search

VisionLearningGroup / visda-2018-public

45 stars 33 forks source link

request for initial model #5

Open chrisliu54 opened 6 years ago

chrisliu54 commented 6 years ago

Nice work! But some initial model required in your code is not provided, such as ssd512_vgg16.pth in this line: https://github.com/VisionLearningGroup/visda-2018-public/blob/1ef76001b1c1763295926ffe5ae5765546080052/detection/pytorch-ssd-mmd-coral/examples/ssd/train.py#L30 Hope all required initial models of this repo be exposed soon.

Ellenisawake commented 6 years ago

Similarly please share ImageNet model used as initialization!

No such file or directory: '/scratch2/model_weights/ssd300_12_imagenet.pth'

MInner commented 6 years ago

My bad. I believe it must be this one, and you convert it to proper format using this script. And I believe that the best file to start with would be this one

chrisliu54 commented 6 years ago

Same problem after following your instruction. image

I suppose it could be caused by pytorch version mismatch. I run this script under pytorch=0.4 and I noticed your pytorch version is older(you are using Variable() API). I suggest you provide us a requirements.txt or environment.yaml(for conda) to make thins clear.

MInner commented 6 years ago

That probably means that, for some reason, one of tensors turn contained nan or value outsize of propeor range, I guess? If you could share somehow what are vales of these tensors, which of them are suspicious and suggest how did they manage to get these values, I might try to guess why it might have happened. This code worked fine for me, I usually do not push broken code intentionally :)

chrisliu54 commented 6 years ago

When I turned into torchcv, most of the problems disappeared. But the so called device-side assert triggered is caused by misconfiguration(the number of class of a model should be #foreground_classes + 1, which included the background as one class).