Can i use it to train a model of ssd512 only? about nncf

[adamantboy]

@aleksei-kashapov

[kshpv]

Dear @adamantboy,

Thank you for your issue.
We found out the bug because of it.
The last pull request should fix it.
It has small fixes so you can apply them manually or if you have more time you may wait till it would be merged.

If it wouldn't work for you please let me know

[adamantboy]

@aleksei-kashapov thanks, it works! However, another question appears, the ap of VOC07 metric is zero always in that the value of batch_detections is all zero, i am sure that the input value fetched from pull_item in COCODataset is right. The output values in terminal are as follows.
batch_detections:
tensor([[[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
...,
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.]]])

all_boxes length: 0
Evaluating detections
VOC07 metric? Yes
classes: ('box',)
cls_ind: 0 cls: box
box_list []
box_list[:, 1]: []
cls_ind + 1: 1
AP for box = 0.0000
Mean AP = 0.0000

It is really weird.

[adamantboy]

I solve this problem when i add 1 to num_classes.

[kshpv]

Awesome!