Attention is all you need
This revision is from 2024/07/28 07:05. You can Restore it.
import torch
import torch.nn as nn
import torch.nn.functional as F
class MultiHeadAttention(nn.Module):
def __init__(self, d_model, n_heads):
super(MultiHeadAttention, self).__init__()
assert d_model % n_heads == 0, "d_model must be divisible by n_heads"
self.d_model = d_model
self.n_heads = n_heads
self.d_k = d_model // n_heads
self.W_q = nn.Linear(d_model, d_model)
self.W_k = nn.Linear(d_model, d_model)
self.W_v = nn.Linear(d_model, d_model)
self.W_o = nn.Linear(d_model, d_model)
def forward(self, Q, K, V, mask=None):
batch_size = Q.size(0)
Q = self.W_q(Q).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
K = self.W_k(K).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
V = self.W_v(V).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
scores = torch.matmul(Q, K.transpose(-2, -1)) / torch.sqrt(torch.tensor(self.d_k, dtype=torch.float32))
if mask is not None:
scores = scores.masked_fill(mask == 0, -1e9)
attention = F.softmax(scores, dim=-1)
output = torch.matmul(attention, V)
output = output.transpose(1, 2).contiguous().view(batch_size, -1, self.d_model)
output = self.W_o(output)
return output, attention
class PositionwiseFeedForward(nn.Module):
def __init__(self, d_model, d_ff):
super(PositionwiseFeedForward, self).__init__()
self.linear1 = nn.Linear(d_model, d_ff)
self.linear2 = nn.Linear(d_ff, d_model)
def forward(self, x):
return self.linear2(F.relu(self.linear1(x)))
class PositionalEncoding(nn.Module):
def __init__(self, d_model, max_len=5000):
super(PositionalEncoding, self).__init__()
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-torch.log(torch.tensor(10000.0)) / d_model))
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0).transpose(0, 1)
self.register_buffer('pe', pe)
def forward(self, x):
x = x + self.pe[:x.size(0), :]
return x
class EncoderLayer(nn.Module):
def __init__(self, d_model, n_heads, d_ff, dropout=0.1):
super(EncoderLayer, self).__init__()
self.self_attn = MultiHeadAttention(d_model, n_heads)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x, mask=None):
attn_output, _ = self.self_attn(x, x, x, mask)
x = self.norm1(x + self.dropout1(attn_output))
ff_output = self.feed_forward(x)
x = self.norm2(x + self.dropout2(ff_output))
return x
class DecoderLayer(nn.Module):
def __init__(self, d_model, n_heads, d_ff, dropout=0.1):
super(DecoderLayer, self).__init__()
self.self_attn = MultiHeadAttention(d_model, n_heads)
self.src_attn = MultiHeadAttention(d_model, n_heads)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = nn.LayerNorm(d_model)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.dropout3 = nn.Dropout(dropout)
def forward(self, x, memory, src_mask=None, tgt_mask=None):
attn1, _ = self.self_attn(x, x, x, tgt_mask)
x = self.norm1(x + self.dropout1(attn1))
attn2, _ = self.src_attn(x, memory, memory, src_mask)
x = self.norm2(x + self.dropout2(attn2))
ff_output = self.feed_forward(x)
x = self.norm3(x + self.dropout3(ff_output))
return x
class Transformer(nn.Module):
def __init__(self, src_vocab_size, tgt_vocab_size, d_model, n_heads, d_ff, n_layers, dropout=0.1):
super(Transformer, self).__init__()
self.encoder = nn.ModuleList([EncoderLayer(d_model, n_heads, d_ff, dropout) for _ in range(n_layers)])
self.decoder = nn.ModuleList([DecoderLayer(d_model, n_heads, d_ff, dropout) for _ in range(n_layers)])
self.src_embed = nn.Sequential(nn.Embedding(src_vocab_size, d_model), PositionalEncoding(d_model))
