Multiple Heads, Feedforward Layer#
|
import torch
import torch.nn as nn
from torch.nn import functional as F
Hyperparameters#
B = 32 # B: how many independent sequences will we process in parallel?
T = 8 # T: what is the maximum context length for predictions?
C = 32 # C: numer of different features analysed (also D = dims)\
H = 4 # H: number of attention heads
max_iters = 5000
eval_interval = 500
learning_rate = 1e-3
device = 'cuda' if torch.cuda.is_available() else 'cpu'
eval_iters = 200
torch.manual_seed(1337)
<torch._C.Generator at 0x7f5fbc33a210>
Data Loading#
!wget https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt
--2024-06-09 02:00:55-- https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 185.199.111.133, 185.199.109.133, 185.199.108.133, ...
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|185.199.111.133|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 1115394 (1.1M) [text/plain]
Saving to: ‘input.txt.1’
input.txt.1 0%[ ] 0 --.-KB/s
input.txt.1 100%[===================>] 1.06M --.-KB/s in 0.06s
2024-06-09 02:00:55 (18.1 MB/s) - ‘input.txt.1’ saved [1115394/1115394]
with open('input.txt', 'r', encoding='utf-8') as f:
text = f.read()
# here are all the unique characters that occur in this text
chars = sorted(list(set(text)))
vocab_size = len(chars)
# create a mapping from characters to integers
stoi = { ch:i for i,ch in enumerate(chars) }
itos = { i:ch for i,ch in enumerate(chars) }
encode = lambda s: [stoi[c] for c in s] # encoder: take a string, output a list of integers
decode = lambda l: ''.join([itos[i] for i in l]) # decoder: take a list of integers, output a string
chars_str = ''.join(chars)
print(f'vocab_size: {vocab_size}')
print(f'vocabulary: {chars_str}')
# Train and test splits
data = torch.tensor(encode(text), dtype=torch.long)
n = int(0.9*len(data)) # first 90% will be train, rest val
train_data = data[:n]
val_data = data[n:]
def get_batch(split):
# generate a small batch of data of inputs x and targets y
data = train_data if split == 'train' else val_data
ix = torch.randint(len(data) - T, (B,))
x = torch.stack([data[i:i+T] for i in ix])
y = torch.stack([data[i+1:i+T+1] for i in ix])
x, y = x.to(device), y.to(device)
return x, y
vocab_size: 65
vocabulary:
!$&',-.3:;?ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz
Attention Head#
class Head(nn.Module):
""" One head of self attention"""
def __init__(self, Ci, Co):
super().__init__()
self.key = nn.Linear(Ci, Co, bias=False)
self.query = nn.Linear(Ci, Co, bias=False)
self.value = nn.Linear(Ci, Co, bias=False)
self.register_buffer('tril', torch.tril(torch.ones(T, T)))
def forward(self, x):
B, T, Ci = x.shape
'''
B - batch # of independant vectors processed
T - time/block/context # of tokens in a context
Ci - channals/dims input # of features in input
'''
k = self.key(x) # (B,T,Co)
q = self.query(x) # (B,T,Co)
# compute attention scores / affinities
wei = q @ k.transpose(-2,-1) # (B,T,Co) @ (B,Co,T) -> (B,T,T)
wei /= C**0.5 # (B,T,T) scaling, bring variance to 1, to prevent softmax clipping
wei = wei.masked_fill(self.tril[:T,:T]==0, float('-inf')) # (B,T,T) Replace upper triangular of wei with -inf
wei = F.softmax(wei, dim=-1) # (B,T,T) -inf -> 0, rest normalized to 1
v = self.value(x) # (B,T,Co)
out = wei @ v # (B,T,T) @ (B,T,Co) = (B,T,Co)
return out
Model#
class BigramLanguageModel(nn.Module):
def __init__(self, B,T,C,H):
super().__init__()
self.B, self.T, self.C, self.H = B,T,C,H
# each token directly reads off the logits for the next token from a lookup table
self.token_embedding_table = nn.Embedding(vocab_size, C) # for every possible token, weights for next token
self.position_embedding_table = nn.Embedding(T, C)
self.heads = nn.ModuleList([Head(Ci=C, Co=C//H) for _ in range(H)]) # 4 heads of 8-dimensional self-attention, for n_embed=32, like a group convolution
self.ffwd = nn.Sequential( # Feedforward network, so the tokens can "think about" what they found in attention.
