nan value predicted sometimes for pred_conf

[vcasecnikovs]

nan value predicted sometimes for pred_conf

[VCasecnikovs]

Hello,
thank You for Your question.
Pred_conf comes from the network. So somewhere in the network something explodes and gives you such a result. Could You tell me Your network configuration, otherwise I would not be able to help You?
Best regards,
Vadims.

[azharshaikh06]

the only change I have made is to the backbone, the number of blocks i am using is 1,2,2,1 instead of 2,8,8,4
hparams = {
"pretrained" : False,
"train_ds" : data_path+"train.txt",
"valid_ds" : data_path+"val.txt",
"img_size" : 800,
"bs" : 2,
"momentum": 0.9,
"wd": 0.001,
"lr": 0.01, #0.0001 for 3layer tinyyolo
"epochs" : 2,
"pct_start" : 10/100,
"iou_threshold": 0.5,
"confidence_threshold": 0.5,

"optimizer" : "Ranger",
"flat_epochs" : 15,
"cosine_epochs" : 10,
"scheduler" : "Cosine Delayed", 

"SAT" : True,
"epsilon" : 0.1,
"SAM" : False,
"ECA" : False,
"WS" : False,
"Dropblock" : True,
"iou_aware" : False,
"coord" : False,
"hard_mish" : False,
"asff" : False,
"repulsion_loss" : False,
"acff" : False,
"bcn" : False,
"mbn" : False,

}

[azharshaikh06]

Could you tell me how to convert this model to tiny-yolov4

[VCasecnikovs]

Hello, I would recommend you to cancel each if them separately and find if there are some problems. Maybe you could do some Traceback where it fails. Also, if You have change a backbone problem could be somewhere there... Take a look, maybe you have some problems there.
About tiny yolo. I haven't implemented it. To implement it you should take only half of the channels in route nodes (concatenation).
Tianxiaomo/pytorch-YOLOv4#125

[VCasecnikovs]

Maybe the error was because you had 0 target images. If yes, the last update should fix it. Sorry for bug. d010272 #11