chinese_ocr icon indicating copy to clipboard operation
chinese_ocr copied to clipboard

如何修改batchsize大小,找了半天没找到

Open ziyuwzf opened this issue 6 years ago • 5 comments

这是densenet测试了十几万张图片之后报的错 2019-01-23 19:51:42.943120: F tensorflow/stream_executor/cuda/cuda_dnn.cc:542] Check failed: cudnnSetTensorNdDescriptor(handle_.get(), elem_type, nd, dims.data(), strides.data()) == CUDNN_STATUS_SUCCESS (3 vs. 0)batch_descriptor: {count: 1 feature_map_count: 128 spatial: 4 0 value_min: 0.000000 value_max: 0.000000 layout: BatchDepthYX} Aborted (core dumped)

ziyuwzf avatar Jan 24 '19 03:01 ziyuwzf

@ziyuwzf faster-rcnn算法,据我了解,除了mmdetection和maskrcnn-benchmark,其他的似乎都没能实现批量训练吧,都是默认batchsize=1. 这是我的粗浅了解,未必正确

fmscole avatar Oct 10 '19 03:10 fmscole

这是densenet测试了十几万张图片之后报的错 2019-01-23 19:51:42.943120: F tensorflow/stream_executor/cuda/cuda_dnn.cc:542] Check failed: cudnnSetTensorNdDescriptor(handle_.get(), elem_type, nd, dims.data(), strides.data()) == CUDNN_STATUS_SUCCESS (3 vs. 0)batch_descriptor: {count: 1 feature_map_count: 128 spatial: 4 0 value_min: 0.000000 value_max: 0.000000 layout: BatchDepthYX} Aborted (core dumped)

您好。想请问下您解决这个问题了吗? 谢谢

xxxxxxxiao avatar Jan 09 '20 09:01 xxxxxxxiao

这个我也不知道:)

---原始邮件--- 发件人: "xxxxxxxiao"<[email protected]> 发送时间: 2020年1月9日(周四) 下午5:31 收件人: "YCG09/chinese_ocr"<[email protected]>; 抄送: "fmscole"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [YCG09/chinese_ocr] 如何修改batchsize大小,找了半天没找到 (#165)

这是densenet测试了十几万张图片之后报的错 2019-01-23 19:51:42.943120: F tensorflow/stream_executor/cuda/cuda_dnn.cc:542] Check failed: cudnnSetTensorNdDescriptor(handle_.get(), elem_type, nd, dims.data(), strides.data()) == CUDNN_STATUS_SUCCESS (3 vs. 0)batch_descriptor: {count: 1 feature_map_count: 128 spatial: 4 0 value_min: 0.000000 value_max: 0.000000 layout: BatchDepthYX} Aborted (core dumped)

您好。想请问下您解决这个问题了吗? 谢谢

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

fmscole avatar Jan 09 '20 10:01 fmscole

@ziyuwzf @xxxxxxxiao 您好,请问您解决这个问题了嘛?谢谢!

Jaiy avatar Oct 29 '20 06:10 Jaiy

这是densenet测试了十几万张图片之后报的错 2019-01-23 19:51:42.943120: F tensorflow/stream_executor/cuda/cuda_dnn.cc:542] Check failed: cudnnSetTensorNdDescriptor(handle_.get(), elem_type, nd, dims.data(), strides.data()) == CUDNN_STATUS_SUCCESS (3 vs. 0)batch_descriptor: {count: 1 feature_map_count: 128 spatial: 4 0 value_min: 0.000000 value_max: 0.000000 layout: BatchDepthYX} Aborted (core dumped)

我想主要有两个原因: 1、单gpu训练、单gpu推理,主要原因是输入的图片太小,例如图片的height resize到32,width大小是1,就出现bug了 2、单gpu训练保存的模型,作为预训练模型用多gpu进行训练

JinMana avatar Mar 31 '21 04:03 JinMana