Experiment on DenseNet with dilation

[Xiangyu-CAS]

Hi, recently, I have carried out some experiment based on DenseNet backbone. In order to produce feature map at stride 8, dilation or a trous trick was used.

However, the results are really confusing. Most of the refinement stages were not working. In original implementation, loss reduce from stage 1 to stage 6, but in my experiment with dilation, loss stay constant from stage 3 to stage 6.

Did you carry out any experiment using dilation and met the same issue ?

Original:

I0824 06:42:20.602033 8724 solver.cpp:228] Iteration 55075, loss = 685.986
I0824 06:42:20.602083 8724 solver.cpp:244] Train net output #0: loss_stage1_L1 = 112.87 (* 1 = 112.87 loss)
I0824 06:42:20.602092 8724 solver.cpp:244] Train net output #1: loss_stage1_L2 = 33.2826 (* 1 = 33.2826 loss)
I0824 06:42:20.602097 8724 solver.cpp:244] Train net output #2: loss_stage2_L1 = 87.7518 (* 1 = 87.7518 loss)
I0824 06:42:20.602102 8724 solver.cpp:244] Train net output #3: loss_stage2_L2 = 27.0587 (* 1 = 27.0587 loss)
I0824 06:42:20.602108 8724 solver.cpp:244] Train net output #4: loss_stage3_L1 = 83.2692 (* 1 = 83.2692 loss)
I0824 06:42:20.602111 8724 solver.cpp:244] Train net output #5: loss_stage3_L2 = 25.0962 (* 1 = 25.0962 loss)
I0824 06:42:20.602116 8724 solver.cpp:244] Train net output #6: loss_stage4_L1 = 81.7719 (* 1 = 81.7719 loss)
I0824 06:42:20.602121 8724 solver.cpp:244] Train net output #7: loss_stage4_L2 = 24.2401 (* 1 = 24.2401 loss)
I0824 06:42:20.602126 8724 solver.cpp:244] Train net output #8: loss_stage5_L1 = 81.4189 (* 1 = 81.4189 loss)
I0824 06:42:20.602130 8724 solver.cpp:244] Train net output #9: loss_stage5_L2 = 23.8976 (* 1 = 23.8976 loss)
I0824 06:42:20.602136 8724 solver.cpp:244] Train net output #10: loss_stage6_L1 = 81.3313 (* 1 = 81.3313 loss)
I0824 06:42:20.602141 8724 solver.cpp:244] Train net output #11: loss_stage6_L2 = 23.9978 (* 1 = 23.9978 loss)

DenseNet + dilation

I0831 21:17:55.839206 31806 solver.cpp:228] Iteration 80925, loss = 822.671
I0831 21:17:55.839253 31806 solver.cpp:244] Train net output #0: loss_stage1_L1 = 114.009 (* 1 = 114.009 loss)
I0831 21:17:55.839259 31806 solver.cpp:244] Train net output #1: loss_stage1_L2 = 29.0028 (* 1 = 29.0028 loss)
I0831 21:17:55.839264 31806 solver.cpp:244] Train net output #2: loss_stage2_L1 = 107.842 (* 1 = 107.842 loss)
I0831 21:17:55.839309 31806 solver.cpp:244] Train net output #3: loss_stage2_L2 = 27.9593 (* 1 = 27.9593 loss)
I0831 21:17:55.839316 31806 solver.cpp:244] Train net output #4: loss_stage3_L1 = 107.783 (* 1 = 107.783 loss)
I0831 21:17:55.839321 31806 solver.cpp:244] Train net output #5: loss_stage3_L2 = 27.661 (* 1 = 27.661 loss)
I0831 21:17:55.839326 31806 solver.cpp:244] Train net output #6: loss_stage4_L1 = 108.645 (* 1 = 108.645 loss)
I0831 21:17:55.839331 31806 solver.cpp:244] Train net output #7: loss_stage4_L2 = 27.7962 (* 1 = 27.7962 loss)
I0831 21:17:55.839336 31806 solver.cpp:244] Train net output #8: loss_stage5_L1 = 108.079 (* 1 = 108.079 loss)
I0831 21:17:55.839340 31806 solver.cpp:244] Train net output #9: loss_stage5_L2 = 27.6918 (* 1 = 27.6918 loss)
I0831 21:17:55.839345 31806 solver.cpp:244] Train net output #10: loss_stage6_L1 = 108.332 (* 1 = 108.332 loss)
I0831 21:17:55.839350 31806 solver.cpp:244] Train net output #11: loss_stage6_L2 = 27.8711 (* 1 = 27.8711 loss)

[Ai-is-light]

@Xiangyu-CAS would you mind sharing your result of your experiment?

[Xiangyu-CAS]

@Ai-is-light ,
ResNet50 (conv4,5 dilation, bn_istrain) +2tstage : 55% (2644 images)
ResNet50 (conv4,5 dilation, bn_istrain) +6tstage : 55% (2644 images)
ResNet50 (conv4,5 dilation, bn_freeze) + N stage: failed <10%

                  DenseNet121(conv4,5 dilation, bn_freeze) +2tstage : 54%
                  DenseNet121(conv4,5 dilation, bn_freeze) +6tstage : 54%
                  DenseNet121(conv4,5 dilation, bn_istrain) +6tstage : 54%

would you mind sharing yours ?

[Ai-is-light]

@Xiangyu-CAS I haven't try the ResNet and DenseNet , but I will do it. I have tried the original network changed by the dilated work. In my work , I need pay more attention to the speed. However I failed when I try to use the BN-layer at the stage of 2-6. In your work, ResNet50 (conv4,5 dilation, bn_istrain) +2tstage : 55% (2644 images), which means you did use bn-layer in training including every layer of stage2-6 ??? I trained the original network and got coco-2644 about 0.584(mAP), and dilate-conv about 0.543 (mAP). And , now, I try to use the mobileNet to train more faster model. So,

mAP/coco-dataset

mAP/coco-dataset

[hellojialee]

Is the performance you reported obtained on COCO2017 or COCO2014?

[Hanhanhan11]

@Xiangyu-CAS I haven't try the ResNet and DenseNet , but I will do it. I have tried the original network changed by the dilated work. In my work , I need pay more attention to the speed. However I failed when I try to use the BN-layer at the stage of 2-6. In your work, ResNet50 (conv4,5 dilation, bn_istrain) +2tstage : 55% (2644 images), which means you did use bn-layer in training including every layer of stage2-6 ??? I trained the original network and got coco-2644 about 0.584(mAP), and dilate-conv about 0.543 (mAP). And , now, I try to use the mobileNet to train more faster model. So,

mAP/coco-dataset

mAP/coco-dataset

Hi, did you use mobileNet to train the model? I have tried it for a long time, I want to ask if you change the cropsize to 128 to train this model? How about the performance? Thanks