基于Xavier（jetson）完成segnet模型训练

已于 2022-04-22 15:49:58 修改
阅读量1.5k 收藏 18
点赞数 1
分类专栏： segnet xavier jetson 文章标签： 深度学习 人工智能
于 2020-11-27 11:43:03 首次发布
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/weixin_44316541/article/details/110222670
版权
segnet 同时被 3 个专栏收录
1 篇文章 0 订阅
订阅专栏
xavier
1 篇文章 0 订阅
订阅专栏
jetson
1 篇文章 0 订阅
订阅专栏

基于Xavier（jetson）完成segnet模型训练

本文基于源码：https://github.com/dusty-nv/jetson-inference，完成segnet语义分割模型的开发（training部分）

环境

硬件：Xavier
JettPack Version：JetPack 4.4
docker container：dustynv/jetson-inference:r32.4.3

训练流程

1、进入源码训练目录

cd /jetson-inference/python/training/segmentation

2、修改数据导入函数（基于VOC格式,数据制作脚本更新中）

def get_dataset(path, image_set, transform):
    """
    load voc datasets
    name: voc
    path: path of datasets
    image_set: "train"
    transform: get_transform(train=True, resolution=resolution)
    """
    num_classes = 2
    ds = VOCSegmentationLoad(path, image_set=image_set, transforms=transform)
    return ds, num_classes

3、模型训练

3.1 源码采用的fcn_resnet18, 但是torchversion只支持fcn_resnet50、fcn_resnet101，所以这里需要自定义fcn_resnet18:
源码支持模型：

__all__ = [ 'fcn_resnet50', 'fcn_resnet101', 'deeplabv3_resnet50', 'deeplabv3_resnet101']

model_urls = {
    'fcn_resnet50_coco': 'https://download.pytorch.org/models/fcn_resnet50_coco-1167a1af.pth',
    'fcn_resnet101_coco': 'https://download.pytorch.org/models/fcn_resnet101_coco-7ecb50ca.pth',
    'deeplabv3_resnet50_coco': 'https://download.pytorch.org/models/deeplabv3_resnet50_coco-cd0a2569.pth',
    'deeplabv3_resnet101_coco': 'https://download.pytorch.org/models/deeplabv3_resnet101_coco-586e9e4e.pth',
}

这里需要自定义模型fcn_resnet18(路径：/usr/local/lib/python3.6/dist-packages/torchvision-0.7.0a0+78ed10c-py3.6-linux-aarch64.egg/torchvision/models/segmentation/segmentation.py)：

def fcn_resnet18(pretrained=False, progress=True, num_classes=21, aux_loss=None, **kwargs):
        """Constructs a Fully-Convolutional Network model with a ResNet-50 backbone.
        Args:
            pretrained (bool): If True, returns a model pre-trained on COCO train2017 which
            contains the same classes as Pascal VOC
            progress (bool): If True, displays a progress bar of the download to stderr
        """
        return _load_model('fcn', 'resnet18', pretrained, progress, num_classes, aux_loss, **kwargs)

resnet18采用Basicblock， resnet101采用Bottleneck，所以需要修改函数:def _segm_resnet(name, backbone_name, num_classes, aux, pretrained_backbone=True):

    backbone = resnet.__dict__[backbone_name](
        pretrained=pretrained_backbone,
        replace_stride_with_dilation=[False, True, True])

改为：

    backbone = resnet.__dict__[backbone_name](
        pretrained=pretrained_backbone,
        replace_stride_with_dilation=[False, False, False])

resnet18输出512，resnet50输出2048，修改

inplanes = 2048

为

inplanes = 512

3.2 回到/jetson-inference/python/training/segmentation目录，开始训练

python3 train.py path_voc_dataset --width=512 --height=320 -b=1 -j=8 --model-dir=models_dir

