try to add more context to learn

[KarolDuracz]

Hi. I tried to add more information to the context than just the average of the previous tokens. I added a table that calculates the distance between tokens and adds to wei at the end. But I don't know if that makes sense. I made a global train_mode variable and during training it is set to True but when generating it is changed to False.

this is outputs after 3 times x 5000 iterations. I take 3 training shots. This is result after third training loop on google colab. Because my laptop is 2-3 x times slower than their machines. :)

step 4999: train loss 1.6128, val loss 1.7871

I'm a rookie on this field. Does it improve anything in the model at all?

For example, we have

xs = tensor([18., 47., 56., 57., 58.,  1., 15., 47.])

cm = torch.zeros((8))
cm[0] = xs[0]-xs[1]
cm[1] = xs[1]-xs[2]
cm[2] = xs[2]-xs[3]
cm[3] = xs[3]-xs[4]
cm[4] = xs[4]-xs[5]
cm[5] = xs[5]-xs[6]
cm[6] = xs[6]-xs[7]
cm # tensor([-29.,  -9.,  -1.,  -1.,  57., -14., -32.,   0.])
#cm[torch.arange(7)] = xs[torch.arange(0,7)] - xs[torch.arange(1,8)]
#tensor([-29.,  -9.,  -1.,  -1.,  57., -14., -32.,   0.])

cm

Nor is defentain by his fouwl and speeceef:
What madam; light I come yours crother,
But acctey's sheath with is see;
Ask his tend, here such my brother menest as actiness, my nugleds, destired, as us as mind.
The flesse; and grace. or meheigness, apppecity-mean's young
contens youtful, patient, you are but,
His duke of bety that dead.
Lad, you lipp's sue, best feast man's break of the feel;
And here seet he is dost not minder hereignes?

BBATH:
Till behear, sit from teaths: I may were thee rude now.

PARILINA:
Luce than he
voicess more conduce fiten:
Thy ham swoon. You from are what heart raim'n.
Low, War, as drengentaged, time fatrom what fight them griet remed justices.

CORIOLANUS:
I say, behow, so much down's gently as virtue,
Wlike deempade your penger, thy swoundship, war, face, what,
Wordespers a much'd of from
thee things then at mure pattal her,
As you set greaties, adoug-mue face I
Hom not bed mistrol!

First Servingman:
As was me noble, my answere reoten,
And moutster mockh'd nor agar
Do heaven, pomplarigeus again!
I robes, O, if sweet the hall hast
hatan's haste full sing then hasts. They capfite, noble eye
Ast can drawleign reessirtate, the reportation,
Whiced sea
Eving inteed to much liest as bear bands! I would, what coburpety than thou's merty thee hath;
Juliet's sepeak these love's hotnow,
I please the sain'st may see is ciest. I'll his husband smeen thee;
He last been it, that it then ure!

HAMINNNUS:
Show not thing's dryy:.
Inreal, when then I torchisoon the friends thet to can?

HARD IIV:
I, what Coast him think the ceque in prinhes,
Truth as plase,
That whit as where these that nappite the greation.

PAHN:
I'll womerch's like, noble,
As night my conscien'st; play speed the horn'd;
And know for your fucel'sim: as he do yes.
Now might now, by her kisnow be this lets; there are her;
Or or grast thou met's deed-indreat be his unbdound,
speak him do slaint, give lettain.

HENRY BOLINNGBROKE:
I wear the Mearage, and sheat, whipt the cucord.

GLOUCESTE

This is the modification

if train_mode == True:
  cm = torch.zeros((block_size, block_size))
  cm = cm - wei[:, :, torch.arange(0,cm.shape[0])]
  rol = cm[:, :, torch.arange(0,cm.shape[0])].roll(-1)
  rol[:,:,cm.shape[0]-1] = float("-inf")
  rol[:,:,cm.shape[0]-1].exp()
  wei = wei + cm

Here's the code

import torch
import torch.nn as nn
from torch.nn import functional as F

# hyperparameters
batch_size = 16 # how many independent sequences will we process in parallel?
block_size = 32 # what is the maximum context length for predictions?
max_iters = 5000
eval_interval = 100
learning_rate = 1e-2
device = 'cuda' if torch.cuda.is_available() else 'cpu'
eval_iters = 200
n_embd = 64
n_head = 4
n_layer = 4
dropout = 0.0
# ------------

torch.manual_seed(1337)

# wget https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt
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

# 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:]

train_mode = True

# data loading
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) - block_size, (batch_size,))
    x = torch.stack([data[i:i+block_size] for i in ix])
    y = torch.stack([data[i+1:i+block_size+1] for i in ix])
    x, y = x.to(device), y.to(device)
    return x, y

@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

class Head(nn.Module):
    """ one head of self-attention """

    def __init__(self, head_size):
        super().__init__()
        self.key = nn.Linear(n_embd, head_size, bias=False)
        self.query = nn.Linear(n_embd, head_size, bias=False)
        self.value = nn.Linear(n_embd, head_size, bias=False)
        self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size)))

