💻 Deep Learning and Practice Final Project
🏹 Stock Prediction using Transformer
Download data from dataset and unzip it into directory "archive".
$ kaggle datasets download -d jacksoncrow/stock-market-datasetReference is implemented in Tensorflow.
Stock predictions with state-of-the-art Transformer and Time Embeddings
| Argument | Description | Option | Default |
|---|---|---|---|
'-e', '--epochs' |
Number of epochs | int | 10 |
'-w', '--warmup' |
Number of epochs for warmup | int | 2 |
'-l', '--learning_rate' |
Learning rate | float | 0.001 |
'-b', '--batch_size' |
Batch size | int | 64 |
'-s', '--seq_len' |
Sequence length (consecutive days) | int | 128 |
'-ne', '--num_encoder' |
Number of transformer encoder in the network | int | 3 |
'-a', '--attn_dim' |
Dimension of single attention output | int | 96 |
'-nh', '--num_heads' |
Number of heads for multi-attention | int | 12 |
'-d', '--dropout_rate' |
Dropout rate | float | 0.3 |
'-hs', '--hidden_size' |
Hidden size between the linear layers in the encoder | int | 256 |
'-loss', '--loss_function' |
Loss function | 'l1' or 'l2' | 'l2' |
'-i', '--inference_only' |
Inference only or not | 'store_true' | False |
'-r', '--root_dir' |
Directory containing the downloaded data | str | 'archive' |
'-v', '--verbosity' |
Verbosity level | 0-2 | 0 |