训练开始

pytorch-segmentation/datasets/__init__.py
Not using distributed mode
Namespace(arch='fcn_resnet18', aux_loss=False, batch_size=1, data='/jetson-inference/data/label/ts_data/', dataset='voc', device='cuda', dist_url='env://', distributed=False, epochs=30, height=320, lr=0.01, model_dir='models/', momentum=0.9, pretrained=False, print_freq=10, resolution=320, resume='', test_only=False, weight_decay=0.0001, width=512, workers=8, world_size=1)
=> training with dataset: 'voc' (train=714, val=179)
=> training with resolution: 512x320, 3 classes
=> training with model: fcn_resnet18
arch:fcn_resnet18
.
.
.
  target = torch.as_tensor(np.asarray(target), dtype=torch.int64)
Test:  [  0/179]  eta: 0:01:05    time: 0.3666  data: 0.3359  max mem: 286
Test:  [100/179]  eta: 0:00:02    time: 0.0227  data: 0.0035  max mem: 286
Test: Total time: 0:00:04
global correct: 98.7
average row correct: ['99.7', '18.9', '65.6']
IoU: ['98.7', '15.7', '53.9']
mean IoU: 56.1
saved checkpoint to:  models/model_29.pth  (56.118% mean IoU, 98.651% accuracy)

生成pytorch模型：

root@jykj-desktop:/jetson-inference/python/training/segmentation# ls models/
model_0.pth   model_14.pth  model_2.pth   model_25.pth  model_4.pth  model_best.pth
model_1.pth   model_15.pth  model_20.pth  model_26.pth  model_5.pth
model_10.pth  model_16.pth  model_21.pth  model_27.pth  model_6.pth
model_11.pth  model_17.pth  model_22.pth  model_28.pth  model_7.pth
model_12.pth  model_18.pth  model_23.pth  model_29.pth  model_8.pth
model_13.pth  model_19.pth  model_24.pth  model_3.pth   model_9.pth

4、 模型转化：

首先，torchversion的fcn具有上采样逻辑，作者的onnx模型无上采样操作，所以，首先需要修改torchversion源码，去掉上采样（否者启动模型时会报错）：
源码路径

/usr/local/lib/python3.6/dist-packages/torchvision-0.7.0a0+78ed10c-py3.6-li
nux-aarch64.egg/torchvision/models/segmentatio/_utils.py

修改def _SimpleSegmentationModel(nn.Module):

def forward(self, x):
        input_shape = x.shape[-2:]
        # contract: features is a dict of tensors
        features = self.backbone(x)

        result = OrderedDict()
        x = features["out"]
        x = self.classifier(x)
        x = F.interpolate(x, size=input_shape, mode='bilinear', align_corners=False)
        result["out"] = x
        if self.aux_classifier is not None:
            x = features["aux"]
            x = self.aux_classifier(x)
            x = F.interpolate(x, size=input_shape, mode='bilinear', align_corners=False)
            result["aux"] = x

        return result

修改为

def forward(self, x):
        input_shape = x.shape[-2:]
        # contract: features is a dict of tensors
        features = self.backbone(x)

        result = OrderedDict()
        x = features["out"]
        x = self.classifier(x)
        if self.train_status:   #训练的时候需要上采样，模型转换的时候去掉
            x = F.interpolate(x, size=input_shape, mode='bilinear', align_corners=False)
        result["out"] = x

        if self.aux_classifier is not None:
            x = features["aux"]
            x = self.aux_classifier(x)
            x = F.interpolate(x, size=input_shape, mode='bilinear', align_corners=False)
            result["aux"] = x

        return result

回到training目录：

/jetson-inference/python/training/segmentation

修改onnx_export.py文件

# model = models.segmentation.__dict__[arch](num_classes=num_classes,
#                                            aux_loss=None,
#                                            pretrained=False,
#                                            export_onnx=True #去掉，torchversion源码不识别
)
model = models.segmentation.__dict__[arch](num_classes=num_classes,
                                           aux_loss=None,
                                           pretrained=False,
                                           train_status=False)	# 这个是去掉长采样的status，可以自定义

开始模型转换

python3 onnx_export.py --input=path_of_model.pth --output=output_model.onnx

完成转化：

Namespace(input='models/model_best.pth', model_dir='', output='fcn_resnet18.onnx')
running on device cuda:0
loading checkpoint:  models/model_best.pth
using model:  fcn_resnet18
num classes:  3
input size:  512x320
exporting model to ONNX...
graph(%input_0 : Float(1:491520, 3:163840, 320:512, 512:1),
.
.
.
  %output_0 : Float(1:480, 3:160, 10:16, 16:1) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%197, %classifier.4.weight, %classifier.4.bias) # /usr/local/lib/python3.6/dist-packages/torch/nn/modules/conv.py:416:0
  return (%output_0)

