Provided checkpoint file cannot be used. about neural-vqa HOT 6 CLOSED

To answer my own question:

What are the appropriate paths for answers_vocab.t7, questions_vocab.t7 and data.t7?

They should be in the data folder.

from neural-vqa.

Training from scratch, i.e. generating answers_vocab.t7, questions_vocab.t7 and data.t7 works. But you need to install torch with

TORCH_LUA_VERSION=LUA51 ./install.sh

Version LUA52 cannot be installed, and LUA21 (the default) cannot allocate enough memory.

from neural-vqa.

@ilija139 I see, thanks! And were you able to fix the nan training loss issue?

from neural-vqa.

To answer your first query, the checkpoint file is old. I added a bunch of things after that (Multi-layer LSTMs, thinner model checkpointing, etc), which is probably why it broke. :/ I'll fix it as soon as I can.

from neural-vqa.

@abhshkdz yes, the nan training loss is solved when training from scratch.
These are the files I've just generated with the current version of the repo. If you want you can link them:

https://www.dropbox.com/s/ounll9zagxk8hp5/answers_vocab.t7?dl=0
https://www.dropbox.com/s/55hk5x2x2fh63sc/data.t7?dl=0
https://www.dropbox.com/s/ty6wm905oexk4rb/questions_vocab.t7?dl=0
https://www.dropbox.com/s/3vj6477ia5xmz2f/vqa_epoch15.00_0.4576.t7?dl=0

Anyway thanks for providing the code, it's easy to understand so it's useful as a starting point.

from neural-vqa.

@ilija139 I'll link these, thanks!

from neural-vqa.