search

chenzhi1992 / TensorRT-SSD

Use TensorRT API to implement Caffe-SSD， SSD（channel pruning）， Mobilenet-SSD
251 stars 84 forks source link

Concat layer and soffmax layer #15

Open quocbh opened 6 years ago

quocbh commented 6 years ago

I'm newbie in TensorRT. I want to deploy SSD on Jetson TX2 using tensor RT. I don't understand why reimplement Concat layer in tensor RT. Can you explain for me. Same question for Softmax Layer. And I need reimplement another layer?

chenzhi1992 commented 6 years ago

Beacause TensorRT only support Concat layer across the channel dimension. It is same for Softmax layer. You can learn more information from tensorRT3.0 document.

quocbh commented 6 years ago

I only need reimplement concat and softmax or another layer?

linux-devil commented 6 years ago

Only Softmax layer @quocbh

quocbh commented 6 years ago

@linux-devil you can run it? You can share some experience. Thanks

JullyZ commented 5 years ago

Beacause TensorRT only support Concat layer across the channel dimension. It is same for Softmax layer. You can learn more information from tensorRT3.0 document.

你好，我可以理解这句话的意思是当axis=1的时候，我们可以用自带的concat.当axis不等于1的时候，需要自己实现。所以在ssd的priorbox concat的时候，需要用自己实现的层。但是在“mbox_loc”中，axis=1, 按道理来讲，我们可以用tensorRT自带的concat去解析。但是，在实验的时候，我把您prototxt中关于“mbox_loc”的type改为“Concat”,在Plugincomplement中，把和mbox_loc相关的注释掉，代码跑不通，模型解析不了，您知道为什么吗？谢谢。