Almost results are Zero.

[kimna4]

Hello, I tried run your code using "ssd_vgg16_train_voc.yml"

But, almost results are zero and doesn't increase.

=======================================================
AP for aeroplane = 0.0000
AP for bicycle = 0.0000
AP for bird = 0.0000
AP for boat = 0.0000
AP for bottle = 0.0000
AP for bus = 0.0000
AP for car = 0.0000
AP for cat = 0.0000
AP for chair = 0.0000
AP for cow = 0.0000
AP for diningtable = 0.0000
AP for dog = 0.0000
AP for horse = 0.0000
AP for motorbike = 0.0000
AP for person = 0.0001
AP for pottedplant = 0.0003
AP for sheep = 0.0000
AP for sofa = 0.0000
AP for train = 0.0000
AP for tvmonitor = 0.0000
Mean AP = 0.0000

Results:
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000

================================================

Above result are printed at epoch 5.
Although the number epoch is too low, there are problems.
Please tell me the way to fix these problems if you know.

Thank you.

[shufanwu]

I met the same problem and i have trained the model nearly 700 epoch for 300 images with batch size 32

[cvtower]

The same problem met.

[anjandeepsahni]

Any solutions yet for this issue?

[kimna4]

I use another SSD code (https://github.com/qijiezhao/pytorch-ssd) instead of this code.
With this code, I can meet the performance which written at SSD paper.

[cvtower]

I use another SSD code (https://github.com/qijiezhao/pytorch-ssd) instead of this code.
With this code, I can meet the performance which written at SSD paper.

Hi @kimna4 ,
Is the imagenet pretrained model needed to reproduce the performance,
Or training from scratch could also repfoduce?

Thanks very much for your help!

[kimna4]

Hi @cvtower ,

You can get the pretrained model from here ( https://github.com/amdegroot/ssd.pytorch ).
I think that the repository is a master code. So you can get a lot of information here.

Thank you

[1453042287]

i don't know why but i use the same model to finetune the BDD100K dataset, and i get a quite well result:
AP for bike = 0.2777
AP for bus = 0.4525
AP for car = 0.4817
AP for motor = 0.2510
AP for person = 0.2840
AP for rider = 0.2626
AP for traffic light = 0.1715
AP for traffic sign = 0.2024
AP for train = 0.0001
AP for truck = 0.4366
Mean AP = 0.2820

[cvtower]

i don't know why but i use the same model to finetune the BDD100K dataset, and i get a quite well result:
AP for bike = 0.2777
AP for bus = 0.4525
AP for car = 0.4817
AP for motor = 0.2510
AP for person = 0.2840
AP for rider = 0.2626
AP for traffic light = 0.1715
AP for traffic sign = 0.2024
AP for train = 0.0001
AP for truck = 0.4366
Mean AP = 0.2820

Hi,
This problem will be met when training from scratch, and the pre-trained model could almost reproduce the result.

[cvtower]

Hi all,

After checking the source code and cfg files, i found that the default .yml cfg file for most network contains only 'test' phase, that is no training will ever happen during default "training".
I could train from scratch normally now.

To solve this issue:

1. modify the .yml file-add train into the phase list , and prepare corresponding datasets
2. if you use pytorch 0.4.0 and met other errors, previous issues will provide solution.

[whuzs]

why i run the demo, it's none result?

ids, count = nms(boxes, scores, self.nms_thresh, self.top_k)
ValueError: not enough values to unpack (expected 2, got 0)

[blueardour]

@whuzs see #15 please

[blueardour]

@whuzs also try: add 'scores.size(0) == 0' in detection.py as follows:
scores = conf_scores[cl][c_mask]
if scores.size(0) == 0 or scores.dim() == 0:
continue

[qiaoyaya2011]

@cvtower hi, after i modify the .yml file-add train into the phase list ,i also meet the same problem.
AP for human0 = 0.0000
AP for head = 0.0000
AP for cloth = 0.0000
AP for fire = 0.0000
Mean AP = 0.0000

Results:
0.000
0.000
0.000
0.000
0.000

but torch version is 1.3.0

[qiaoyaya2011]

@cvtower i have solved it

[goodlixu]

@cvtower i have solved it

hello,
I met the same problem, can you tell me how did you solve this problem?

[foreverYoungGitHub]

I think the initialization weights in the master branch has some issues. That cause the problem for low AP when we train from scratch. But it should be fixed by the dev branch. Please try the code in the dev branch. Will close the issues for now.