Has anyone met this error when I use learn = ConvLearner(data, models.tvm.inception_v3, metrics=accuracy,callback_fns=ShowGraph). It looks ony work on resnet series. When I use inception or others. It error like:
RuntimeError: Error(s) in loading state_dict for Sequential:
While copying the parameter named “0.0.conv.weight”, whose dimensions in the model are torch.Size([32, 3, 3, 3]) and whose dimensions in the checkpoint are torch.Size([32, 3, 3, 3]).
RuntimeError: size mismatch, m1: [8 x 277248], m2: [768 x 1000] at /opt/conda/conda-bld/pytorch-nightly_1539421063549/work/aten/src/THC/generic/THCTensorMathBlas.cu:266
ConvLearner hasn’t been tested apart from resnets for now, so it doesn’t support other models for now. This will be added in future developments.
Hi, is there any sense of timing for when other models such as densenet*, resnext* and inception might be added to v1?
I started to get something working by looking at 0.7 but it quickly became a frankenstein and t’was easier to switch back to v0.7 for now. The docs still say “On top of the models offered by torchivision, the fastai library has …” so perhaps there is a less direct way to use these models, if so it would be great to have a pointer as to how to go about making it work with the current codebase.
You have to add their metadata, like for the resnet, which tells the library where to cut for the head and where to split for differential learning rates.
We won’t be working on this any time soon, so it probably won’t be available until someone makes a PR for that.
OK thanks for the reply, much appreciated.