Almost results are Zero. about ssds.pytorch HOT 16 CLOSED

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

from ssds.pytorch.

The same problem met.

from ssds.pytorch.

Any solutions yet for this issue?

from ssds.pytorch.

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.

from ssds.pytorch.

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!

from ssds.pytorch.

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

from ssds.pytorch.

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

from ssds.pytorch.

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.

from ssds.pytorch.

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.

from ssds.pytorch.

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)

from ssds.pytorch.

@whuzs see #15 please

from ssds.pytorch.

@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

from ssds.pytorch.

@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

from ssds.pytorch.

@cvtower i have solved it

from ssds.pytorch.

@cvtower i have solved it

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

from ssds.pytorch.

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.

from ssds.pytorch.