self.tgt_embed = nn.Sequential(nn.Embedding(tgt_vocab_size, d_model), PositionalEncoding(d_model))
self.out = nn.Linear(d_model, tgt_vocab_size)
def forward(self, src, tgt, src_mask=None, tgt_mask=None):
enc_output = self.src_embed(src)
for layer in self.encoder:
enc_output = layer(enc_output, src_mask)
dec_output = self.tgt_embed(tgt)
for layer in self.decoder:
dec_output = layer(dec_output, enc_output, src_mask, tgt_mask)
output = self.out(dec_output)
return output
# Example usagesrc_vocab_size = 1000
tgt_vocab_size = 1000
d_model = 512
n_heads = 8
d_ff = 2048
n_layers = 6
dropout = 0.1
model = Transformer(src_vocab_size, tgt_vocab_size, d_model, n_heads, d_ff, n_layers, dropout)
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
import random
import math
class MultiHeadAttention(nn.Module):
def __init__(self, d_model, n_heads):
super(MultiHeadAttention, self).__init__()
assert d_model % n_heads == 0, "d_model must be divisible by n_heads"
self.d_model = d_model
self.n_heads = n_heads
self.d_k = d_model // n_heads
self.W_q = nn.Linear(d_model, d_model)
self.W_k = nn.Linear(d_model, d_model)
self.W_v = nn.Linear(d_model, d_model)
self.W_o = nn.Linear(d_model, d_model)
def forward(self, Q, K, V, mask=None):
batch_size = Q.size(0)
Q = self.W_q(Q).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
K = self.W_k(K).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
V = self.W_v(V).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(self.d_k)
if mask is not None:
scores = scores.masked_fill(mask == 0, -1e9)
attention = F.softmax(scores, dim=-1)
output = torch.matmul(attention, V)
output = output.transpose(1, 2).contiguous().view(batch_size, -1, self.d_model)
output = self.W_o(output)
return output, attention
class PositionwiseFeedForward(nn.Module):
def __init__(self, d_model, d_ff):
super(PositionwiseFeedForward, self).__init__()
self.linear1 = nn.Linear(d_model, d_ff)
self.linear2 = nn.Linear(d_ff, d_model)
def forward(self, x):
return self.linear2(F.relu(self.linear1(x)))
class PositionalEncoding(nn.Module):
def __init__(self, d_model, max_len=5000):
super(PositionalEncoding, self).__init__()
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0).transpose(0, 1)
self.register_buffer('pe', pe)
def forward(self, x):
x = x + self.pe[:x.size(0), :]
return x
class EncoderLayer(nn.Module):
def __init__(self, d_model, n_heads, d_ff, dropout=0.1):
super(EncoderLayer, self).__init__()
self.self_attn = MultiHeadAttention(d_model, n_heads)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x, mask=None):
attn_output, _ = self.self_attn(x, x, x, mask)
x = self.norm1(x + self.dropout1(attn_output))
ff_output = self.feed_forward(x)
x = self.norm2(x + self.dropout2(ff_output))
return x
class DecoderLayer(nn.Module):
def __init__(self, d_model, n_heads, d_ff, dropout=0.1):
super(DecoderLayer, self).__init__()
self.self_attn = MultiHeadAttention(d_model, n_heads)
self.src_attn = MultiHeadAttention(d_model, n_heads)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = nn.LayerNorm(d_model)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.dropout3 = nn.Dropout(dropout)
def forward(self, x, memory, src_mask=None, tgt_mask=None):
attn1, _ = self.self_attn(x, x, x, tgt_mask)
x = self.norm1(x + self.dropout1(attn1))
attn2, _ = self.src_attn(x, memory, memory, src_mask)
x = self.norm2(x + self.dropout2(attn2))
ff_output = self.feed_forward(x)
x = self.norm3(x + self.dropout3(ff_output))
return x
class Transformer(nn.Module):
def __init__(self, src_vocab_size, tgt_vocab_size, d_model, n_heads, d_ff, n_layers, dropout=0.1):
