happywu / sequence-level-semantics-aggregation Goto Github PK

您好，您提供的预训练模型的链接点不进去，能更新一下吗，谢谢

How can i make ImageSets for my datasets?
Thank you!

Hey I ran the follwing command using Python 3.5.2 Ubuntu 16.04

python experiments/selsa/test.py --cfg experiments/selsa/cfgs/resnet_v1_101_rcnn_selsa_aug.yaml --test-pretrained ./model/pretrained_model/selsa_rcnn_vid

And I got the following error

Traceback (most recent call last):
File "/home/indoordesk/project/Sequence-Level-Semantics-Aggregation/experiments/selsa/test.py", line 16, in
import test
File "/home/indoordesk/project/Sequence-Level-Semantics-Aggregation/experiments/selsa/../../rcnn_selsa/test.py", line 48, in
from function.test_rcnn import test_rcnn
File "/home/indoordesk/project/Sequence-Level-Semantics-Aggregation/experiments/selsa/../../rcnn_selsa/function/test_rcnn.py", line 19, in
from symbols import *
File "/home/indoordesk/project/Sequence-Level-Semantics-Aggregation/experiments/selsa/../../rcnn_selsa/symbols/init.py", line 1, in
import resnet_v1_101_rcnn_selsa
ImportError: No module named 'resnet_v1_101_rcnn_selsa'

In your code of augmentation:

if overlap.min() < min_iou and max_iou < overlap.max():

In this line, the max_iou in INF. So max_iou will never be < overlap.max(). As a result, whatever the overlap is, the crop will be kept.

Is this a bug? Our you intend to keep all random crops?

Hi, I have troubles with the order of the 300 pooled roi features (i.e., symbol 'roi_pool' in "resnet_v1_101_rcnn_selsa.py"). When I try to change the order of these roi features at the test phase, the mAP drops significantly (If I exchange the first 100 features (0-99 in 'roi_pool ') and the final 100 features (200-299 in 'roi_pool'), the mAP will drop from 82.7 to 3.8......). I only chage the order of the 300 features and their values are untouched. The order of the pooled roi features should not to be the reason for the very bad result. Any suggestions? Thanks

When I try to use batch size greater than 1 for training, error occurs. If I want to change batch size bigger than 1, could you offer some advice how to modify code?Thx~

Hi,
The pre_weights link seems failure...

Could you please share the links for the datasets? I have been able to find the link to the VID dataset, but not for the DET dataset.

Can you release them again?
Thanks

Hey
Any suggestion how to do transfer learning to humans detention?
Also so goods dataset for it?
Thanks!

i am confused by the meaning of the txt file inside /ImageSets/, and i really want to know if I have some misunderstandings.

Hi! Is there any config file like this one to not perform data augmentation at the inference time? What should I modify to achieve it?

Thanks in advance!

From where can i reduce the batch size for testing and training purpose?

Hi! Can you provide the weights to test SELSA with ResNeXt-101? I'm interested in replicating those results from the paper.

Thanks in advance!