 0.09449291229248047 _00016_ 2.0.relu_in errRate = 0.09449291229248047 _00016_ 2.0.relu_out errRate = 0.0994093045592308 _00017_ 2.0.conv2_in errRate = 0.0994093045592308 _00017_ 2.0.conv2_out errRate = 0.1016751155257225 _00018_ 2.0.bn2_in errRate = 0.1016751155257225 _00018_ 2.0.bn2_out errRate = 0.12491264194250107 _00019_ 2.0.downsample.0_in errRate = 0.05876987800002098 _00019_ 2.0.downsample.0_out errRate = 0.06935565173625946 _00020_ 2.0.downsample.1_in errRate = 0.06935565173625946 _00020_ 2.0.downsample.1_out errRate = 0.0953071191906929 _00021_ 2.0.relu_in errRate = 0.1112576499581337 _00021_ 2.0.relu_out errRate = 0.10760968178510666 _00022_ 2.1.conv1_in errRate = 0.10760968178510666 _00022_ 2.1.conv1_out errRate = 0.12161954492330551 _00023_ 2.1.bn1_in errRate = 0.12161954492330551 _00023_ 2.1.bn1_out errRate = 0.149844229221344 _00024_ 2.1.relu_in errRate = 0.149844229221344 _00024_ 2.1.relu_out errRate = 0.1467655748128891 _00025_ 2.1.conv2_in errRate = 0.1467655748128891 _00025_ 2.1.conv2_out errRate = 0.1631915271282196 _00026_ 2.1.bn2_in errRate = 0.1631915271282196 _00026_ 2.1.bn2_out errRate = 0.19003801047801971 _00027_ 2.1.relu_in errRate = 0.1514744907617569 _00027_ 2.1.relu_out errRate = 0.13444244861602783 _00028_ 3.0.conv1_in errRate = 0.13444244861602783 _00028_ 3.0.conv1_out errRate = 0.15132983028888702 _00029_ 3.0.bn1_in errRate = 0.15132983028888702 _00029_ 3.0.bn1_out errRate = 0.1939898133277893 _00030_ 3.0.relu_in errRate = 0.1939898133277893 _00030_ 3.0.relu_out errRate = 0.1948491930961609 _00031_ 3.0.conv2_in errRate = 0.1948491930961609 _00031_ 3.0.conv2_out errRate = 0.20783396065235138 _00032_ 3.0.bn2_in errRate = 0.20783396065235138 _00032_ 3.0.bn2_out errRate = 0.25121334195137024 _00033_ 3.0.downsample.0_in errRate = 0.13444244861602783 _00033_ 3.0.downsample.0_out errRate = 0.14449596405029297 _00034_ 3.0.downsample.1_in errRate = 0.14449596405029297 _00034_ 3.0.downsample.1_out errRate = 0.19246523082256317 _00035_ 3.0.relu_in errRate = 0.22027139365673065 _00035_ 3.0.relu_out errRate = 0.21942254900932312 _00036_ 3.1.conv1_in errRate = 0.21942254900932312 _00036_ 3.1.conv1_out errRate = 0.2183808833360672 _00037_ 3.1.bn1_in errRate = 0.2183808833360672 _00037_ 3.1.bn1_out errRate = 0.26192155480384827 _00038_ 3.1.relu_in errRate = 0.26192155480384827 _00038_ 3.1.relu_out errRate = 0.26256048679351807 _00039_ 3.1.conv2_in errRate = 0.26256048679351807 _00039_ 3.1.conv2_out errRate = 0.2696223556995392 _00040_ 3.1.bn2_in errRate = 0.2696223556995392 _00040_ 3.1.bn2_out errRate = 0.3138006627559662 _00041_ 3.1.relu_in errRate = 0.2700217366218567 _00041_ 3.1.relu_out errRate = 0.23995940387248993 _00042_ 4.0.conv1_in errRate = 0.23995940387248993 _00042_ 4.0.conv1_out errRate = 0.25094619393348694 _00043_ 4.0.bn1_in errRate = 0.25094619393348694 _00043_ 4.0.bn1_out errRate = 0.31458067893981934 _00044_ 4.0.relu_in errRate = 0.31458067893981934 _00044_ 4.0.relu_out errRate = 0.3183543086051941 _00045_ 4.0.conv2_in errRate = 0.3183543086051941 _00045_ 4.0.conv2_out errRate = 0.3262660503387451 _00046_ 4.0.bn2_in errRate = 0.3262660503387451 _00046_ 4.0.bn2_out errRate = 0.38016772270202637 _00047_ 4.0.downsample.0_in errRate = 0.23995940387248993 _00047_ 4.0.downsample.0_out errRate = 0.24140916764736176 _00048_ 4.0.downsample.1_in errRate = 0.24140916764736176 _00048_ 4.0.downsample.1_out errRate = 0.3190271258354187 _00049_ 4.0.relu_in errRate = 0.3478679060935974 _00049_ 4.0.relu_out errRate = 0.35211071372032166 _00050_ 4.1.conv1_in errRate = 0.35211071372032166 _00050_ 4.1.conv1_out errRate = 0.3489462733268738 _00051_ 4.1.bn1_in errRate = 0.3489462733268738 _00051_ 4.1.bn1_out errRate = 0.40637654066085815 _00052_ 4.1.relu_in errRate = 0.40637654066085815 _00052_ 4.1.relu_out errRate = 0.40399107336997986 _00053_ 4.1.conv2_in errRate = 0.40399107336997986 _00053_ 4.1.conv2_out errRate = 0.4024793803691864 _00054_ 4.1.bn2_in errRate = 0.4024793803691864 _00054_ 4.1.bn2_out errRate = 0.47079962491989136 _00055_ 4.1.relu_in errRate = 0.4131106734275818 _00055_ 4.1.relu_out errRate = 0.37025389075279236 _00056_up1.up_in errRate = 0.37025389075279236 _00056_up1.up_out errRate = 0.3204860985279083 _00057_up1.conv.double_conv.0_in errRate = 0.2926936745643616 _00057_up1.conv.double_conv.0_out errRate = 0.3043934106826782 _00058_up1.conv.double_conv.1_in errRate = 0.3043934106826782 _00058_up1.conv.double_conv.1_out errRate = 0.4010332226753235 _00059_up1.conv.double_conv.2_in errRate = 0.4010332226753235 _00059_up1.conv.double_conv.2_out errRate = 0.4012783467769623 _00060_up1.conv.double_conv.3_in errRate = 0.4012783467769623 _00060_up1.conv.double_conv.3_out errRate = 0.3816970884799957 _00061_up1.conv.double_conv.4_in errRate = 0.3816970884799957 _00061_up1.conv.double_conv.4_out errRate = 0.4565137028694153 _00062_up1.conv.double_conv.5_in errRate = 0.4565137028694153 _00062_up1.conv.double_conv.5_out errRate = 0.4541933238506317 _00063_up2.up_in errRate = 0.4541933238506317 _00063_up2.up_out errRate = 0.4256354868412018 _00064_up2.conv.double_conv.0_in errRate = 0.26447153091430664 _00064_up2.conv.double_conv.0_out errRate = 0.23553727567195892 _00065_up2.conv.double_conv.1_in errRate = 0.23553727567195892 _00065_up2.conv.double_conv.1_out errRate = 0.31906989216804504 _00066_up2.conv.double_conv.2_in errRate = 0.31906989216804504 _00066_up2.conv.double_conv.2_out errRate = 0.32121607661247253 _00067_up2.conv.double_conv.3_in errRate = 0.32121607661247253 _00067_up2.conv.double_conv.3_out errRate = 0.307621031999588 _00068_up2.conv.double_conv.4_in errRate = 0.307621031999588 _00068_up2.conv.double_conv.4_out errRate = 0.36907580494880676 _00069_up2.conv.double_conv.5_in errRate = 0.36907580494880676 _00069_up2.conv.double_conv.5_out errRate = 0.3802216649055481 _00070_up3.up_in errRate = 0.3802216649055481 _00070_up3.up_out errRate = 0.3672914206981659 _00071_up3.conv.double_conv.0_in errRate = 0.18738088011741638 _00071_up3.conv.double_conv.0_out errRate = 0.18433697521686554 _00072_up3.conv.double_conv.1_in errRate = 0.18433697521686554 _00072_up3.conv.double_conv.1_out errRate = 0.21873363852500916 _00073_up3.conv.double_conv.2_in errRate = 0.21873363852500916 _00073_up3.conv.double_conv.2_out errRate = 0.20668698847293854 _00074_up3.conv.double_conv.3_in errRate = 0.20668698847293854 _00074_up3.conv.double_conv.3_out errRate = 0.20921824872493744 _00075_up3.conv.double_conv.4_in errRate = 0.20921824872493744 _00075_up3.conv.double_conv.4_out errRate = 0.25048312544822693 _00076_up3.conv.double_conv.5_in errRate = 0.25048312544822693 _00076_up3.conv.double_conv.5_out errRate = 0.2546943128108978 _00077_up4.up_in errRate = 0.2546943128108978 _00077_up4.up_out errRate = 0.2466486543416977 _00078_up4.conv.double_conv.0_in errRate = 0.09144531190395355 _00078_up4.conv.double_conv.0_out errRate = 0.10355861485004425 _00079_up4.conv.double_conv.1_in errRate = 0.10355861485004425 _00079_up4.conv.double_conv.1_out errRate = 0.12308644503355026 _00080_up4.conv.double_conv.2_in errRate = 0.12308644503355026 _00080_up4.conv.double_conv.2_out errRate = 0.10969457775354385 _00081_up4.conv.double_conv.3_in errRate = 0.10969457775354385 _00081_up4.conv.double_conv.3_out errRate = 0.10458693653345108 _00082_up4.conv.double_conv.4_in errRate = 0.10458693653345108 _00082_up4.conv.double_conv.4_out errRate = 0.12041313201189041 _00083_up4.conv.double_conv.5_in errRate = 0.12041313201189041 _00083_up4.conv.double_conv.5_out errRate = 0.11157391220331192 _00084_outc.conv_in errRate = 0.11157391220331192 _00084_outc.conv_out errRate = 0.04279392957687378 _00085_sigmoid_in errRate = 0.04279392957687378 _00085_sigmoid_out errRate = 0.4959424138069153
热门帖子
精华帖子