When using TensorRT to deploy Resnet34, it reports error as below:
"
Caffe Paser: Relu negative slope is not supported
error parsing layer type Relu index 4
"
Doesn’t TensorRT support leaky Relu now?
In our models, leaky Relu is widely used
When using TensorRT to deploy Resnet34, it reports error as below:
"
Caffe Paser: Relu negative slope is not supported
error parsing layer type Relu index 4
"
Doesn’t TensorRT support leaky Relu now?
In our models, leaky Relu is widely used
Hi,
TensorRT will support Leaky relu from v5.1.
If acceptable, you can update your lrelu to relu+scale.
Thanks.
Hi AastaLLL,
So when will v5.1 be released?
Leaky Relu does better in our dataset and models comparing to Relu, so will not change to Relu
Hi,
You can use relu + scale to approximate.
In general, it can give you the same result.
Thanks.
Hi
(i have replied twice, but i don’t know why it says error and delete my reply!!! )
The thing is that we widely use leaky Relu in our models, according to your ways, we will modify it to Relu and re-train the models to check the results, then modify it back to leaky Relu in your next version, our guys think it is not worth the effort.
Another thing is that other platforms support leaky Relu
Hi,
This should not be an issue.
Leaky relu, relu and scaling are all weight-free layers.
It is controlled by the hyperparameter.
Thanks.
Hi,
1 )So what is the scale in relu + scale? we talked, but don’t understand
Please provide specific solution
Hi,
1. For example, y = max(0.01x,x)
yi = xi if xi >0 -> generate this with standard relu
xi*ai if xi <0 -> generate this with scale layer, scaling factor is ai
2. My colleague have sent you a private message. Please check.
Thanks.
Hi
We have tied max(x, a*x), now the model can be supported, we will check the final results of the model later, thanks