super(Transformer, self).__init__()
self.encoder = nn.ModuleList([EncoderLayer(d_model, n_heads, d_ff, dropout) for _ in range(n_layers)])
self.decoder = nn.ModuleList([DecoderLayer(d_model, n_heads, d_ff, dropout) for _ in range(n_layers)])
self.src_embed = nn.Sequential(nn.Embedding(src_vocab_size, d_model), PositionalEncoding(d_model))
self.tgt_embed = nn.Sequential(nn.Embedding(tgt_vocab_size, d_model), PositionalEncoding(d_model))
self.out = nn.Linear(d_model, tgt_vocab_size)
def forward(self, src, tgt, src_mask=None, tgt_mask=None):
enc_output = self.src_embed(src)
for layer in self.encoder:
enc_output = layer(enc_output, src_mask)
dec_output = self.tgt_embed(tgt)
for layer in self.decoder:
dec_output = layer(dec_output, enc_output, src_mask, tgt_mask)
output = self.out(dec_output)
return output
class MockDataset(Dataset):
def __init__(self, src_vocab_size, tgt_vocab_size, num_samples, max_len):
self.src_vocab_size = src_vocab_size
self.tgt_vocab_size = tgt_vocab_size
self.num_samples = num_samples
self.max_len = max_len
self.data = self.generate_data()
def generate_data(self):
data = []
for _ in range(self.num_samples):
src_len = random.randint(1, self.max_len)
tgt_len = random.randint(1, self.max_len)
src = torch.randint(1, self.src_vocab_size, (src_len,))
tgt = torch.randint(1, self.tgt_vocab_size, (tgt_len,))
data.append((src, tgt))
return data
def __len__(self):
return self.num_samples
def __getitem__(self, idx):
return self.data[idx]
def pad_sequence(sequences, pad_value=0):
max_len = max([seq.size(0) for seq in sequences])
padded_seqs = []
for seq in sequences:
padded_seq = torch.full((max_len,), pad_value, dtype=torch.long)
padded_seq[:seq.size(0)] = seq
padded_seqs.append(padded_seq)
return torch.stack(padded_seqs)
def collate_fn(batch):
src_seqs, tgt_seqs = zip(*batch)
src_padded = pad_sequence(src_seqs)
tgt_padded = pad_sequence(tgt_seqs)
return src_padded, tgt_padded
def train(model, dataloader, criterion, optimizer, device, num_epochs):
model.train()
for epoch in range(num_epochs):
total_loss = 0
for batch, (src, tgt) in enumerate(dataloader):
src, tgt = src.to(device), tgt.to(device)
src_mask = (src != 0).unsqueeze(1).unsqueeze(2)
tgt_mask = (tgt != 0).unsqueeze(1).unsqueeze(2) & torch.triu(torch.ones((1, tgt.size(1), tgt.size(1))), diagonal=1).bool().to(device)
optimizer.zero_grad()
output = model(src, tgt[:, :-1], src_mask, tgt_mask[:, :-1, :-1])
loss = criterion(output.contiguous().view(-1, tgt_vocab_size), tgt[:, 1:].contiguous().view(-1))
loss.backward()
optimizer.step()
total_loss += loss.item()
if batch % 100 == 0:
print(f"Epoch {epoch+1}, Batch {batch}, Loss: {loss.item():.4f}")
avg_loss = total_loss / len(dataloader)
print(f"Epoch {epoch+1} complete. Average Loss: {avg_loss:.4f}")
# Hyperparameterssrc_vocab_size = 1000
tgt_vocab_size = 1000
d_model = 512
n_heads = 8
d_ff = 2048
n_layers = 6
dropout = 0.1
batch_size = 32
num_epochs = 5
learning_rate = 0.0001
# Create mock dataset and dataloaderdataset = MockDataset(src_vocab_size, tgt_vocab_size, num_samples=10000, max_len=50)
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn)
# Initialize model, loss function, and optimizerdevice = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = Transformer(src_vocab_size, tgt_vocab_size, d_model, n_heads, d_ff, n_layers, dropout).to(device)
criterion = nn.CrossEntropyLoss(ignore_index=0)
optimizer = optim.Adam(model.parameters(), lr=learning_rate, betas=(0.9, 0.98), eps=1e-9)
# Train the modeltrain(model, dataloader, criterion, optimizer, device, num_epochs)