model exported to:  fcn_resnet18.onnx

至此，xavier上的segnet模型训练与转化部分完成，可用于模型训练与测试
模型检测结果（自定义输出结构）：

{"frame_id": 199, "time": "2020-11-27 01:57:48", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 200, "time": "2020-11-27 01:57:48", "objects": [{"class_id": 1, "name": "something", "pixel_num": 160}]}
{"frame_id": 201, "time": "2020-11-27 01:57:48", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 202, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 203, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 204, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 160}]}
{"frame_id": 205, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 206, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 2}, {"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 207, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 160}]}
{"frame_id": 208, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 209, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 210, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 211, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 212, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 213, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 160}]}
{"frame_id": 214, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 215, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 216, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 4}]}
{"frame_id": 217, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 218, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 219, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 220, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 3}, {"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 221, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 222, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 223, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 3}, {"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 224, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 225, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 1, "name": "something", "pixel_num": 3}, {"class_id": 2, "name": "machine", "pixel_num": 3}]}
{"frame_id": 226, "time": "2020-11-27 01:57:49", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 227, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 228, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 229, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 4}]}
{"frame_id": 230, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 1, "name": "something", "pixel_num": 160}]}
{"frame_id": 231, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 232, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 233, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 234, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 235, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 4}]}
{"frame_id": 236, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 237, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 238, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 239, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 240, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 241, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 242, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 2}]}
{"frame_id": 243, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 244, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 1, "name": "something", "pixel_num": 3}, {"class_id": 2, "name": "machine", "pixel_num": 1}]}
{"frame_id": 245, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 1, "name": "something", "pixel_num": 160}]}
{"frame_id": 246, "time": "2020-11-27 01:57:50", "objects": [{"class_id": 1, "name": "something", "pixel_num": 1}, {"class_id": 2, "name": "machine", "pixel_num": 3}]}

5、QA

1、training报错

AttributeError: module 'torchvision.models.segmentation.segmentation' has no attribute 'fcn_resnet18'

解决方案：修改源码，添加fcn_resnet18，参考3.1
2、training 报错

NotImplementedError: Dilation > 1 not supported in BasicBlock

解决方案：由于resnet18采用basicblock， 修改 replace_stride_with_dilation=[False, False, False])
3、training报错：

RuntimeError: input and target batch or spatial sizes don't match: target [1 x 320 x 512], input [1 x 3 x 10 x 16] at /media/nvidia/WD_NVME/PyTorch/JetPack_4.4/GA/pytorch-v1.6.0/aten/src/THCUNN/generic/SpatialClassNLLCriterion.cu:2

解决方案：训练时需要添加上采样，模型转化时去掉上采样

        if self.train_status:   #训练的时候需要上采样，模型转换的时候去掉
            x = F.interpolate(x, size=input_shape, mode='bilinear', align_corners=False)

4、推理报错

ERROR: builtin_op_importers.cpp:3460 In function importUpsample:
[8] Assertion failed: scales_input.is_weights()
[TRT]    failed to parse ONNX model '/usr/local/bin/networks/FCN-ResNet18-Pascal-VOC-512x320/fcn_resnet18.onnx'
[TRT]    device GPU, failed to load networks/FCN-ResNet18-Pascal-VOC-512x320/fcn_resnet18.onnx
[TRT]    segNet -- failed to load.
jetson.inference -- segNet failed to load network
Exception in thread Thread-1:
Traceback (most recent call last):
 File "/usr/lib/python3.6/threading.py", line 916, in _bootstrap_inner
   self.run()
 File "/usr/lib/python3.6/threading.py", line 864, in run
   self._target(*self._args, **self._kwargs)
 File "./segnet.py", line 151, in process
   net = jetson.inference.segNet(opt.network, sys.argv)
Exception: jetson.inference -- segNet failed to load network

解决方案：模型转换时，去掉上采样
加上上采样的onnx网络结构：

  %201 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%200, %classifier.0.weight)
 %202 = BatchNormalization[epsilon = 9.99999974737875e-06, momentum = 0.899999976158142](%201, %classifier.1.weight, %classifier.1.bias, %classifier.1.running_mean, %classifier.1.running_var)
 %203 = Relu(%202)
 %204 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%203, %classifier.4.weight, %classifier.4.bias)
 %205 = Unsqueeze[axes = [0]](%131)
 %206 = Unsqueeze[axes = [0]](%134)
 %207 = Concat[axis = 0](%205, %206)
 %208 = Constant[value = <Tensor>]()
 %209 = Cast[to = 1](%207)
 %210 = Shape(%204)
 %211 = Slice[axes = [0], ends = [9223372036854775807], starts = [2]](%210)
 %212 = Cast[to = 1](%211)
 %213 = Div(%209, %212)
 %214 = Concat[axis = 0](%208, %213)
 %output_0 = Upsample[mode = 'linear'](%204, %214)
 return %output_0

