Hey all,
I have my network topology saved in buffers, as follows:
register_buffer(“in_channels”, torch::tensor(in_channels, torch::dtype(torch::kInt64)));
register_buffer(“classes”, torch::tensor(classes, torch::dtype(torch::kInt64)));
register_buffer(“depth”, torch::tensor(depth, torch::dtype(torch::kInt64)));
register_buffer(“channel_base”, torch::tensor(channel_base, torch::dtype(torch::kInt64)));
register_buffer(“bottleneck_depth”, torch::tensor(bottleneck_depth, torch::dtype(torch::kInt64)));
After model initialisation and training, i save the model as follows:
torch::save(trained_model, model_file.generic_string());
To my supprise, the registered buffer parameters are not saved at all when i load a new model with
torch::load(inference_model, model_file.generic_string());.
How can i solve this problem? My network topology changes all the time in function of the type of images presented. it is crucial that the network topology is saved in the *.pt file itself.
Thanks in advance!
I.