This page presents selected functions provided by tools.py which is commonly used in the project.
Some times models are trained by the another (or old version) repo. When importing the pretrained model to new repo, the name of certain variables /paramters might be changed.
The following commands can be used to browse items in the pretrained model and renaming specific parameters (Remove the brackets when testing, brackets only indicates the enclosed one can be replaced with other string).
- looking up item in the model file.
# cd weights/pytorch-resnet50/
# download pretrained model by
# wget https://download.pytorch.org/models/resnet50-19c8e357.pth
python tools.py --keyword verbose --verbose_list all --old [weights/pytorch-resnet50/resnet50-19c8e357.pth]
# save to name into text
python tools.py --keyword verbose --verbose_list all --old [weights/pytorch-resnet50/resnet50-19c8e357.pth] | awk '{ print $1}' | tee name_list.txt
- renaming parameter
Export pytorch official resnet model to Detectron2 format as initialization model. Edit your own mapping_from.txt and mapping_from.txt file based on the naming space (which can be browsed by above command)
python tools.py --keyword update[,raw] --mf [weights/det-resnet50/mapping_from.txt] --mt [weights/det-resnet50/mapping_to.txt] --old [weights/pytorch-resnet50/resnet50-19c8e357.pth] --new [weights/det-resnet50/official-r50.pth]
Add raw in the keyword to generate the model file with/without state_dict key.
Another example for LSQ 2bit model converting:
python ../model-quantization/tools.py --keyword update,raw --mf weights/det-resnet50/mf_lsq_2bit.txt --mt weights/det-resnet50/mt_lsq_2bit.txt --old weights/pytorch-resnet50/lsq_best_model_a2w2-new.pth --new weights/det-resnet50/lsq_best_model_a2w2.pth