去掉上采样的onnx网络结构：

 %194 = Relu(%193)
 %195 = Conv[dilations = [1, 1], group = 1, kernel_shape = [3, 3], pads = [1, 1, 1, 1], strides = [1, 1]](%194, %classifier.0.weight)
 %196 = BatchNormalization[epsilon = 9.99999974737875e-06, momentum = 0.899999976158142](%195, %classifier.1.weight, %classifier.1.bias, %classifier.1.running_mean, %classifier.1.running_var)
 %197 = Relu(%196)
 %output_0 = Conv[dilations = [1, 1], group = 1, kernel_shape = [1, 1], pads = [0, 0, 0, 0], strides = [1, 1]](%197, %classifier.4.weight, %classifier.4.bias)
 return %output_0
}
打豆豆的蓝猫
关注
专栏目录
NVIDIA Jetson Xavier NX 实现官方Jeston-inference深度学习样例
zbb297918657的博客
05-29 8856
一、jetson-inference相关项目组件的下载 首先附上官方提供的jetson_inference项目文件的Github仓库地址，大家可以自行前往下载 jetson-inference下载地址：https://github.com/dusty-nv/jetson-inference 另外，在该地址下包含详细的官方关于这个项目文件的相关介绍和使用方法，大家如果觉得我说的啰嗦，可直接前往该地址一步一步的安装调试。 还要多说一点的是，很多人也会选择直接在终端执行git操作进行克隆jetson-infere
YolactEdge：边缘设备上的实时实例分割（Xavier: 30 FPS, RTX 2080 Ti：170 FPS）
阿木寺的博客
12-23 1242
带有ResNet-101的YolactEdge在Jetson AGX Xavier上的速度高达30.8 FPS，在RTX 2080 Ti上的速度为172.7 FPS，AP性能超强！速度是目前主流方法的3-5倍，代码于1小时前刚刚开源！ 注：文末附【图像分割】学习交流群 YolactEdge: Real-time Instance Segmentation on the Edge (Jetson AGX Xavier: 30 FPS, RTX 2080 Ti: 170 FPS) 作者单位：UC Davis,
MXNet之模型训练配置
weixin_39451323的博客
06-27 1973
模型训练配置1.问题定义2参数及训练配置2.1 参数初始化2.2优化函数设置2.3 模型保存2.4 训练日志保存2.5 选择并定义评价指标2.6 多GPU训练3. 迁移学习4.断点训练5. minist项目演练 一个完整的训练代码主要由3部分组成,数据读取,网络结构搭建和模型训练配置. 训练模型之前要定义问题,这个过程想当于为今后的算法设计,优化确定一个总体方向.定义问题需要结合项目需求,数据特...
【pytorch | bug | 版本不兼容】：RuntimeError: Expected 2D (unbatched) or 3D (batched) input to conv1d, but g
你的问题在于，读书太少而想得太多。
05-01 6716
解决【pytorch | bug | 版本不兼容】：RuntimeError: Expected 2D (unbatched) or 3D (batched) input to conv1d, but got input of size: [4, 128, 160, 120]
超越英伟达的，不会是另一款GPU？这家深圳公司发布全球首款数据流AI芯片
AI科技大本营
06-25 1024
2020年6月23日，鲲云科技在深圳举行产品发布会，发布全球首款数据流AI芯片CAISA，定位于高性能AI推理，已完成量产。鲲云通过自主研发的数据流技术在芯片实测算力上实现了技术突破，较...
Jetson Inference安装测试
sjjzym1314
02-18 1936
jetson inference、ros deeplearing的安装及分割模型的测试
Jetson-AGX-Xavier上的量化深层神经网络：如何创建，训练和量化网络，然后将其集成到后期图像处理中并生成用于定位Jetson AGX Xavier的CUDA C ++代码
02-05
在本文中，我们将深入探讨如何在NVIDIA Jetson AGX Xavier平台上进行深度学习模型的创建、训练、量化以及集成到图像处理系统中。Jetson AGX Xavier是一款强大的嵌入式计算平台，专为边缘计算和AI应用设计，具有高效...
基于Jetson Xavier nx和Jetson nano部署yolov5工地头盔和人头检测系统C++源码含运行说明.zip
最新发布
10-09
