Peng Chen comments

Results 72 comments of


                                            Peng Chen

ABCNet deployment

hi, @shuangyichen From my perspective, I suggest not exporting the BezierAlign into onnx. Instead, just export other code with normal operators such as conv/fc/bn/relu/interp. When inference on embeded platform or...

ABCNet deployment

Yes. Though it involves with some extra effort, it somehow provides a feasible way.

when I train RCAN,something wrong:RuntimeError: Expected 4-dimensional input for 4-dimensional weight 3 3 1, but got 3-dimensional input of size [1, 184, 270] instead

Same error on my machine

any example on the project?

Hi, @xuanyoya I notice you add a readme in the arch subfolder. For the "capacity": "[1st level buffer size, 2nd level buffer size, ...]", does it mean the 1st level...

AQD

Good. And PRs are welcomed.

loss become infinite while training quant models

HI, @RaidenE1 There maybe many reasons for the loss NAN. I also frequently met such a problem. My general ways are: 1. empoly single GPU for training to see any...

Can I use tensorRT to accelerate Fcos model trained use pytorch?

This problem seem to be very likely casued by the upsample operator. When employing interpolate/upsample operator, TensorRT requires to explict width/height rather than scale factors are used. To solver the...

Can I use tensorRT to accelerate Fcos model trained use pytorch?

@jinfagang You mean the extra output from the FCOS branch besides the centerness/cls/bbox? For any output blobs, you can fetch them by using the name as the index.

Can I use tensorRT to accelerate Fcos model trained use pytorch?

@jinfagang The ROI layers are not exported. Current ouput contains the logits, centerness, bbox_reg, top_feat from the FCOS branch and the basis from the basis_module branch. To generate the final...

Can I use tensorRT to accelerate Fcos model trained use pytorch?

There are limited number of boundingbox that need to pass through the ROI process. As far as I can remeber, the filter boundingbox is generally no more than 10 boxes...