Skip to content

RNN

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

RNN

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
RNN.ipynb
RNN.ipynb
 
 

README.md

Time Series Prediction with RNN (PyTorch)

This project implements a Recurrent Neural Network (RNN) using PyTorch to perform time series forecasting on a synthetic sinusoidal dataset.

The goal is to build an end-to-end sequence learning pipeline including:

  • Synthetic data generation
  • Custom sequence dataset creation
  • RNN model definition
  • Training and evaluation
  • Visualization of predictions

Dataset

Synthetic Sinusoidal Data

The dataset is generated using the sine function:

[ y = \sin(x) ]

  • Number of samples: 1000
  • Sequence length: 50
  • Input: A sliding window of 50 time steps
  • Target: The next value in the sequence

Each training sample looks like:

Input : [y(t), y(t+1), ..., y(t+49)] Target : y(t+50)

This transforms a continuous signal into a supervised learning problem.

Model Architecture

The model is a simple vanilla RNN:

  • RNN layer:
    • Input size: 1
    • Hidden size: 16
    • Number of layers: 1
  • Fully connected layer:
    • Maps hidden state → output value

Forward Pass Logic

  • The RNN processes the full sequence.
  • Only the last time step output is used.
  • A linear layer maps it to the final prediction.

Training Details

  • Framework: PyTorch
  • Loss Function: Mean Squared Error (MSE)
  • Optimizer: Adam
  • Learning Rate: 0.001
  • Batch Size: 32
  • Epochs: 20

Results

Training loss decreases rapidly, showing that the RNN successfully learns the sinusoidal pattern.

Epoch Final Loss
1 0.1947
5 0.0097
10 0.0012
20 0.0003

The model generalizes well to unseen ranges of the sine function.

Testing & Visualization

The trained model is tested on two unseen intervals:

  • Test 1: sin(100 → 110)
  • Test 2: sin(120 → 130)

For each test:

  • The model receives 50 values.
  • It predicts the next time step.

The final plot shows:

  • Training data
  • Test sequences
  • Predicted future values

This demonstrates the RNN’s ability to extrapolate temporal patterns.

How to Run

  1. Install dependencies:
pip install torch torchvision matplotlib numpy jupyter
  1. Run the script:
jupyter notebook RNN.ipynb

Key Concepts Demonstrated

  • Sequence-to-one prediction
  • Sliding window dataset
  • RNN hidden states
  • Time series forecasting
  • PyTorch DataLoader usage

Possible Improvements

  • Replace RNN with LSTM or GRU for better long-term memory
  • Normalize data for more stable training
  • Predict multiple future steps (sequence-to-sequence)
  • Add GPU support (.to(device))
  • Save and load trained models

Conclusion

This project is a minimal yet complete example of using RNNs for time series prediction in PyTorch. It demonstrates how sequential data can be transformed into a supervised learning problem and how neural networks can learn temporal dependencies.