weno-ai/leaf

1import torch2import torch.nn as nn3from torch.nn import functional as F4import json5import os6 7# --- Hyperparameters ---8# These are the settings for our model. You can experiment with these values.9batch_size = 32  # How many sequences to process in parallel10block_size = 8  # Maximum context length for predictions11max_iters = 3000  # Number of training iterations12eval_interval = 300  # How often to evaluate the model13learning_rate = 1e-2  # The learning rate for the optimizer14device = 'cuda' if torch.cuda.is_available() else 'cpu'  # Use GPU if available15eval_iters = 200  # Number of iterations for evaluation16n_embd = 32  # The dimension of the token embeddings17n_head = 4  # The number of attention heads in the Multi-Head Attention block18n_layer = 4  # The number of Transformer blocks19dropout = 0.0  # Dropout rate for regularization20 21# --- Data Preparation ---22# To use this code, you need to create a file named 'dataset.jsonl'23# in the same directory as this script. Each line of the file should be a JSON object24# with 'header' and 'formal_statement' keys, like the example you provided.25file_path = 'dataset.jsonl'26 27# Process the JSONL data from the file.28corpus = ""29try:30    with open(file_path, 'r') as f:31        for line in f:32            data_point = json.loads(line)33            # Combine the 'header' and 'formal_statement' fields.34            # We add a newline character to separate the two parts of the text.35            corpus += data_point['header'] + '\n' + data_point['formal_statement'] + '\n'36except FileNotFoundError:37    print(f"Error: The file '{file_path}' was not found. Please create it and add your data.")38    exit()39except json.JSONDecodeError:40    print(f"Error: There was a problem parsing a line in '{file_path}'. Make sure each line is a valid JSON object.")41    exit()42except KeyError:43    print(f"Error: A line in '{file_path}' does not have the 'header' or 'formal_statement' keys. Please check your JSONL file format.")44    exit()45 46# Check if the corpus is empty after loading the file.47if not corpus:48    print(f"Error: The corpus is empty. This could be because '{file_path}' is empty or contains no valid text.")49    exit()50 51# Here we create a simple character-level tokenizer.52# The vocabulary consists of all unique characters in the text.53chars = sorted(list(set(corpus)))54vocab_size = len(chars)55stoi = {ch: i for i, ch in enumerate(chars)}56itos = {i: ch for i, ch in enumerate(chars)}57# Fix the bug in the encode function. The loop variable was 's' instead of 'c'.58encode = lambda s: [stoi[c] for c in s]59decode = lambda l: ''.join([itos[i] for i in l])60 61# Convert the entire text into a PyTorch tensor.62data = torch.tensor(encode(corpus), dtype=torch.long)63 64# Create a simple train/validation split.65n = int(0.9 * len(data))66train_data = data[:n]67val_data = data[n:]68 69# --- Helper Functions ---70# This function gets a random batch of data from either the training or validation set.71def get_batch(split):72    data = train_data if split == 'train' else val_data73    # Generate random starting indices for each sequence in the batch.74    ix = torch.randint(len(data) - block_size, (batch_size,))75    # Stack the sequences to create a batch.76    x = torch.stack([data[i:i + block_size] for i in ix])77    y = torch.stack([data[i + 1:i + block_size + 1] for i in ix])78    x, y = x.to(device), y.to(device)79    return x, y80 81# This function is used to estimate the model's loss on both the train and validation sets.82# It uses torch.no_grad() to make the process more efficient as we're not training.83@torch.no_grad()84def estimate_loss():85    out = {}86    model.eval()  # Set the model to evaluation mode.87    for split in ['train', 'val']:88        losses = torch.zeros(eval_iters)89        for k in range(eval_iters):90            X, Y = get_batch(split)91            logits, loss = model(X, Y)92            losses[k] = loss.item()93        out[split] = losses.mean()94    model.train()  # Set the model back to training mode.95    return out96 97# --- The Self-Attention Mechanism ---98# This is a single attention head.99class Head(nn.Module):100    def __init__(self, head_size):101        super().__init__()102        # Linear layers to project the input into key, query, and value vectors.103        self.key = nn.Linear(n_embd, head_size, bias=False)104        self.query = nn.Linear(n_embd, head_size, bias=False)105        self.value = nn.Linear(n_embd, head_size, bias=False)106        # A buffer to store a lower-triangular matrix, which prevents future tokens from107        # "seeing" past tokens (decoder-style attention).108        self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size)))109        # Dropout layer for regularization.110        self.dropout = nn.Dropout(dropout)111 112    def forward(self, x):113        B, T, C = x.shape114        k = self.key(x)  # (B, T, head_size)115        q = self.query(x)  # (B, T, head_size)116        117        # Compute the affinity scores (weights).118        # (q @ k.transpose(-2, -1)) is matrix multiplication of q and k transpose.119        wei = q @ k.transpose(-2, -1) * C**-0.5  # (B, T, head_size) @ (B, head_size, T) -> (B, T, T)120        # Apply the lower-triangular mask to enforce causality.121        wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf'))122        # Apply softmax to get the attention weights.123        wei = F.softmax(wei, dim=-1)124        self.dropout(wei)125 126        v = self.value(x)  # (B, T, head_size)127        out = wei @ v  # (B, T, T) @ (B, T, head_size) -> (B, T, head_size)128        return out129 130# This combines multiple attention heads in parallel.131class MultiHeadAttention(nn.Module):132    def __init__(self, num_heads, head_size):133        super().