hua-wu-que

question about backward parameter

1

Hi! In deform_conv_cuda.c line 206, why you recompute the columns by calling "deformable_im2col( THCState_getCurrentStream(state), THCudaTensor_data(state, input_n), THCudaTensor_data(state, offset_n), nInputPlane, inputHeight, inputWidth, kH, kW, padH, padW, dH, dW, dilationH, dilationW, im2col_step,...

wrong demo result

6

There is double sided effect in the extracted mesh as well as point cloud in the demo code. The other side is not correct, how can this be fixed? Thank...

Hi! I tried twice to launch the training following your setup. But the embedding will blow up after around 60k steps. (the embedding goes to hundreds in magnitude). Did you...

an illegal memory access was encountered

1

Hi! I could not finish the training because I encountered this error non-indefinitely(usually after trained for several epochs). I use a small batch size so I don't think it is...

Hi, I found following potential waste computation: the for loop goes from -2 -> 2 "for (int dy = -2; dy

Can anyone confirm this line is a bug for stride!=1 case? in line 162 of file dcn_v2_im2col_cuda.cu: "const float *data_offset_ptr = data_offset + (b_col * deformable_group + deformable_group_index) * 2...

has anyone trained on all category?

2

Hi! I wonder if anyone has experience train one model on all category instead of one model for each category? What are the performance be like? Thank you!

influence of diffrent normalization

2

Hi there, I want to ask your advice regarding the normalization layers. I am implementing a network similar to yours, but I found if I use batch norm layer, during...

does anyone ever compared the speed of this deformed 3d conv with vanilla 3dconv in pytorch? Is there huge(10x) difference?

through out error after launch training

1

The training can be launch but after some time it will always broken. The output looks like this: epoch 1 step 100 loss 0.38945767 - time per batch 0.7466 epoch...