Couldn't find any detections about dsod

[lzx1413]

DSOD/DSOD300_pascal.py

Line 334 in 2d33da5

'test_initialization': True,

turn this False and everything will be OK.

[szq0214]

Hi @miraclewkf, see BVLC/caffe#4594. I think it's an SSD bug and has already been fixed. I recommend you to use the latest version of SSD (https://github.com/weiliu89/caffe/tree/ssd).

[shudct]

Hi, @miraclewkf , I meet the same questions as you. And I have already used the latest version of SSD. Have you solved the problem? thank you

[Tmono]

I meet the same questions...And I change the batch size to a small one, it didn't work either. ：(

[moyans]

My detection_eval only 0.045 in 3.1w Iteration， But the loss is normal

[szq0214]

@moyans, Please use our default parameters setting. If you only have one GPU device, just change "batch_size" to 4 with other parameters constant. I guess this is because your "accum_batch_size" is set to a very small value.

[moyans]

Thank to @szq0214 , I change the "accum_batch_size" like to "batch_size" , I set to 128 now，But it is really slow speed。

[edwwang]

@miraclewkf @shudct Hi, have you solved this problem? I have already used the lattest ssd version and tried to modify the batch_size to 4 and 16 , but the question remains the same . Btw, my machine has got 4 GPUs , and I have changed the contents in examples/dsod/DSOD300_pascal.py from ' gpus = "0,1,2,3,4,5,6,7" ' to ' gpus = "0,1,2,3" '.

[shudct]

@edwwang The information 'Couldn't find any detections' is just a warning which means in this image there is no object detected. DSOD trains from scratch so it's very possible to detect no objects for the first test. And I ignore the warning, there is no that for the next test.

[edwwang]

@shudct Ok i know. Thanks for your answer.

[BestSongEver]

@moyans
Hi,I'm facing the same problem with u, achiving a detection_eval = 0.0555858. Have u solve the problem? And how?
Please help
Thx a lot : ).

[moyans]

@BestSongEver when I set ‘accum_batch_size’ to 128（before I set it was equal to "batch_size: 4" ），the detection_eval become normal.