nn.Linear(C, C),
nn.ReLU(),
)
self.lm_head = nn.Linear(C, vocab_size)
def forward(self, idx, targets=None):
tok_emb = self.token_embedding_table(idx) # (B,T,C=n_embed)
pos_emb = self.position_embedding_table(torch.arange(self.T, device=device)) # (T,C): [0,1,2..T-1]
x = tok_emb + pos_emb # (B,T,C)
x = torch.cat([h(x) for h in self.heads], dim=-1) # (B,T,C), Multi head self attention
x = self.ffwd(x) # (B,T,C), Per token level. B,T act as batch dimensions
logits = self.lm_head(x) # (B,T,vocab_size)
if targets is None:
loss = None
else:
B, T, C = logits.shape
logits = logits.view(B*T, C)
targets = targets.view(B*T)
loss = F.cross_entropy(logits, targets)
return logits, loss
def generate(self, idx, max_new_tokens):
for _ in range(max_new_tokens): # idx is (B, T) array of indices in the current context
idx_cond = idx[:, -self.T:] # crop the last block_size tokens for input
logits, loss = self(idx_cond) # get the predictions
logits = logits[:, -1, :] # (B,T,C) -> (B, C)
probs = F.softmax(logits, dim=-1) # (B, C)
idx_next = torch.multinomial(probs, num_samples=1) # sample from the distribution acc to prob (B, 1)
idx = torch.cat((idx, idx_next), dim=1) # New idx is concat (B, T+1)
return idx
model = BigramLanguageModel(B,T,C,H)
m = model.to(device)
Training#
@torch.no_grad()
def estimate_loss():
out = {}
model.eval()
for split in ['train', 'val']:
losses = torch.zeros(eval_iters)
for k in range(eval_iters):
X, Y = get_batch(split)
logits, loss = model(X, Y)
losses[k] = loss.item()
out[split] = losses.mean()
model.train()
return out
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
for iter in range(max_iters):
if iter % eval_interval == 0: # every once in a while evaluate the loss on train and val sets
losses = estimate_loss()
print(f"step {iter}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")
xb, yb = get_batch('train') # sample a batch of data
# evaluate the loss
logits, loss = model(xb, yb)
optimizer.zero_grad(set_to_none=True)
loss.backward()
optimizer.step()
step 0: train loss 4.1996, val loss 4.1995
step 500: train loss 2.5993, val loss 2.6077
step 1000: train loss 2.4629, val loss 2.4651
step 1500: train loss 2.3974, val loss 2.3951
step 2000: train loss 2.3297, val loss 2.3470
step 2500: train loss 2.3018, val loss 2.3221
step 3000: train loss 2.2828, val loss 2.2936
step 3500: train loss 2.2495, val loss 2.2721
step 4000: train loss 2.2435, val loss 2.2468
step 4500: train loss 2.2286, val loss 2.2411
Inference#
context = torch.ones((1, T), dtype=torch.long, device=device) # start with '\n\n\n\n' as seed
out_ints = m.generate(context, max_new_tokens=500)[0].tolist() # output list of ints
print(decode(out_ints))
Ba hill ind vexe
Whe the hot mes fin.
Cy hirad I four shat son yald hat lods guk- have ave lithr
GLOull
Wllld, with.
BANTAUCHAR:
I ork sak's willl eger aepale ganed ay wouce
song thy in noduace being uliths tot upiord--mard meme the fles,
Tharr, leanven-ty's thy on it in weancepte, digus I the souts pof ther cork, if woth' that he the bre tolf,
An live:
Mrins,
What,
Whorend;
Ant to mal ry fa of thaw a tes tir, to tis argito-, setolf brntens?
I' sit mem, thand yought is an he frifperend tepefur