Indicated download isn't adequate to run true_poetry

[exo-pla-net]

After running

rsync --verbose rsync://78.46.86.149:873/biggan/2019-12-13-gwern-gpt-2-1.5b-poetry-model-500522.tar.xz ./

And extracting into directory "poetry", true_poetry.py throws an error that there is there is no config.json file.

[summerstay]

Sorry! I've added all the extra files into a subfolder called "poetry." There might be a couple of extra ones in there, only needed for using the tensorflow model.

[exo-pla-net]

Hmm, true_poetry.py now throws the error:

Error no file named ['pytorch_model.bin', 'tf_model.h5', 'model.ckpt.index'] found in directory poetry or from_tf set to False

The pertinent rsync downloaded file is "model-500522.data-00000-of-00001". Seems there needs to be a conversion?

[summerstay]

I just got the information about how to do the conversion.

The file you have there is for tensorflow. T0 convert it for pytorch, try the following:

%env OPENAI_GPT2_CHECKPOINT_PATH=/content/true_poetry/gwern

!transformers-cli convert --model_type gpt2 --tf_checkpoint model-500522 --config /content/true_poetry/gwern/config.json --pytorch_dump_output /

On Tue, May 12, 2020 at 3:46 PM exo-pla-net [email protected] wrote:

Hmm, true_poetry.py now throws the error:

Error no file named ['pytorch_model.bin', 'tf_model.h5', 'model.ckpt.index'] found in directory poetry or from_tf set to False

The pertinent rsync downloaded file is "model-500522.data-00000-of-00001". Seems there needs to be a conversion?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/summerstay/true_poetry/issues/2#issuecomment-627554781, or unsubscribe https://github.com/notifications/unsubscribe-auth/AA3NRIRWFVGBMVSGO5FQ3O3RRGRTDANCNFSM4M6LINQA .

[exo-pla-net]

Thanks for the pointers -- I got it running!

For posterity, here are the steps I took:

Clone repo:

git clone https://github.com/summerstay/true_poetry

Install the requirements:

• Inside repo directory, create requirements.txt with contents:

torch
transformers
tensorflow

• Enter a python virtual environment. Then install: pip install -r requirements.txt

Download the model:

• Navigate to "poetry" directory.
• Download (will take a while): rsync --progress rsync://78.46.86.149:873/biggan/2019-12-13-gwern-gpt-2-1.5b-poetry-model-500522.tar.xz ./

Convert the model:

• In "poetry" directory, run: sudo transformers-cli convert --model_type gpt2 --tf_checkpoint model-500522 --config config.json --pytorch_dump_output /
• It should spit out a "pytorch_model.bin" and a "config.json".
• It might spit them out in a strange place, like the base directory of "Macintosh HD". Move them to the "poetry" folder.

Run true_poetry!

python true_poetry.py

 True poetry we together create, 
 and then we read the poems, we
 discuss the poems, then
 we talk. And we are in our own
 poetic worlds again.
 And I have to confess, it makes
 my blood to see the way
 you hold the world in such esteem.

[summerstay]

Great! Thanks for the detailed writeup, I'm sure it will help others.

[qazbergoef]

Hi,

I used the steps of exo-pla-net to get true_poetry running. At least, to a certain point. The rhymes and model loaded successful. After inserting text at the starting prompt I receive an error:

python true_poetry.py 
rhymes loaded
model loaded
starting prompt: Some text
Traceback (most recent call last):
  File "true_poetry.py", line 526, in <module>
    (probs, past) = expand_node(prompt, None) 
  File "true_poetry.py", line 380, in expand_node
    logits, past = model(**inputs, past=past)
  File "/venv/gpt-2/truePoetryMaster/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
TypeError: forward() got an unexpected keyword argument 'past'

Is it something I did wrong, can someone help me get it working? I did an extensive search online, but cannot get my head around it. I tried to look for a solution in the torch module, but until so far, I do not understand what is going wrong and get it to work. Thanks!

@exo-pla-net, thanks for the clear description!

[summerstay]

It looks like you are using a newer version of huggingface transformers than this was written for. I found this comment:

"It is a known issue. Huggingface transformers 4.0 has a breaking change: past becomes past_key_values in GPT-2 models. You can use transformers 3.* to walk around or wait for our fix that is coming soon."

So try renaming "past" to past_key_values and see if that fixes it.

[summerstay]

I have updated the program to work with the Huggingface 4.0.

[sebhudson89]

Hi, I'm trying to convert my own pre-trained model but I don't have a config.json file in my model folder. How do I make one? Thank you.