ans_offset[:,
0]).cuda() if use_cuda else Variable(ans_offset[:,
0])
end_ans = Variable(ans_offset[:,
1]).cuda() if use_cuda else Variable(
ans_offset[:, 1])
appear = Variable(appear).cuda() if use_cuda else Variable(appear)
start, end, start_attn, end_attn = fusion_net(context, q, appear)
loss = criterion(start_attn, start_ans) + criterion(
end_attn, end_ans)
loss.backward()
nn.utils.clip_grad_norm(fusion_net.parameters(), 10)
optimizer.step()
start, end, scores = decode(start.data.cpu(), end.data.cpu(), 1)
f1_score, exact_match_score = batch_score(start, end, ans_offset)
if batch % args.display_freq == 0:
print(
'epoch: %d | batch: %d/%d| loss: %f | f1: %f | exact: %f' %
(epoch, batch, len(train_engine), loss.data[0], f1_score,
exact_match_score))
batch += 1
valid_f1, valid_exact = 0, 0
fusion_net.eval()
for context, q, ans_offset, appear in valid_engine:
context = Variable(context).cuda() if use_cuda else Variable(
context)
q = Variable(q).cuda() if use_cuda else Variable(q)
start_ans = Variable(
fusion_net = torch.load('model.cpt')
if use_cuda:
fusion_net = fusion_net.cuda()
fusion_net.eval()
valid_f1, valid_exact = 0, 0
for context, q, ans_offset, appear in valid_engine:
context = Variable(context).cuda() if use_cuda else Variable(context)
q = Variable(q).cuda() if use_cuda else Variable(q)
start_ans = Variable(ans_offset[:, 0]).cuda() if use_cuda else Variable(ans_offset[:, 0])
end_ans = Variable(ans_offset[:, 1]).cuda() if use_cuda else Variable(ans_offset[:, 1])
appear = Variable(appear).cuda() if use_cuda else Variable(appear)
start, end, start_attn, end_attn = fusion_net(context, q, appear)
start, end, scores = decode(start.data.cpu(), end.data.cpu(), 1, 20)
f1_score, exact_match_score = batch_score(start, end, ans_offset)
valid_f1 += f1_score
valid_exact += exact_match_score
print('valid_f1: %f | valid_exact: %f'%(
valid_f1/len(valid_engine), valid_exact/len(valid_engine)
))
f = open('predict.csv', 'w')
f.write('id,answer\n')
for i in tqdm(range(len(test_engine))):
context, q, ans_offset, appear = test_engine[i]
context = Variable(context).cuda() if use_cuda else Variable(context)
q = Variable(q).cuda() if use_cuda else Variable(q)
appear = Variable(appear).cuda() if use_cuda else Variable(appear)
start, end, start_attn, end_attn = fusion_net(context.unsqueeze(0), q.unsqueeze(0), appear.unsqueeze(0))