train.py

import torch
import torch.nn as nn
import torch.optim as optim
import os
import json
from tqdm import tqdm, trange
import time

# Generate simple training data
training_text = open("train_data.txt", encoding="utf-8").read()
chars = sorted(list(set(training_text)))  # Unique characters
char_to_idx = {ch: i for i, ch in enumerate(chars)}
idx_to_char = {i: ch for i, ch in enumerate(chars)}

# Model parameters
input_size = len(chars)
hidden_size = 32
output_size = len(chars)
sequence_length = 5
epochs = 1000
learning_rate = 0.0001
model_path = "tiny_llm.pth"

# Create training data (input-output pairs)
train_data = []
for i in range(len(training_text) - sequence_length):
    input_seq = training_text[i : i + sequence_length]
    target_char = training_text[i + sequence_length]
    train_data.append((torch.tensor([char_to_idx[ch] for ch in input_seq]), char_to_idx[target_char]))

# Define the simple RNN model
class SimpleRNN(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super(SimpleRNN, self).__init__()
        self.hidden_size = hidden_size
        self.rnn = nn.RNN(input_size, hidden_size, batch_first=True)
        self.fc = nn.Linear(hidden_size, output_size)

    def forward(self, x, hidden):
        x = torch.nn.functional.one_hot(x, num_classes=input_size).float()
        out, hidden = self.rnn(x.unsqueeze(0), hidden)
        out = self.fc(out[:, -1, :])  # Take last time step's output
        return out, hidden

# Load model if available
if os.path.exists(model_path):
    model = torch.load(model_path, weights_only=False)
    with open("vocab.json", "r") as f:
        chars = json.loads(f.read())
    char_to_idx = {ch: i for i, ch in enumerate(chars)}
    idx_to_char = {i: ch for i, ch in enumerate(chars)}
    print("Loaded pre-trained model.")
else:
    print("Training new model...")
    # Initialize the model
    model = SimpleRNN(input_size, hidden_size, output_size)
    criterion = nn.CrossEntropyLoss()
    optimizer = optim.Adam(model.parameters(), lr=learning_rate)
    for epoch in range(epochs):
        try:
            total_loss = 0
            hidden = torch.zeros(1, 1, hidden_size)
            
            pbar = tqdm(train_data, desc=f"Epoch={epoch}, Loss=N/A")
            count = 0
            for input_seq, target in pbar:
                count += 1
                optimizer.zero_grad()
                output, hidden = model(input_seq, hidden.detach())
                loss = criterion(output, torch.tensor([target]))
                loss.backward()
                optimizer.step()
                total_loss += loss.item()
                pbar.desc = f"Epoch={epoch}, Loss={total_loss / count:.12f}"
            
            pbar.close()
            time.sleep(1)
        except KeyboardInterrupt:
            break
    
    hidden = torch.zeros(1, 1, hidden_size)
    output, hidden = model(input_seq, hidden.detach())

    # Save the trained model
    torch.save(model, model_path)
    with open("vocab.json", "w") as f:
        f.write(json.dumps(chars))
    print("Model saved.")

# Text generation function
def generate_text(start_text, length=10000):
    model.eval()
    hidden = torch.zeros(1, 1, hidden_size)
    input_seq = torch.tensor([char_to_idx[ch] for ch in start_text])

    generated_text = start_text
    for _ in trange(length):
        output, hidden = model(input_seq, hidden)
        predicted_idx = output.argmax().item()
        generated_text += idx_to_char[predicted_idx]
        input_seq = torch.cat((input_seq[1:], torch.tensor([predicted_idx])))

    return generated_text

# Generate some text

while True:
    print("LLM Output: ", generate_text(input("Ask LLM: ")))