Jasminexjf / LSTM-Classification-Pytorch

Text classification based on LSTM on R8 dataset for pytorch implementation

---

Text Classification based on LSTM on R8 Dataset-[pytorch implementation]

1. Details of file fold:

• data/
• data/train_txt/*.txt
• data/train_txt.txt
• data/train_label.txt
• data/test_txt/*.txt
• data/test_txt.txt
• data/test_label.txt

2. File description:

file	description
data/train_txt/	training text fold
data/test_txt/	testing text fold
data/train_txt.txt	file name list for training text
data/test_txt.txt	file name list for testing text
data/train_label.txt	label list for training text
data/test_label.txt	label list for testing text

3. Running example:

Environment: python 3

Requirements:

pytorch

running example:

python main.py

output:

4. Dataset:

We use the following dataset for our example: R8 Dataset.link.

5. Notes:

• This code is largely inspired from the original repo: jiangqy with a few modifications:
  • Use the Adam optimizer
  • Custom dataloader
  • Select the right output after the LSTM/GRU (masking)
  • Attention layer included, from here. The attention mechanism is not the full one though, it is a simplified version. No new parameters are learned in the process as it just combines the rnn output with the hidden states using matrix multiplications. This simplified attention mechanism still provides a significant improvements over no attention at all.
• The entire sentences are used. If you want to cap the number of words, please, update the parameter sentence_len = None to sentence_len = n. You should get pretty descent results even with only 32 words max per sentence.
• There is no early stopping. The code just runs for n epochs and saves the results at the end. But it is easy enough to select the best model based on the evaluation phase.