default training

[summerstay]

I ran train.py with the default parameters. When I test it on sentences, even ones directly from the training set, the resulting sentences it generates seem pretty much random. Should I use a different set of training parameters?

[michalivne]

Can you provide more details? How do you test the model? What happens when you run test script? Micha Livne

…

On Apr 29, 2020, at 06:02, summerstay ***@***.***> wrote:  I ran train.py with the default parameters. When I test it on sentences, even ones directly from the training set, the resulting sentences it generates seem pretty much random. Should I use a different set of training parameters? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[summerstay]

I run train.py. The only parameter I add is the path to the ptb training file, so it is using all the defaults.
To test it, I am using a very short program I wrote that loads the trained model, converts the sentence to indexes (using w2i) and then calls model.encode to create a z vector (by default this seems to be length 16, which seems too small) and then use model.inference to decode that z vector. Here is the core part of the code:

z, mean, std = model.encode(inputsent,inputlen, return_mean=True, return_std=True)

output = model.inference(z=z)

print(idx2word(output[0], i2w=i2w, pad_idx=w2i['<pad>']))

every time I run this, it prints out a different sentence that seems unrelated to the sentence I input.