FOTS.PyTorch
FOTS.PyTorch copied to clipboard
jiangxiluning
Training Environment
Could you give a little bit more information about the training environment. When I use torch v0.4.1, first I get an error about CTCLoss as it is included after the pytorch 1.0 version. To overcome that, I've installed the warp-ctc from this link, and used that instead. Afterwards, when I start training, I got an error saying TypeError: initializer for ctype 'struct THIntTensor *' must be a pointer to same type, not cdata 'struct THCudaLongTensor *' which I could not find a solution for. While executing build.sh with torch v1.0.1, it says ImportError: torch.utils.ffi is deprecated. Please use cpp extensions instead. First, I googled it and found a solution in this link. I have, however, got another error saying TypeError: dist must be a Distribution instance after that, so I couldn't execute the build.sh this time.
I am facing this issue as well. build.sh does not work on pytorch 1.0 so I used pytorch 0.4.1. However pretrainedmodels uses torchvision 0.4 which requires pytorch 1.0+
@jiangxiluning Could you please specify the version pretrainedmodels to use?
@bhargavaurala I've first installed pytorch 0.4.1 to execute build.sh, then I've installed pytorch 1.0+. There may be something wrong with this approach, but at least it works :)
According to "build.sh" script, it is building roi_align, which is not used by FOTS author and it is not used in this repository too. So i was able to train and evaluate without building the roi_align module using pytorch 1.3.
According to "build.sh" script, it is building roi_align, which is not used by FOTS author and it is not used in this repository too. So i was able to train and evaluate without building the roi_align module using pytorch 1.3.
Did you train successfully? My training process is too slow, can i borrow your model weight to see the effect?thanks a lot