基于Jetson Xavier nx和Jetson nano部署yolov5工地头盔和人头检测系统C++源码含运行说明.zip基于Jetson Xavier nx和Jetson nano部署yolov5工地头盔和人头检测系统C++源码含运行说明.zip基于Jetson Xavier nx和Jetson...
Jetson Xavier nx 和 Jetson nano 中 Yolov5 头盔检测的 Python 训练和推理实现
04-01
总结，本文详细介绍了在Jetson Xavier NX和Jetson Nano上使用Yolov5进行头盔检测的流程，涵盖了从环境配置、数据准备到模型训练和推理的整个过程。这个话题结合了计算机视觉、深度学习和嵌入式系统的知识，为开发...
基于jetson AGX Xavier的GLIR相机ros驱动spinnaker
04-01
通过上述步骤，我们可以将FLIR GLIR相机集成到基于Jetson AGX Xavier的ROS系统中，利用Spinnaker SDK的强大功能进行高效的图像采集和处理。这为各种机器人和自动化应用提供了坚实的基础，例如视觉导航、目标检测和...
jetpack 4.6 刷机 英伟达jetson agx Xavier 完整刷机记录
03-01
本文是关于如何使用NVIDIA JetPack 4.6对英伟达Jetson AGX Xavier开发板进行刷机的详细教程。刷机过程中涉及到的主要工具是JetPack SDK Manager，这是一个用于管理NVIDIA嵌入式平台软件开发工具包的工具。整个过程...
ValueError: Expected input batch_size (1944) to match target batch_size (1962).
claywt的博客
11-08 1356
训练数据报错，说输入输出batch_size不一致，这里面有可能是tokenizer分词导致的，我这里模型要求输入和输出的句子长度一致，但是如果仅仅用len判读input和target两句话是不够的。如果是单纯的判断两句话长度是一致的，但是tokenizer分词后长度是不一致的。可以通过tokenizer分词过滤句子。
PyTorch(总)——PyTorch遇到令人迷人的BUG与记录
热门推荐
hudongloop的博客
06-28 12万+
这篇博客就用来记录在使用pytorch时遇到的BUG，虽然年纪大了，但是调出BUG还是令人兴奋^_^! BUG1： 在使用NLLLoss()激活函数时，NLLLoss用来做n类分类的，一般最后一层网络为LogSoftmax，如果其他的则需要使用CrossEntropyLoss。其使用格式为：loss(m(input), target)，其中input为2DTensor大小为（minibatch...
Pytorch学习笔记（二）使用Pytorch的常见错误汇总
ROM的博客
07-31 7万+
那些年我们一起踩过的坑 _, pred = output.data.cpu().topk(1, dim=1) RuntimeError: cuda runtime error (59) : device-side assert triggered at /pytorch/aten/src/THC/generic/THCTensorCopy.c:70...
成功解决RuntimeError: multi-target not supported at /pytorch/aten/src/THCUNN/generic/
梦坠凡尘
03-17 5994
计算交叉熵时报错： RuntimeError: multi-target not supported at /pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:15 维度不匹配导致 input, target的维度与F.cross_entropy()函数要求的维度不匹配 解决方法： F.cross_entropy( input, target.squeeze() ) ...
RuntimeError: 1only batches of spatial targets supported (non-empty 3D tensors) but got targets of s
m0_46378271的博客
10-31 5535
目录 在运行UNet3+进行多分类训练的时候报错 RuntimeError: 1only batches of spatial targets supported (non-empty 3D tensors) but got targets of size xxx 1. 修改train.py 2. 修改predict.py 3. 修改eval.py 然后运行下面语句就可以开始训练了： 在运行UNet3+进行多分类训练的时候报错 RuntimeError: 1only batches of .
记录Pytorch中遇到的问题
qq_36228172的博客
03-01 1万+
在feature层到classifier层中，若出现如下错误： 则需要改变FC层的入口参数，如此图中可以改为： x = x.view(out.size(0), -1) self.linear = nn.Linear(320, 10) # 320为入口参数 ...
评论 22
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值