pages/17_RNN.py

import streamlit as st
import torch
import torch.nn as nn
import numpy as np

# Introduction
st.title('RNN Character Prediction')
st.write("""
This app demonstrates how to train a Recurrent Neural Network (RNN) to predict the next character in a given string. 
The RNN learns the sequence of characters from a provided text and generates text based on the learned patterns.
You can choose different options for training the RNN and see how it affects the generated text.
""")

# User input for the training string
sequence = st.text_area('Enter the training string:', 'In the vast expanse of the digital realm.')
chars = list(set(sequence))
data_size, vocab_size = len(sequence), len(chars)

# Create mappings from characters to indices and vice versa
char_to_idx = {ch: i for i, ch in enumerate(chars)}
idx_to_char = {i: ch for i, ch in enumerate(chars)}

# Convert the sequence to indices
indices = np.array([char_to_idx[ch] for ch in sequence])

class RNN(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super(RNN, self).__init__()
        self.hidden_size = hidden_size
        self.i2h = nn.Linear(input_size + hidden_size, hidden_size)
        self.i2o = nn.Linear(input_size + hidden_size, output_size)
        self.softmax = nn.LogSoftmax(dim=1)

    def forward(self, input, hidden):
        combined = torch.cat((input, hidden), 1)
        hidden = self.i2h(combined)
        output = self.i2o(combined)
        output = self.softmax(output)
        return output, hidden

    def init_hidden(self):
        return torch.zeros(1, self.hidden_size)

# Hyperparameters
n_hidden = 128
learning_rate = 0.005

# Initialize the model, loss function, and optimizer
rnn = RNN(vocab_size, n_hidden, vocab_size)
criterion = nn.NLLLoss()

# Define training options
options = {
    'Quick Train (100 epochs)': 100,
    'Medium Train (500 epochs)': 500,
    'Long Train (1000 epochs)': 1000
}

train_option = st.selectbox('Select training option:', list(options.keys()))
n_epochs = options[train_option]
optimizer = torch.optim.Adam(rnn.parameters(), lr=learning_rate)

def char_tensor(char):
    if char not in char_to_idx:
        raise ValueError(f"Character '{char}' not in vocabulary.")
    tensor = torch.zeros(1, vocab_size)
    tensor[0][char_to_idx[char]] = 1
    return tensor

# Training function
def train_model(n_epochs):
    for epoch in range(n_epochs):
        hidden = rnn.init_hidden()
        rnn.zero_grad()
        loss = 0

        for i in range(data_size - 1):
            input_char = char_tensor(sequence[i])
            target_char = torch.tensor([char_to_idx[sequence[i + 1]]], dtype=torch.long)

            output, hidden = rnn(input_char, hidden)
            loss += criterion(output, target_char)

        loss.backward()
        optimizer.step()

        if epoch % 10 == 0:
            st.write(f'Epoch {epoch} loss: {loss.item() / (data_size - 1)}')

    st.write("Training complete.")

# Train the model
if st.button('Train Model'):
    train_model(n_epochs)

def generate(start_char, predict_len=100):
    if start_char not in char_to_idx:
        raise ValueError(f"Start character '{start_char}' not in vocabulary.")
    hidden = rnn.init_hidden()
    input_char = char_tensor(start_char)
    predicted_str = start_char

    for _ in range(predict_len):
        output, hidden = rnn(input_char, hidden)
        topv, topi = output.topk(1)
        predicted_char_idx = topi[0][0].item()
        predicted_char = idx_to_char[predicted_char_idx]
        predicted_str += predicted_char
        input_char = char_tensor(predicted_char)

    return predicted_str

start_char = st.text_input('Enter a starting character:', 'h')
predict_len = st.slider('Select the length of the generated text:', min_value=10, max_value=200, value=50)

if st.button('Generate Text'):
    try:
        generated_text = generate(start_char, predict_len)
        st.write('Generated Text:')
        st.text(generated_text)
    except ValueError as e:
        st.error(str(e))