- Notifications
You must be signed in to change notification settings - Fork1
GPT (Decoder only Transformer - from scratch) generated fake/phoney taxonomies (based on NCBI taxonomy dataset)
License
suvash/taxophoney
Folders and files
| Name | Name | Last commit message | Last commit date | |
|---|---|---|---|---|
Repository files navigation
GPT (Decoder only Transformer - from scratch) generated fake/phoney taxonomies, trained on NCBI taxonomy dataset, included in this repository.
- Pytorch - 1.12.1+cu116 (with CUDA support - for reasonably short training runs)
$ python gpt.pyUsing device: cudastep 0: train loss 4.4625, val loss 4.4653step 500: train loss 2.0843, val loss 2.1280step 1000: train loss 1.5394, val loss 1.5920step 1500: train loss 1.3097, val loss 1.3789step 2000: train loss 1.1842, val loss 1.2741step 2500: train loss 1.1017, val loss 1.2182step 3000: train loss 1.0408, val loss 1.1938step 3500: train loss 0.9831, val loss 1.1692step 4000: train loss 0.9382, val loss 1.1591step 4500: train loss 0.8935, val loss 1.1392step 4999: train loss 0.8545, val loss 1.1383The model training and sampling script can be used to train the model and generate(sample) a lot of names afterwards. Some of the names have been included in thetaxophoney.txt file included in the repo.
Naturally, some of these names makes one wonder what they could look like. I've used theStable Diffusion v1-5 Model by RunwayML to generate the images for some of the names. The generation prompt only includes the common name (inside the parens) and not the scientific names, since they didn't help with plausible images.
About
GPT (Decoder only Transformer - from scratch) generated fake/phoney taxonomies (based on NCBI taxonomy dataset)
Topics
Resources
License
Uh oh!
There was an error while loading.Please reload this page.