Closed wwjwy closed 1 year ago
The config in MMRazor only define the distillation configuration. If you want modify settings of MMSeg, you can modify config in MMSeg, then import them to MMRazor config by mmseg::_base_/datasets/cityscapes_modify.py
in the front.
As for how to change MMSeg config, you can refer mmseg doc
The config in MMRazor only define the distillation configuration. If you want modify settings of MMSeg, you can modify config in MMSeg, then import them to MMRazor config by
mmseg::_base_/datasets/cityscapes_modify.py
in the front. As for how to change MMSeg config, you can refer mmseg doc
Thank you for your help! it runs well now.
When i distill transformer seg models --->CNN seg model, there is something wrong. assert preds_S.shape[-2:] == preds_T.shape[-2:] The transformer and CNN has different output shape, can you show me how to modify the config file?
assert preds_S.shape[-2:] == preds_T.shape[-2:]
used to make sure feature shape between teacher and student are the same, to make sure the KD loss can be calculate. For example, MSE(Tensor shape [2, 3, 4, 4], Tensor shape [2, 3, 5, 5])
can not get a result. As far as I know, many Transformer based backbone like PVT, Swin, ..., their backbone output is also with shape [B, C, H, W], it can used to distill with CNN based backbone.
So, my advice is
1) check your preds_S.shape
and preds_T.shape
, make sure preds_S.shape[-2:] == preds_T.shape[-2:]
, if not, check your config, make sure your recorders
get right feature.
2) if shape between teacher and student are different, but you still want to distill them, you can align dimension by using connector
in MMRazor, example config
Thank you for your help!
After training the cwd model, i run the tools/test.py, but i can only get evaluation metrics. How can i get the segmentation images predicted by the distilled student model?
MMRazor tools/test.py
seems not provide --show
, but MMSeg provide, like this. You can try use MMSeg tools/test.py
with --show
, or if your server has no GUI, you can save the image to disk by --show-dir
.
Thanks
I have noticed your question about how to transfer distill model
to student only model
, however, there are no such tools in MMRazor now. I may develop some tools to meet this requirement later.
Thank you very much for your help!
Hi, I add tools to convert distill model
to student only model
, you can refer #381.
Based on my own dataset, I trained a model using mmseg1.x. Now I want to realize knowledge distillation, resnet101+deeplab--->resnet18+deeplab. I configured the mmrazor environment, but in the distill config file https://github.com/open-mmlab/mmrazor/blob/bbb58f1a5c2fe2878484856767dba540092bc7bf/configs/distill /mmseg/cwd/cwd_logits_pspnet_r101-d8_pspnet_r18-d8_4xb2-80k_cityscapes-512x1024.py#L2 I don't know where I can modify the default settings of mmseg, such as the number of categories to be segmented, data set selection, etc.? Can you give me some help?