def train_skipgram(vocab, sg_loader):
losses = []
loss_fn = nn.L1Loss()
model = SkipGram(len(vocab),
file=embeddings_bin) if load_emb else SkipGram(
len(vocab), embed_size, simple)
print(model)
optimizer = optim.SGD(model.parameters(), lr=learning_rate)
for epoch in range(n_epoch):
total_loss = 0.0
for i, sample_batched in enumerate(sg_loader):
sample_batched = sample_batched[0]
model.zero_grad()
log_probs = model(sample_batched[:, :-1], average)
loss = loss_fn(log_probs, Variable(sample_batched[:, -1].float()))
loss.backward()
optimizer.step()
total_loss += loss.data
losses.append(total_loss.item())
print('Epoch:', epoch, 'Loss:', total_loss.item())
save_params(emb=model, losses=losses, vocab=vocab)
# Early stopping
if len(losses) > 2 and losses[-1] > losses[-2]:
break
return model, losses
def train_skipgram(vocab, sg_loader):
losses = []
loss_fn = nn.L1Loss()
model = SkipGram(len(vocab), embed_size, simple)
print(model)
if load_prev:
try:
model.load_state_dict(torch.load(model_file))
except:
print('Could not load file')
optimizer = optim.SGD(model.parameters(), lr=learning_rate)
for epoch in range(n_epoch):
total_loss = 0.0
for i, sample_batched in enumerate(sg_loader):
sample_batched = sample_batched[0]
in_w_var = Variable(sample_batched[:, 0])
ctx_w_var = Variable(sample_batched[:, 1])
# print(in_w_var.shape)
model.zero_grad()
log_probs = model(in_w_var, ctx_w_var)
loss = loss_fn(log_probs, Variable(sample_batched[:, 2].float()))
loss.backward()
optimizer.step()
total_loss += loss.data
losses.append(total_loss.item())
print('Epoch:', epoch, 'Loss:', total_loss.item())
save_params(vocab, model, losses)
return model, losses
