在github本仓库页面的Settings -> Secrets -> Actions -> New repository secret中,添加GOOGLE_SCHOLAR_ID变量:name=GOOGLE_SCHOLAR_ID、value=SCHOLAR_ID。
在github本仓库页面的Action中,点击*"I understand my workflows, go ahead and enable them"*启用workflows by clicking *"。本action将会谷歌学术引用的统计量数据gs_data.json到本仓库的google-scholar-stats分支中。每次修改main分支的内容会触发该action。本action也会在每天08:00 UTC定时触发。
@Andrew-Qibin I am trying to test the results based on the given pre-trained model. But it seems that tthe .pth provided is saved from .state_dict(), which cannot be used to load the structure of the network. Could you please show how can I evaluate the network? Thank you.
Hello, thanks for your excellent job. I use it to train voc datasets. After 40000 iterations, the results are as follows. It seems the miou only 7%? Can you help me, where I am wrong?
2021-09-15 03:23:16,957 SSD.trainer INFO: Total training time: 2 days, 9:11:49 (5.1477 s / it)
2021-09-15 03:23:16,981 SSD INFO: Start evaluating...
2021-09-15 03:23:17,036 SSD.inference INFO: Evaluating voc_2007_test dataset(4952 images):
2021-09-15 03:24:05,197 SSD.inference WARNING: Number of images that were gathered from multiple processes is not a contiguous set. Some images might be missing from the evaluation
2021-09-15 03:24:37,639 SSD.inference INFO: mAP: 0.0723
aeroplane : 0.0652
bicycle : 0.0396
bird : 0.1066
boat : 0.1150
bottle : 0.0238
bus : 0.0395
car : 0.0527
cat : 0.1141
chair : 0.1031
cow : 0.0233
diningtable : 0.0628
dog : 0.0737
horse : 0.0834
motorbike : 0.0736
person : 0.0704
pottedplant : 0.0180
sheep : 0.1083
sofa : 0.0568
train : 0.1090
tvmonitor : 0.1067
I try to use pretrained weigth of mobilenet_v2 pascal voc and test using the given config file. However, it seems like the model and weight are mismatch. I have tried to manually adjust the channel numbers of extra layers to match the weight but accuracy is not correct.
ed.
Thank you for your wonderful research. Regarding some of your code, I have a few confusions. When I selected mobilenet_v2ussd320_voc0712.yaml as the configuration file, I calculated the model parameter count to be 3.43M, while in Table 8 of your paper, the model parameter count is 4.3M. Is there anything else that needs to be changed in the configuration file or elsewhere?? Looking forward to your reply very much, thank you.....