Deep comedy

Report

RNN char level: Embedding - LSTM - Dense - Dense

• Good metric results
• Prone to overfitting
• No rhymes
• No padding

RNN syllabe level: Embedding - LSTM - Dense - Dense

• https://www.groundai.com/project/neural-poetry-learning-to-generate-poems-using-syllables/1
• Very bad results
• not converging loss
• padding on/off

RNN syllabe level: LSTM - Dense - Dense

• Removing Embedding gives a slightly improvement, but not yet comparable to the char level method
• Uneffective metric loss

3 Input Char model with 1 LSTM:

• https://towardsdatascience.com/generating-haiku-with-deep-learning-dbf5d18b4246
• slightly improvement, but not yet comparable to the char level method
• Good performances with syllable count
• No Rhymes

3 Input Syllable model with 1 LSTM:

• https://towardsdatascience.com/generating-haiku-with-deep-learning-dbf5d18b4246
• Significantly worse than the 3 Input Char model
• No Rhymes
• Bad syllable counting

Transformer

• https://arxiv.org/pdf/1706.03762.pdf
• (generation) https://huggingface.co/blog/how-to-generate
• explanation pos encoding https://kazemnejad.com/blog/transformer_architecture_positional_encoding/