- Split the training file into 5 folds
source/train_folds.py - Tokenize and padd the data
- Embedding layer( embedding dim = 128, max_len=300, batch_size=4096,learning_rate=0.001).
- Two Bidirectional GRU layers with 300 units each
- One Conv1D layers with 300 filters
- One GlobalAveragePooling1D
- One GlobalMaxPooling1D
- Concatenate the max_pool and avg_pool
- Dense layer with 1024 units
- Batchnormalization 11 Dropout 12 Dense layer with 20 units
- Check
notebooks/enzyme.ipynb - Pseudo_labeling: Small improvement from 88.99 to 89.15
- After making predictions, add the predicted test samples to the original training data, then re-train again
- Submit prediction
- Put your data into input folder, and your models into models folder
- The model was trained using Google Colab TPU
- App demo using GradioML:
