Huggin Face Conversational error: error: argument --model: invalid choice: 'models/' (choose from 'openai-gpt', 'gpt2')

Question

I´m trying to replicate the results of this repo:

https://github.com/huggingface/transfer-learning-conv-ai

For that I'm following the basic example that is not based on docker:

git clone https://github.com/huggingface/transfer-learning-conv-ai
cd transfer-learning-conv-ai
pip install -r requirements.txt
python -m spacy download en

Then I try:

python3 interact.py --model models/

And there I get this error:

  np_resource = np.dtype([("resource", np.ubyte, 1)])
usage: interact.py [-h] [--dataset_path DATASET_PATH]
                   [--dataset_cache DATASET_CACHE] [--model {openai-gpt,gpt2}]
                   [--model_checkpoint MODEL_CHECKPOINT]
                   [--max_history MAX_HISTORY] [--device DEVICE] [--no_sample]
                   [--max_length MAX_LENGTH] [--min_length MIN_LENGTH]
                   [--seed SEED] [--temperature TEMPERATURE] [--top_k TOP_K]
                   [--top_p TOP_P]
interact.py: error: argument --model: invalid choice: 'models/' (choose from 'openai-gpt', 'gpt2')

First thing I notice is that there was not any "models" directory hence I created one and tried again, got the same error.

Second thing I tried was to download the model as in the repo it specifies:

We make a pretrained and fine-tuned model available on our S3 here

From that link I tried:

wget https://s3.amazonaws.com/models.huggingface.co/transfer-learning-chatbot/finetuned_chatbot_gpt.tar.gz

And uncompress the files both in the main directory and in the models directory and tried again.

For the third time, I tried and got the same error.

This is the current structure of my working dir:

Dockerfile   config.json                   interact.py              pytorch_model.bin       train.py
LICENCE      convai_evaluation.py          merges.txt               requirements.txt        utils.py
README.md    example_entry.py              model_training_args.bin  special_tokens.txt      vocab.json
__pycache__  finetuned_chatbot_gpt.tar.gz  models                   test_special_tokens.py

EDIT

Tried kimbo´s sugestion:

python3 interact.py --model gpt2

I get this error now:

 File "interact.py", line 154, in <module>
    run()
  File "interact.py", line 114, in run
    raise ValueError("Interacting with GPT2 requires passing a finetuned model_checkpoint")
ValueError: Interacting with GPT2 requires passing a finetuned model_checkpoint

Also tried just runing:

python3 interact.py

For that I have not get any error, it seems to get stuck at this point:

INFO:/home/lramirez/transfer-learning-conv-ai/utils.py:Download dataset from https://s3.amazonaws.com/datasets.huggingface.co/personachat/personachat_self_original.json
INFO:/home/lramirez/transfer-learning-conv-ai/utils.py:Tokenize and encode the dataset

I have been there for about 30min

Are you still stuck on this or did my answer work? – kimbo Mar 05 '20 at 00:48 — kimbo, Mar 05 '20 at 00:48
@kimbo have not tested it yet, Will accept it once I do it – Luis Ramon Ramirez Rodriguez Mar 05 '20 at 12:54 — Luis Ramon Ramirez Rodriguez, Mar 05 '20 at 12:54

kimbo · Answer 1 · 2020-03-04T03:42:23.043

New Update

It's taking forever to tokenize the dataset because it's tokenizing the entire dataset, which is a 200 MB JSON file.

To make it MUCH faster, just load part of the dataset.

Open up utils.py and change the tokenize function:

def tokenize(obj):
    if isinstance(obj, str):
        return tokenizer.convert_tokens_to_ids(tokenizer.tokenize(obj))
    if isinstance(obj, dict):
        return dict((n, tokenize(o)) for n, o in obj.items())
    limit = 100  # <- this is the number of items in the dataset to load
    return list(tokenize(o) for o in obj[:limit])  # <- change it here

That will only load the first 100 items in the dataset.

Old Answer

When I'm unsure how to use a python script (or anything you run from the command line, really), I usually try a couple things to figure it out.

python script.py -h or python script.py --help. Often that will print out an explanation of the arguments the script is expecting and how to run it.
If it's an executable command you installed, I always try man <executable>. Probably won't work in this case since you just cloned the repo from GitHub and didn't install anything.
If I still don't understand how to use the script because the above didn't work, I go online and look for some documentation (a Github README, a wiki, readthedocs, etc)
If it's documented poorly, I just look at the source code. Sometimes I skip straight to this part because for smaller stuff it's often quicker.

In this case, I read the README on Github and that didn't tell me all that much, so I took a look at interact.py. If you look starting at line 139 (https://github.com/huggingface/transfer-learning-conv-ai/blob/master/interact.py#L139), it appears they're in a while loop, waiting for you to input something to feed to the model.

/end update

This part:

(choose from 'openai-gpt', 'gpt2')

should tell you all you need to know.

Try running

python3 interact.py --model gpt2

or

python3 interact.py --model openai-gpt

I tried your sugestion, got a new error with that one. I updated the question with the details. — Luis Ramon Ramirez Rodriguez, Mar 01 '20 at 12:52
I think the model is waiting for you to input something (try typing something into the terminal and then pressing Enter) — kimbo, Mar 04 '20 at 03:04
@LuisRamonRamirezRodriguez Never mind, tignore my last comment. See updated answer. — kimbo, Mar 04 '20 at 03:43

score 0 · Answer 2 · answered Jun 16 '20 at 22:22

This error is because HuggingFace don't have a fine-tuned model of GPT2 on ConvAI dataset, they only did it for GPT

if args.model == 'gpt2':
            raise ValueError("Interacting with GPT2 requires passing a finetuned model_checkpoint")

You can find this part of the code in the interact.py file, so if you want to have it on gpt2, you need to fine-tune it yourself. I was just struggling with this righ now.

Huggin Face Conversational error: error: argument --model: invalid choice: 'models/' (choose from 'openai-gpt', 'gpt2')

2 Answers2

New Update

Old Answer