search

luca-medeiros / boxinst_lightning

Adelaidet's Boxinst model implemented in Pytorch Lightning
2 stars 2 forks source link

why read GT mask #1

Open StephenZhao1 opened 3 days ago

StephenZhao1 commented 3 days ago

Great job !!! I have read your datasets class, this part means you use the gt mask. But as far as I know, boxinst just use bbox labels for predicting instance mask. Waiting for your response.Thx image

luca-medeiros commented 2 days ago

Not used for training. You can just remove it if you like. https://github.com/luca-medeiros/boxinst_lightning/blob/17aaff534ef3dc0a862d4bf69424e13c0b1feaa2/data/dataset.py#L202-L226

when collating, masks are not included.

StephenZhao1 commented 2 days ago

After reading your source code carefully, I have found your collate function usage. Thx for your quick response. Great job again !!!

StephenZhao1 commented 2 days ago

without modifying network, below error happens. image the code I use pretrained model file is like: image Where should I change to fix this bug? Thx a lot.

StephenZhao1 commented 2 days ago

I have found this bug. The channel of backbone is [512, 1024, 2048] in config file. image But in your code, the channel is [48, 136, 384] image

StephenZhao1 commented 17 hours ago

Any update? I think the problem was caused by timm verison?

luca-medeiros commented 9 hours ago

Hey Stephen, this repo is not maintained. Feel free to fork it and make a PR if you wish.