__init__()134        # Create a list of `Head` modules.135        self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)])136        # A final linear layer to project the concatenated output of all heads.137        self.proj = nn.Linear(num_heads * head_size, n_embd)138        self.dropout = nn.Dropout(dropout)139 140    def forward(self, x):141        # Concatenate the output from each head.142        out = torch.cat([h(x) for h in self.heads], dim=-1)143        out = self.dropout(self.proj(out))144        return out145 146# This is a simple feed-forward network.147class FeedFoward(nn.Module):148    def __init__(self, n_embd):149        super().__init__()150        # A simple linear-ReLU-linear stack.151        self.net = nn.Sequential(152            nn.Linear(n_embd, 4 * n_embd),153            nn.ReLU(),154            nn.Linear(4 * n_embd, n_embd),155            nn.Dropout(dropout),156        )157 158    def forward(self, x):159        return self.net(x)160 161# This is a single Transformer block, composed of Multi-Head Attention and a Feed-Forward network.162class TransformerBlock(nn.Module):163    def __init__(self, n_embd, n_head):164        super().__init__()165        head_size = n_embd // n_head166        # The attention mechanism.167        self.sa = MultiHeadAttention(n_head, head_size)168        # The feed-forward network.169        self.ffwd = FeedFoward(n_embd)170        # Layer normalization layers.171        self.ln1 = nn.LayerNorm(n_embd)172        self.ln2 = nn.LayerNorm(n_embd)173 174    def forward(self, x):175        # Apply self-attention with a residual connection and layer normalization.176        x = x + self.sa(self.ln1(x))177        # Apply feed-forward with another residual connection and layer normalization.178        x = x + self.ffwd(self.ln2(x))179        return x180 181# --- The Main Language Model ---182class LanguageModel(nn.Module):183    def __init__(self):184        super().__init__()185        # A token embedding table: each integer token gets a vector representation.186        self.token_embedding_table = nn.Embedding(vocab_size, n_embd)187        # A positional embedding table: each position gets a vector representation.188        self.position_embedding_table = nn.Embedding(block_size, n_embd)189        # A sequence of Transformer blocks.190        self.blocks = nn.Sequential(*[TransformerBlock(n_embd, n_head) for _ in range(n_layer)])191        # A final layer normalization.192        self.ln_f = nn.LayerNorm(n_embd)193        # A linear layer to project the final embeddings to the vocabulary size.194        self.lm_head = nn.Linear(n_embd, vocab_size)195 196    def forward(self, idx, targets=None):197        B, T = idx.shape198 199        # Get token embeddings and positional embeddings.200        tok_emb = self.token_embedding_table(idx)  # (B, T, C)201        pos_emb = self.position_embedding_table(torch.arange(T, device=device))  # (T, C)202        # Add them together to get the final embeddings.203        x = tok_emb + pos_emb  # (B, T, C)204        # Pass through the Transformer blocks.205        x = self.blocks(x)206        x = self.ln_f(x)207        # Project to the vocabulary size.208        logits = self.lm_head(x)  # (B, T, vocab_size)209 210        loss = None211        if targets is not None:212            # Reshape for cross-entropy loss calculation.213            B, T, C = logits.shape214            logits = logits.view(B * T, C)215            targets = targets.view(B * T)216            loss = F.cross_entropy(logits, targets)217 218        return logits, loss219 220    # A function to generate text.221    def generate(self, idx, max_new_tokens):222        # idx is (B, T) tensor of indices in the current context.223        for _ in range(max_new_tokens):224            # Crop idx to block_size, as the model has a limited context.225            idx_cond = idx[:, -block_size:]226            # Get predictions.227            logits, loss = self(idx_cond)228            # Focus only on the last time step.229            logits = logits[:, -1, :]230            # Apply softmax to get probabilities.231            probs = F.softmax(logits, dim=-1)232            # Sample from the distribution.233            idx_next = torch.multinomial(probs, num_samples=1)234            # Append the new token to the sequence.235            idx = torch.cat((idx, idx_next), dim=1)236        return idx237 238# --- Training and Generation ---239model = LanguageModel()240m = model.to(device)241 242# Create a PyTorch optimizer.243optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)244 245# Main training loop.246for iter in range(max_iters):247    # Every few iterations, evaluate the loss on both splits.248    if iter % eval_interval == 0:249        losses = estimate_loss()250        print(f"step {iter}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")251 252    # Sample a batch of data.253    xb, yb = get_batch('train')254 255    # Forward pass: compute loss.256    logits, loss = model(xb, yb)257    # Backward pass: compute gradients.258    optimizer.zero_grad(set_to_none=True)259    loss.backward()260    # Update the model parameters.261    optimizer.step()262 263# --- Generate new text from the trained model ---264context = torch.zeros((1, 1), dtype=torch.long, device=device)265generated_text_indices = m.generate(context, max_new_tokens=20)266print("\nGenerated text:")267print(decode(generated_text_indices[0].tolist()))