I have implemented Transformer model based on nn.Transformer
Unfortunately, my model is not learning. I have LSTM based networks that show good learning on the same sequence to sequence dataset. Can you please suggest what can be wrong?
class TransformerBase(nn.Module):
def __init__(self, input_size, output_size, hidden_size=256):
super(TransformerBase, self).__init__()
self.input_size = input_size
self.output_size = output_size
self.hidden_size = hidden_size
self.transformer_model = nn.Transformer(d_model=hidden_size)
self.embedding_input = nn.Embedding(self.input_size, hidden_size)
self.embedding_output = nn.Embedding(self.output_size, hidden_size)
self.pos_encoder = PositionalEncoding(hidden_size, 0.1)
self.fc_out = nn.Linear(hidden_size, self.output_size)
def forward(self, src, trg):
embedded_input = self.pos_encoder(self.embedding_input(src) * math.sqrt(self.hidden_size))
embedding_output = self.pos_encoder(self.embedding_output(trg) * math.sqrt(self.hidden_size))
tgt_mask = generate_square_subsequent_mask(trg.shape[0])
x = self.transformer_model(src=embedded_input, tgt=embedding_output, tgt_mask=tgt_mask.to(device))
return self.fc_out(x)
def generate_square_subsequent_mask(sz: int):
mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0))
return mask
# https://pytorch.org/tutorials/beginner/transformer_tutorial.html
class PositionalEncoding(nn.Module):
def __init__(self, d_model, dropout=0.1, max_len=5000):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(p=dropout)
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 self.dropout(x)
def train(model, iterator, optimizer, criterion, clip):
model.train()
epoch_loss = 0
for i, batch in enumerate(iterator):
# Get input and targets and get to cuda
src = batch.src.to(device)
trg = batch.trg.to(device)
optimizer.zero_grad()
output = model(src, trg[:-1])
output = output.reshape(-1, output.shape[-1]).contiguous()
trg = trg[1:].reshape(-1).contiguous()
loss = criterion(output, trg)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), clip)
optimizer.step()
epoch_loss += loss.item()
return epoch_loss / len(iterator)
criterion = nn.CrossEntropyLoss(ignore_index=pad_idx)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)