        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        B,T,C = x.shape
        k = self.key(x)   # (B,T,C)
        q = self.query(x) # (B,T,C)
        # compute attention scores ("affinities")
        wei = q @ k.transpose(-2,-1) * C**-0.5 # (B, T, C) @ (B, C, T) -> (B, T, T)
        # wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf')) # (B, T, T)
        # wei = F.softmax(wei, dim=-1) # (B, T, T)
        # wei = self.dropout(wei)
        # perform the weighted aggregation of the values
        v = self.value(x) # (B,T,C)
        #####
        global train_mode
        if train_mode == True:
          cm = torch.zeros((block_size, block_size))
          #print(wei.shape)
          #print(C)
          #print(torch.arange(0,cm.shape[0]))
          cm = cm - wei[:, :, torch.arange(0,cm.shape[0])]
          
          #print(cm.shape, wei.shape)
          # roll
          rol = cm[:, :, torch.arange(0,cm.shape[0])].roll(-1)
          # change last column
          rol[:,:,cm.shape[0]-1] = float("-inf")
          # and move to 0 by exp
          rol[:,:,cm.shape[0]-1].exp()
          #print(cm.shape, wei.shape)
          wei = wei + cm
        #print(" afeter ", cm.shape, wei.shape)
        wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf')) # (B, T, T)
        wei = F.softmax(wei, dim=-1) # (B, T, T)
        wei = self.dropout(wei)
        out = wei @ v # (B, T, T) @ (B, T, C) -> (B, T, C)
        return out

class MultiHeadAttention(nn.Module):
    """ multiple heads of self-attention in parallel """

    def __init__(self, num_heads, head_size):
        super().__init__()
        self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)])
        self.proj = nn.Linear(n_embd, n_embd)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        out = torch.cat([h(x) for h in self.heads], dim=-1)
        out = self.dropout(self.proj(out))
        return out

class FeedFoward(nn.Module):
    """ a simple linear layer followed by a non-linearity """

    def __init__(self, n_embd):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(n_embd, 4 * n_embd),
            nn.ReLU(),
            nn.Linear(4 * n_embd, n_embd),
            nn.Dropout(dropout),
        )

    def forward(self, x):
        return self.net(x)

class Block(nn.Module):
    """ Transformer block: communication followed by computation """

    def __init__(self, n_embd, n_head):
        # n_embd: embedding dimension, n_head: the number of heads we'd like
        super().__init__()
        head_size = n_embd // n_head
        self.sa = MultiHeadAttention(n_head, head_size)
        self.ffwd = FeedFoward(n_embd)
        self.ln1 = nn.LayerNorm(n_embd)
        self.ln2 = nn.LayerNorm(n_embd)

    def forward(self, x):
        x = x + self.sa(self.ln1(x))
        x = x + self.ffwd(self.ln2(x))
        return x

# super simple bigram model
class BigramLanguageModel(nn.Module):

    def __init__(self):
        super().__init__()
        # each token directly reads off the logits for the next token from a lookup table
        self.token_embedding_table = nn.Embedding(vocab_size, n_embd)
        self.position_embedding_table = nn.Embedding(block_size, n_embd)
        self.blocks = nn.Sequential(*[Block(n_embd, n_head=n_head) for _ in range(n_layer)])
        self.ln_f = nn.LayerNorm(n_embd) # final layer norm
        self.lm_head = nn.Linear(n_embd, vocab_size)

    def forward(self, idx, targets=None):
        B, T = idx.shape

        # idx and targets are both (B,T) tensor of integers
        tok_emb = self.token_embedding_table(idx) # (B,T,C)
        pos_emb = self.position_embedding_table(torch.arange(T, device=device)) # (T,C)
        x = tok_emb + pos_emb # (B,T,C)
        x = self.blocks(x) # (B,T,C)
        x = self.ln_f(x) # (B,T,C)
        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):
        global train_mode
        train_mode = False
        # idx is (B, T) array of indices in the current context
        for _ in range(max_new_tokens):
            # crop idx to the last block_size tokens
            #print(idx)
            idx_cond = idx[:, -block_size:]
            # get the predictions
            logits, loss = self(idx_cond)
            #print(logits, loss)
            # focus only on the last time step
            logits = logits[:, -1, :] # becomes (B, C)
            # apply softmax to get probabilities
            probs = F.softmax(logits, dim=-1) # (B, C)
            # sample from the distribution
            idx_next = torch.multinomial(probs, num_samples=1) # (B, 1)
            # append sampled index to the running sequence
            idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)
        return idx

model = BigramLanguageModel()
m = model.to(device)
# print the number of parameters in the model
print(sum(p.numel() for p in m.parameters())/1e6, 'M parameters')

# create a PyTorch optimizer
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)

Training loop

for iter in range(max_iters):

    # every once in a while evaluate the loss on train and val sets
    if iter % eval_interval == 0 or iter == max_iters - 1:
        losses = estimate_loss()
        print(f"step {iter}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")

    # sample a batch of data
    xb, yb = get_batch('train')

    # evaluate the loss
    logits, loss = model(xb, yb)
    optimizer.zero_grad(set_to_none=True)
    loss.backward()
    optimizer.step()

Example

train_mode = False
context = torch.zeros((1, 1), dtype=torch.long, device=device)
print(decode(m.generate(context, max_new_tokens=2000)[0].tolist()))