MaybeShewill-CV

MNN-2.0.0 release CUDA bacnend outputs error

19

# 平台(如果交叉编译请再附上交叉编译目标平台): # Platform(Include target platform as well if cross-compiling): Ubuntu 16.04 GCC-8.2 CUDA-10.2 MNN-Version: MNN-2.0.0 release # 编译方式: # Compiling Method -DCMAKE_BUILD_TYPE=Release -DMNN_CUDA=ON Model's output was wrong with MNN-2.0.0...

Could you please release more information about the training details such as batch_size, learning rate , epoch nums if it matters nothing? Thanks a lot

关于Generalization to large models章节中的参数问题

14

首先感谢大佬 @ycszen 的工作：） 文章中有些我觉得写的不是很清楚的地方想问一下，关于generalization to large models章节中提到的两个参数α和d。  关于α,我觉得指的是进入segmentation heads模块之前的channel expansion倍数，如下图所示  那么这个d参数指的是什么呢，文章中说d参数控制的是模型的深度，那么这个d是指代上文中的什么参数呢，是module repeat的次数吗. 还有点疑惑就是table5下方的描述α控制模型的channel capacity而d控制layer nums.然后下文中说α是width multiplier，d是depth multiplier。请问这个地方上下文的表述是否一致。  这两个参数该如何理解更为准确呢。还请大佬有时间不吝赐教...

Hi I'm very interested in your great work. I wonder whether the training dataset be released in the future?:)

onnx support

1

I wonder if the excellent model is able to converted into onnx ?

metric depth estimation vs metric3d

1

Does the method have detailed results on metric measurements, especially in comparison with the metric3d method?

Could you please tell me the ipm function need external camera parameters?

mobile-sam mnn inference performance

2

Wonder if mobile-sam can reach the inference performance mentioned in origin paper using mnn backend. As mentioned in origin paper encoding part cost 8ms, decoding cost 4ms. I've tested it...