1import torch
2import torch.nn as nn
3from torch.nn import functional as F
4
5# hyperparameters
6batch_size = 64 # how many independent sequences will we process in parallel?
7block_size = 256 # what is the maximum context length for predictions?
8max_iters = 5000
9eval_interval = 500
10learning_rate = 3e-4
11device = 'cuda' if torch.cuda.is_available() else 'cpu'
12eval_iters = 200
13n_embd = 384
14n_head = 6
15n_layer = 6
16dropout = 0.2
17# ------------
18
19torch.manual_seed(1337)
20
21# wget https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt
22with open('input.txt', 'r', encoding='utf-8') as f:
23    text = f.read()
24
25# here are all the unique characters that occur in this text
26chars = sorted(list(set(text)))
27vocab_size = len(chars)
28# create a mapping from characters to integers
29stoi = { ch:i for i,ch in enumerate(chars) }
30itos = { i:ch for i,ch in enumerate(chars) }
31encode = lambda s: [stoi[c] for c in s] # encoder: take a string, output a list of integers
32decode = lambda l: ''.join([itos[i] for i in l]) # decoder: take a list of integers, output a string
33
34# Train and test splits
35data = torch.tensor(encode(text), dtype=torch.long)
36n = int(0.9*len(data)) # first 90% will be train, rest val
37train_data = data[:n]
38val_data = data[n:]
39
40# data loading
41def get_batch(split):
42    # generate a small batch of data of inputs x and targets y
43    data = train_data if split == 'train' else val_data
44    ix = torch.randint(len(data) - block_size, (batch_size,))
45    x = torch.stack([data[i:i+block_size] for i in ix])
46    y = torch.stack([data[i+1:i+block_size+1] for i in ix])
47    x, y = x.to(device), y.to(device)
48    return x, y
49
50@torch.no_grad()
51def estimate_loss():
52    out = {}
53    model.eval()
54    for split in ['train', 'val']:
55        losses = torch.zeros(eval_iters)
56        for k in range(eval_iters):
57            X, Y = get_batch(split)
58            logits, loss = model(X, Y)
59            losses[k] = loss.item()
60        out[split] = losses.mean()
61    model.train()
62    return out
63
64class Head(nn.Module):
65    """ one head of self-attention """
66
67    def __init__(self, head_size):
68        super().__init__()
69        self.key = nn.Linear(n_embd, head_size, bias=False)
70        self.query = nn.Linear(n_embd, head_size, bias=False)
71        self.value = nn.Linear(n_embd, head_size, bias=False)
72        self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size)))
73
74        self.dropout = nn.Dropout(dropout)
75
76    def forward(self, x):
77        # input of size (batch, time-step, channels)
78        # output of size (batch, time-step, head size)
79        B,T,C = x.shape
80        k = self.key(x)   # (B,T,hs)
81        q = self.query(x) # (B,T,hs)
82        # compute attention scores ("affinities")
83        wei = q @ k.transpose(-2,-1) * k.shape[-1]**-0.5 # (B, T, hs) @ (B, hs, T) -> (B, T, T)
84        wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf')) # (B, T, T)
85        wei = F.softmax(wei, dim=-1) # (B, T, T)
86        wei = self.dropout(wei)
87        # perform the weighted aggregation of the values
88        v = self.value(x) # (B,T,hs)
89        out = wei @ v # (B, T, T) @ (B, T, hs) -> (B, T, hs)
90        return out
91
92class MultiHeadAttention(nn.Module):
93    """ multiple heads of self-attention in parallel """
94
95    def __init__(self, num_heads, head_size):
96        super().__init__()
97        self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)])
98        self.proj = nn.Linear(head_size * num_heads, n_embd)
99        self.dropout = nn.Dropout(dropout)
100
101    def forward(self, x):
102        out = torch.cat([h(x) for h in self.heads], dim=-1)
103        out = self.dropout(self.proj(out))
104        return out
105
106class FeedFoward(nn.Module):
107    """ a simple linear layer followed by a non-linearity """
108
109    def __init__(self, n_embd):
110        super().__init__()
111        self.net = nn.Sequential(
112            nn.Linear(n_embd, 4 * n_embd),
113            nn.ReLU(),
114            nn.Linear(4 * n_embd, n_embd),
