Stoopidity/HarryPotterLLM

1import torch2import torch.nn as nn3from torch.nn import functional as F4import gradio as gr5 6# hyperparameters7batch_size = 16 # how many independent sequences will we process in parallel?8block_size = 32 # what is the maximum context length for predictions?9max_iters = 500010eval_interval = 10011learning_rate = 1e-312device = 'cuda' if torch.cuda.is_available() else 'cpu'13eval_iters = 20014n_embd = 6415n_head = 416n_layer = 417dropout = 0.018# ------------19 20torch.manual_seed(1337)21 22# wget https://github.com/Cral-Cactus/HarryPotterLLM/raw/main/HarryPotter.txt23with open('HarryPotter.txt', 'r', encoding='utf-8') as f:24    text = f.read()25 26# here are all the unique characters that occur in this text27chars = sorted(list(set(text)))28vocab_size = len(chars)29# create a mapping from characters to integers30stoi = { ch:i for i,ch in enumerate(chars) }31itos = { i:ch for i,ch in enumerate(chars) }32encode = lambda s: [stoi[c] for c in s] # encoder: take a string, output a list of integers33decode = lambda l: ''.join([itos[i] for i in l]) # decoder: take a list of integers, output a string34 35# Train and test splits36data = torch.tensor(encode(text), dtype=torch.long)37n = int(0.9*len(data)) # first 90% will be train, rest val38train_data = data[:n]39val_data = data[n:]40 41# data loading42def get_batch(split):43    # generate a small batch of data of inputs x and targets y44    data = train_data if split == 'train' else val_data45    ix = torch.randint(len(data) - block_size, (batch_size,))46    x = torch.stack([data[i:i+block_size] for i in ix])47    y = torch.stack([data[i+1:i+block_size+1] for i in ix])48    x, y = x.to(device), y.to(device)49    return x, y50 51@torch.no_grad()52def estimate_loss():53    out = {}54    model.eval()55    for split in ['train', 'val']:56        losses = torch.zeros(eval_iters)57        for k in range(eval_iters):58            X, Y = get_batch(split)59            logits, loss = model(X, Y)60            losses[k] = loss.item()61        out[split] = losses.mean()62    model.train()63    return out64 65class Head(nn.Module):66    """ one head of self-attention """67 68    def __init__(self, head_size):69        super().__init__()70        self.key = nn.Linear(n_embd, head_size, bias=False)71        self.query = nn.Linear(n_embd, head_size, bias=False)72        self.value = nn.Linear(n_embd, head_size, bias=False)73        self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size)))74 75        self.dropout = nn.Dropout(dropout)76 77    def forward(self, x):78        B,T,C = x.shape79        k = self.key(x)   # (B,T,C)80        q = self.query(x) # (B,T,C)81        # compute attention scores ("affinities")82        wei = q @ k.transpose(-2,-1) * C**-0.5 # (B, T, C) @ (B, C, T) -> (B, T, T)83        wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf')) # (B, T, T)84        wei = F.softmax(wei, dim=-1) # (B, T, T)85        wei = self.dropout(wei)86        # perform the weighted aggregation of the values87        v = self.value(x) # (B,T,C)88        out = wei @ v # (B, T, T) @ (B, T, C) -> (B, T, C)89        return out90 91class MultiHeadAttention(nn.Module):92    """ multiple heads of self-attention in parallel """93 94    def __init__(self, num_heads, head_size):95        super().__init__()96        self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)])97        self.proj = nn.Linear(n_embd, n_embd)98        self.dropout = nn.Dropout(dropout)99 100    def forward(self, x):101        out = torch.cat([h(x) for h in self.heads], dim=-1)102        out = self.dropout(self.proj(out))103        return out104 105class FeedFoward(nn.Module):106    """ a simple linear layer followed by a non-linearity """107 108    def __init__(self, n_embd):109        super().__init__()110        self.net = nn.Sequential(111            nn.Linear(n_embd, 4 * n_embd),112            nn.ReLU(),113            nn.Linear(4 * n_embd, n_embd),114            nn.Dropout(dropout),115        )116 117    def forward(self, x):118        return self.net(x)119 120class Block(nn.Module):121    """ Transformer block: communication followed by computation """122 123    def __init__(self, n_embd, n_head):124        # n_embd: embedding dimension, n_head: the number of heads we'd like125        super().__init__()126        head_size = n_embd // n_head127        self.sa = MultiHeadAttention(n_head, head_size)128        self.ffwd = FeedFoward(n_embd)129        self.ln1 = nn.LayerNorm(n_embd)130        self.ln2 = nn.LayerNorm(n_embd)131 132    def forward(self, x):133        x = x + self.sa(self.ln1(x))134        x = x + self.ffwd(self.ln2(x))135        return x136 137# super simple bigram model138class BigramLanguageModel(nn.Module):139 140    def __init__(self):141        super().__init__()142        # each token directly reads off the logits for the next token from a lookup table143        self.token_embedding_table = nn.Embedding(vocab_size, n_embd)144        self.position_embedding_table = nn.Embedding(block_size, n_embd)145        self.blocks = nn.Sequential(*[Block(n_embd, n_head=n_head) for _ in range(n_layer)])146        self.ln_f = nn.LayerNorm(n_embd) # final layer norm147        self.lm_head = nn.Linear(n_embd, vocab_size)148 149    def forward(self, idx, targets=None):150        B, T = idx.shape151 152        # idx and targets are both (B,T) tensor of integers153        tok_emb = self.token_embedding_table(idx) # (B,T,C)154        pos_emb = self.position_embedding_table(torch.arange(T, device=device)) # (T,C)155        x = tok_emb + pos_emb # (B,T,C)156        x = self.blocks(x) # (B,T,C)157        x = self.ln_f(x) # (B,T,C)158        logits = self.lm_head(x) # (B,T,vocab_size)159 160        if targets is None:161            loss = None162        else:163            B, T, C = logits.shape164            logits = logits.view(B*T, C)165            targets = targets.view(B*T)166            loss = F.cross_entropy(logits, targets)167 168        return logits, loss169 170    def generate(self, idx, max_new_tokens):171        # idx is (B, T) array of indices in the current context172        for _ in range(max_new_tokens):173            # crop idx to the last block_size tokens174            idx_cond = idx[:, -block_size:]175            # get the predictions176            logits, loss = self(idx_cond)177            # focus only on the last time step178            logits = logits[:, -1, :] # becomes (B, C)179            # apply softmax to get probabilities180            probs = F.softmax(logits, dim=-1) # (B, C)181            # sample from the distribution182            idx_next = torch.multinomial(probs, num_samples=1) # (B, 1)183            # append sampled index to the running sequence184            idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)185        return idx186 187model = BigramLanguageModel()188m = model.to(device)189m.load_state_dict(torch.load("state.txt",map_location = torch.device(device)))190 191def generate_text(context,mt):192    return decode(m.generate(torch.tensor([encode(context)]), max_new_tokens=mt)[0].tolist())193 194iface = gr.Interface(195    fn=generate_text,196    inputs=[197        gr.Textbox(label="Prompt", placeholder="Put something here!!!"),198        gr.Slider(minimum=1, maximum=1000, step=1, label="Number of characters to generate", value=100)199    ],200    outputs=gr.Textbox(label="Generated Text"),201    title="Name of your bot",202    description="Add a description here!"203)204 205# Launch the interface206if __name__ == "__main__":207    iface.launch()