Fine-tuned and presented by Hannah Bohle, November 2020
Warning: The following GPT-2 model was pretrained on an unknown German corpus by an anonymous user at huggingface. Therefore, we cannot rule out embedding biases and enabling unintended negative uses induced by the corpus. In addition, the German novel used for fine-tuning contains explicit language. When using the model, please be aware that all content used for pretraining and fine-tuning will affect the generated text.
For installation see: huggingface
Most models from transformers can either be based on Tensorflow, Pytorch or both. The provided GPT-2 is (currently only) based on Pytorch (November 2020).
- Installation inside venv: 1. TF and/or 2. Pytorch and 3. transformers
- install dependencies
- load the dataset
- prepare the dataset and build a TextDataset
- load the pre-trained German-GPT-2 model and tokenizer 1.initialize Trainer with Training Arguments
- train and save the model
- test the model
Model pre-trained on German text and fine-tuned on novel "Spieltrieb" by Juli Zeh. Data contains only parts of the novel with randomized chapters to prevent copyright violations.
Txt-file consists of 79886 words. Txt file can be found in 'data' folder.
For text preprocessing it is important to use the same tokenizer as was used for pretraining the model. In this case we use the "german-gpt2" tokenizer as provided by an anonymous user at huggingface.
tokenizer = AutoTokenizer.from_pretrained("anonymous-german-nlp/german-gpt2")Caution: We use a GPT-2 pretrained on an unknown German corpus. Therefore, we cannot rule out embedding biases and enabling unintended negative uses.
For the English version, huggingface states that GPT2 "was trained on millions of webpages with a causal language modeling objective" However, it is unclear which sources were used to pretrain the German model and how large the corpus was.
model = AutoModelWithLMHead.from_pretrained("anonymous-german-nlp/german-gpt2")Model is trained and saved and able to generate lyrics. Unfortunately the pretrained model cannot be saved in github because of its size (~500MB).
Neural nets calculate probabilities. In its basic versions, such as using greedy search, the neural net simply selects the word with the highest probability as its next word. As a result, AI-generated texts often suffer from repetitions.
There are different possibilities to make AI-generated text sound more human like. For instance, by penalizing n-gram repetitions or by add some randomness to text generation, e.g. varying the temperature (for different temperature options, see also my LSTM RNN).
In GPT-2 some advanced sampling methods were introduced which were used for generating the lyrics which are presented in the beginning of the Readme, namely top_k and top_p.
outputs = german_model.generate(inputs, max_length = max_length, do_sample=True, top_p=0.95, top_k=50,num_return_sequences=2)top_k => In Top-K sampling, the K most likely next words are filtered and the probability mass is redistributed among only those K next words.
top_p => Having set p=0.95, Top-p sampling picks the minimum number of words to exceed together p=.95% of the probability mass
num_return sequences makes it possible to generate several samples from the provided prompt.
When prompted "Ada liebte ihre Katze" the GPT-2 generated the following lyrics. It might sound a bit biblical and old-fashioned but pretty human-like in my opinion.
"Ada liebte ihre Katze und den Esel, und nun wird sie von ihrer Mutter und vom Haus gewiesen, daß sie ihr Vater wird, und sie wird den Namen ihres Vaters Jela nennen. Und sein Vater war ein Richter in seiner Stadt. Und Jela sagte: Zu wem soll ich hinaufgehen, du schöner Mann? Er antwortete: Mein Herr, wir haben keinen Umgang miteinander. Mein Herr, so sollst du hinaufgehen und zu ihm hinabgehen und sagen: Ich bin Jela, dein Bruder; wir wollen hinfahren, daß wir hingehen. Und er sah Jela und seinen Vater an und sagte: Ich werde hingehen, Jela zu suchen; und ich werde zu ihr hinabgehen und sagen: Ich"
According to huggingface, "Text generation is currently possible with GPT-2, OpenAi-GPT, CTRL, XLNet, Transfo-XL and Reformer in PyTorch and for most models in Tensorflow as well. [...] GPT-2 is usually a good choice for open-ended text generation because it was trained on millions of webpages with a causal language modeling objective." It would be really interesting to compare the quality of generated text by German pretrained text generation models.
- The pretrained model was provided by an anonymous user at huggingface
- Limitations and biases in text generation
- For text generation sampling techniques see the interesting huggingface blog post by Patrick von Platen
- There are more textgeneration models from huggingface, some of them pretrained on different european languages. (What's the name of the French version of the Bi-directional Encoders for Transformers (BERT)? CamemBERT!)