Saving and loading VarStores of Sequential models about tch-rs

[LaurentMazare]

Actually I think this is the expected behavior (but the naming may be a bit confusing). vs.load(f) loads the weights for all the tensors currently in vs from file f. In your case vs is empty so nothing gets loaded and vs is still empty afterwards.
If you modify your code as follows to create the net (and so create the variables) before calling vs.load, their weights will actually be loaded from the file. Printing the var-store is not enough to confirm this as even without the vs.load call it would return the proper variable names and shapes hence I made italso print the weight values currently stored in one of the variables. You will notice that these values are identical between cargo run and the later call to cargo run weights.pt.
Let me know if this doesn't make sense, also maybe vs.load should be called vs.load_weight_values or something like that to be more explicit about what it does.

fn main() -> failure::Fallible<()> {
    let args: Vec<String> = std::env::args().collect();
    let mut vs = nn::VarStore::new(Device::Cpu);
    if args.len() < 2 {
      train(&mut vs)?;
      vs.save("weights.pt")?;
    } else {
      let _ = net(&vs.root());
      vs.load(args[1].as_str())?;
    }

    println!("{:#?}", vs.root());
    println!("{:#?}", Vec::<f64>::from(&vs.root().get("bias").unwrap()));
    Ok(())
}

[LaurentMazare]

Closing now, the doc has been updated to be clearer on this. Feel free to reopen if you encouter further issues.

[grtlr]

Thanks for the quick response and the great clarification. Thinking about it now, it makes a lot of sense to do it that way. I think either vs.load or vs.load_weight_values is fine, maybe we could add a brief explanation in the documentation. With this in mind, vs.load_weight_values might encourage more users to actually read the documentation 😇.

[kenkoooo]

I had the same problem and solved it by this thread! Thanks a lot!