115            nn.Dropout(dropout),
116        )
117
118    def forward(self, x):
119        return self.net(x)
120
121class Block(nn.Module):
122    """ Transformer block: communication followed by computation """
123
124    def __init__(self, n_embd, n_head):
125        # n_embd: embedding dimension, n_head: the number of heads we'd like
126        super().__init__()
127        head_size = n_embd // n_head
128        self.sa = MultiHeadAttention(n_head, head_size)
129        self.ffwd = FeedFoward(n_embd)
130        self.ln1 = nn.LayerNorm(n_embd)
131        self.ln2 = nn.LayerNorm(n_embd)
132
133    def forward(self, x):
134        x = x + self.sa(self.ln1(x))
135        x = x + self.ffwd(self.ln2(x))
136        return x
137
138class GPTLanguageModel(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 table
143        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 norm
147        self.lm_head = nn.Linear(n_embd, vocab_size)
148
149        # better init, not covered in the original GPT video, but important, will cover in followup video
150        self.apply(self._init_weights)
151
152    def _init_weights(self, module):
153        if isinstance(module, nn.Linear):
154            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
155            if module.bias is not None:
156                torch.nn.init.zeros_(module.bias)
157        elif isinstance(module, nn.Embedding):
158            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
159
160    def forward(self, idx, targets=None):
161        B, T = idx.shape
162
163        # idx and targets are both (B,T) tensor of integers
164        tok_emb = self.token_embedding_table(idx) # (B,T,C)
165        pos_emb = self.position_embedding_table(torch.arange(T, device=device)) # (T,C)
166        x = tok_emb + pos_emb # (B,T,C)
167        x = self.blocks(x) # (B,T,C)
168        x = self.ln_f(x) # (B,T,C)
169        logits = self.lm_head(x) # (B,T,vocab_size)
170
171        if targets is None:
172            loss = None
173        else:
174            B, T, C = logits.shape
175            logits = logits.view(B*T, C)
176            targets = targets.view(B*T)
177            loss = F.cross_entropy(logits, targets)
178
179        return logits, loss
180
181    def generate(self, idx, max_new_tokens):
182        # idx is (B, T) array of indices in the current context
183        for _ in range(max_new_tokens):
184            # crop idx to the last block_size tokens
185            idx_cond = idx[:, -block_size:]
186            # get the predictions
187            logits, loss = self(idx_cond)
188            # focus only on the last time step
189            logits = logits[:, -1, :] # becomes (B, C)
190            # apply softmax to get probabilities
191            probs = F.softmax(logits, dim=-1) # (B, C)
192            # sample from the distribution
193            idx_next = torch.multinomial(probs, num_samples=1) # (B, 1)
194            # append sampled index to the running sequence
195            idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)
196        return idx
197
198model = GPTLanguageModel()
199m = model.to(device)
200# print the number of parameters in the model
201print(sum(p.numel() for p in m.parameters())/1e6, 'M parameters')
202
203# create a PyTorch optimizer
204optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
205
206for iter in range(max_iters):
207
208    # every once in a while evaluate the loss on train and val sets
209    if iter % eval_interval == 0 or iter == max_iters - 1:
210        losses = estimate_loss()
211        print(f"step {iter}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")
212
213    # sample a batch of data
214    xb, yb = get_batch('train')
215
216    # evaluate the loss
217    logits, loss = model(xb, yb)
218    optimizer.zero_grad(set_to_none=True)
219    loss.backward()
220    optimizer.step()
221
222# generate from the model
223context = torch.zeros((1, 1), dtype=torch.long, device=device)
224print(decode(m.generate(context, max_new_tokens=500)[0].tolist()))
225#open('more.txt', 'w').write(decode(m.generate(context, max_new_tokens=10000)[0].tolist()))
226