def get_output_vector(self, qlist=None, batch_size=64):
with torch.no_grad():
for i in range(0, len(qlist), batch_size):
st = i
if i + batch_size < len(qlist):
et = i + batch_size
else:
et = len(qlist)
# features
t = get_specific_comp_list("title", qlist[st:et])
dt = get_specific_comp_list("desc_text", qlist[st:et])
dc = get_specific_comp_list("desc_code", qlist[st:et])
t = torch.tensor(t).long()
dt = torch.tensor(dt).long()
dc = torch.tensor(dc).long()
if self.args.cuda:
t, dt, dc = t.cuda(), dt.cuda(), dc.cuda()
t = self.title_embed(t.cuda() if torch.cuda.is_available() else t)
dt = self.desc_text_embed(dt.cuda() if torch.cuda.is_available() else dt)
dc = self.desc_code_embed(dc.cuda() if torch.cuda.is_available() else dc)
if self.args.static:
t = Variable(t)
dt = Variable(dt)
dc = Variable(dc)
t = t.unsqueeze(1) # (N, Ci, W, D)
dt = dt.unsqueeze(1) # (N, Ci, W, D)
dc = dc.unsqueeze(1) # (N, Ci, W, D)
t = [F.relu(conv(t)).squeeze(3) for conv in self.convs_t] # [(N, Co, W), ...]*len(Ks)
dt = [F.relu(conv(dt)).squeeze(3) for conv in self.convs_dt] # [(N, Co, W), ...]*len(Ks)
dc = [F.relu(conv(dc)).squeeze(3) for conv in self.convs_dc] # [(N, Co, W), ...]*len(Ks)
t = [F.max_pool1d(i, i.size(2)).squeeze(2) for i in t] # [(N, Co), ...]*len(Ks)
dt = [F.max_pool1d(i, i.size(2)).squeeze(2) for i in dt] # [(N, Co), ...]*len(Ks)
dc = [F.max_pool1d(i, i.size(2)).squeeze(2) for i in dc] # [(N, Co), ...]*len(Ks)
x_t, x_dt, x_dc = torch.cat(t, 1), torch.cat(dt, 1), torch.cat(dc, 1)
for j in range(et - st):
qlist[st + j].title = (Variable(x_t).data).cpu().numpy()[j]
qlist[st + j].desc_text = (Variable(x_dt).data).cpu().numpy()[j]
qlist[st + j].desc_code = (Variable(x_dc).data).cpu().numpy()[j]
return qlist
def train(train_iter, dev_iter, model, args, global_train_step):
if args.cuda:
model.cuda()
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
steps = 0
model.train()
for epoch in range(1, args.epochs + 1):
print("\n#epoch %s" % epoch, get_current_time())
for batch in train_iter:
# features
t = np.array(get_specific_comp_list("title", batch))
dt = np.array(get_specific_comp_list("desc_text", batch))
dc = np.array(get_specific_comp_list("desc_code", batch))
# label
target = get_specific_comp_list("tags", batch)
t = torch.tensor(t).long()
dt = torch.tensor(dt).long()
dc = torch.tensor(dc).long()
target = torch.tensor(target).float()
if args.cuda:
t, dt, dc, target = t.cuda(), dt.cuda(), dc.cuda(), target.cuda()
optimizer.zero_grad()
logit = model(t, dt, dc)
# debug
# print("logit.reshape(-1) shape %s"%logit.reshape(-1).shape)
# print("logit.reshape(-1).[:10] %s"%logit.reshape(-1)[:10])
# print("target.reshape(-1) shape %s"%target.reshape(-1).shape)
# print("target.reshape(-1)[:10] %s"%target.reshape(-1)[:10])
loss = nn.BCELoss()
loss = loss(logit.reshape(-1), target.reshape(-1))
loss.backward()
optimizer.step()
steps += 1
global_train_step += 1
if steps % args.log_interval == 0:
sys.stdout.write('\rBatch[{}] - loss: {:.10f}'.format(steps, loss))
if global_train_step % args.save_interval == 0:
print("\nglobal_train_step {} - step {} - loss {:.10f}".format(global_train_step, steps, loss),
get_current_time())
save(model, args.save_dir, 'snapshot', global_train_step)
return model, global_train_step
def eval_compute_f1_5(data_iter, model, args):
with torch.no_grad():
model.eval()
f1_5_dict = dict()
print("type of f1_5 %s" % type(f1_5_dict))
cnt = 0
for batch in data_iter:
id_list = [x.qid for x in batch]
# features
t = get_specific_comp_list("title", batch)
dt = get_specific_comp_list("desc_text", batch)
dc = get_specific_comp_list("desc_code", batch)
# label
target = np.array(get_specific_comp_list("tags", batch))
t = torch.tensor(t).long()
dt = torch.tensor(dt).long()
dc = torch.tensor(dc).long()
target = torch.tensor(target).float()
if args.cuda:
t, dt, dc, target = t.cuda(), dt.cuda(), dc.cuda(), target.cuda()
logit = model(t, dt, dc)
if torch.cuda.is_available():
f1_batch, cnt_batch = evaluate_batch_f1_5(pred=logit.cpu().detach().numpy(),
label=target.cpu().detach().numpy())
else:
f1_batch, cnt_batch = evaluate_batch_f1_5(pred=logit.detach().numpy(),
label=target.detach().numpy())
for i in range(len(id_list)):
f1_5_dict[id_list[i]] = f1_batch[i]
cnt += cnt_batch
print("type of f1_5 %s" % type(f1_5_dict))
return f1_5_dict, cnt
def eval(data_iter, model, args, topk_list):
with torch.no_grad():
model.eval()
pre = [0.0] * len(topk_list)
rc = [0.0] * len(topk_list)
f1 = [0.0] * len(topk_list)
cnt = 0
for batch in data_iter:
# features
t = get_specific_comp_list("title", batch)
dt = get_specific_comp_list("desc_text", batch)
dc = get_specific_comp_list("desc_code", batch)
# label
target = np.array(get_specific_comp_list("tags", batch))
t = torch.tensor(t).long()
dt = torch.tensor(dt).long()
dc = torch.tensor(dc).long()
target = torch.tensor(target).float()
if args.cuda:
t, dt, dc, target = t.cuda(), dt.cuda(), dc.cuda(), target.cuda()
logit = model(t, dt, dc)
if torch.cuda.is_available():
pre_batch, rc_batch, f1_batch, cnt_batch = evaluate_batch(pred=logit.cpu().detach().numpy(),
label=target.cpu().detach().numpy(),
topk_list=topk_list)
else:
pre_batch, rc_batch, f1_batch, cnt_batch = evaluate_batch(pred=logit.detach().numpy(),
label=target.detach().numpy(),
topk_list=topk_list)
for idx, topk in enumerate(topk_list):
pre[idx] += pre_batch[idx]
rc[idx] += rc_batch[idx]
f1[idx] += f1_batch[idx]
cnt += cnt_batch
return pre, rc, f1, cnt