This repository encapsulates the process of modeling and forecasting of time series based deep learning,making it super easy to use.
train_file_path = r"dataset/pollution.csv"
#data is univariate tite series
data = pd.read_csv(train_file_path, header=0, index_col=0)["pollution"].values
#get train_data and test_data
train_data, test_data = divide_train_test(data)
#initialise model parameters
ts = ts_model()
# fit and predict
preds, reals = ts.fit_transform(train_data, test_data)
# plot prediction result
ts.plot_predict_result(preds, reals)Some model parameters are diagrammed below：
| The meaning of model parameters |
|---|
| wnd_len : int ,default=24 |
| the length of sliding window.The sequence in sliding window is the single input sequence of LSTM. |
| pred_len : int ,default=24 |
| prediction sequence length |
| net : str ,default=LSTM |
| Net to model time series.Choices include 'LSTM' or 'RNN' or "GRU" |
| rnn_input_size : int ,default=8 |
| embedding dimension,used to map input feature(for univariate,feature is one) to a high dimension |
| rnn_hid_size : int ,default=64. |
| the dimension of hidden layer in RNN,LSTM,GRU |
| batch_size : int, default=32 |
| Batch size to use during SGD optimization. |
| lr : float, default=1E-3 |
| Learning rate used for optimization. |
| n_epochs : int, default=299 |
| Number of epochs to use during optimization. |
| optimizer : str, default='Adam' |
| Optimizer to use during SGD optimization. Choices include 'Adam' or 'SGD'. |
| criterion : str, default='MSELoss' |
| Prediction loss function to use. |
| train_proportion : float,default=0.8 |
| the proportion of train data used to train model.The rest part is used for validation |
| not_use_visdom : bool,default=True |
| not use visdom to visualize the loss in training process,when False ensure run"python -m visdom.server" firstly |
| cuda : bool, default=False |
| Whether or not to use CUDA. |
| dropout_rate : float, default=0.5. |
| Dropout rate for hidden layers. |
| verbose : bool, default=True |
| Print out loss information. |
| fit_transform(train_data,test_data) | train model on train_data and predict on test_data |
|---|---|
| fit(train_data) | train model on train_data |
| transform(test_data) | predict on test_data |
| plot_predict_result(preds, real) | plot prediction result |
| save_best_model(save_path): | save file on specified path |
| get_min_loss() | return minimum validation loss |
If you have any questions, please open an issue.