Closed hbo-lambda closed 6 months ago
nemonameless
commented
1 year ago coco数据集上均没有此类似报错，建议重新检查下数据集制作，排除空标注或不合理标注框，同时reader里去除mosaic再试试
mosaic_epoch: &mosaic_epoch -1
worker_num: 4
TrainReader:
sample_transforms:
- Decode: {}
- MosaicPerspective: {mosaic_prob: 0.0, target_size: *input_size}
.......https://github.com/PaddlePaddle/PaddleYOLO/issues/43
此外单卡训总bs太小了，gt是不是每张图也很少？这样容易mosaic后无框。建议扩大bs，模型不用选p6，你单卡训精度都不一定有常规p5模型开大bs训来的高
hbo-lambda
commented
1 year ago 43
此外单卡训总bs太小了，gt是不是每张图也很少？这样容易mosaic后无框。建议扩大bs，模型不用选p6，你单卡训精度都不一定有常规p5模型开大bs训来的高
为了寻找报错的原因，所以将bs搞成了1。我用的是tt100k，它的图像是2048*2048的，道路标识又有点小，所以用了p6。我是coco数据集的，可以运行。依照提示修改了mosaic_prob后，用自己的数据集就会卡住。换成voc数据格式后，也会报那个错。
检查了数据集后，没有发现空标注，xmin、ymin、xmax、ymax的数据也合规。贴一个xml数据
<annotation>
<folder>train</folder>
<filename>36319.jpg</filename>
<size>
<width>2048</width>
<height>2048</height>
<depth>3</depth>
</size>
<segmented>0</segmented>
<object>
<name>i2</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>1550</xmin>
<ymin>933</ymin>
<xmax>1567</xmax>
<ymax>952</ymax>
</bndbox>
</object>
<object>
<name>i4</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>1273</xmin>
<ymin>912</ymin>
<xmax>1296</xmax>
<ymax>934</ymax>
</bndbox>
</object>
<object>
<name>p10</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>1496</xmin>
<ymin>992</ymin>
<xmax>1509</xmax>
<ymax>1006</ymax>
</bndbox>
</object>
<object>
<name>pl40</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>1513</xmin>
<ymin>932</ymin>
<xmax>1529</xmax>
<ymax>951</ymax>
</bndbox>
</object>
<object>
<name>pl50</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>1460</xmin>
<ymin>933</ymin>
<xmax>1476</xmax>
<ymax>950</ymax>
</bndbox>
</object>
<object>
<name>pl70</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>1440</xmin>
<ymin>929</ymin>
<xmax>1457</xmax>
<ymax>949</ymax>
</bndbox>
</object>
<object>
<name>pn</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>1478</xmin>
<ymin>932</ymin>
<xmax>1497</xmax>
<ymax>953</ymax>
</bndbox>
</object>
<object>
<name>pne</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>1140</xmin>
<ymin>906</ymin>
<xmax>1164</xmax>
<ymax>930</ymax>
</bndbox>
</object>
<object>
<name>im</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>1529</xmin>
<ymin>935</ymin>
<xmax>1546</xmax>
<ymax>954</ymax>
</bndbox>
</object>
</annotation>{
"images": [{"file_name": "36319.jpg", "width": 2048, "height": 2048, "id": 1761}],
"annotations": [{"segmentation": [[1550, 933, 1550, 952, 1567, 952, 1567, 933]], "area": 323, "iscrowd": 0, "image_id": 1761, "bbox": [1550, 933, 17, 19], "category_id": 8, "id": 4781, "ignore": 0}]
}
"category": [{"supercategory": "none", "id": 1, "name": "i1"},....]方便发下tt100k数据集吗，或者可以使用别的ppyoloe+或rtdetr模型
问题确认 Search before asking
请提出你的问题 Please ask your question
配置文件修改：
因为报错，所以在358行代码中添加了打印
第一个batch有值，第二个batch就成了[]，好纳闷
执行的命令:
有可能是我有遗漏，也许是本来就有bug。同样的数据集，换成 python tools/train.py -c configs/ppyoloe/ppyoloe_plus_crn_s_80e_coco.yml 就可以执行训练