def evaluate_test(select_net_e, rank_net_e):
list_gen = ListGenerator(rel_file=Letor07Path +
'/relation.test.fold%d.txt' %
(letor_config['fold']),
config=letor_config)
map_v = 0.0
map_c = 0.0
with torch.no_grad():
for X1, X1_len, X1_id, X2, X2_len, X2_id, Y, F in \
list_gen.get_batch(data1=loader.query_data, data2=loader.doc_data):
#print_log(X1.shape, X2.shape, Y.shape)
X1, X1_len, X2, X2_len, Y, F = to_device(X1,
X1_len,
X2,
X2_len,
Y,
F,
device=select_device)
X1, X2, X1_len, X2_len, X2_pos = select_net_e(
X1, X2, X1_len, X2_len, X1_id, X2_id)
X2, X2_len = utils.data_adaptor(X2, X2_len, select_net, rank_net,
letor_config)
#print_log(X1.shape, X2.shape, Y.shape)
pred = rank_net_e(X1, X2, X1_len, X2_len, X2_pos)
map_o = utils.eval_MAP(pred.tolist(), Y.tolist())
#print_log(pred.shape, Y.shape)
map_v += map_o
map_c += 1.0
map_v /= map_c
print_log('[Test]', map_v)
return map_v
rank_loss_list = []
start_t = time.time()
for i in range(1000):
# One Step Forward
X1, X1_len, X1_id, X2, X2_len, X2_id, Y, F = pair_gen.get_batch(
data1=loader.query_data, data2=loader.doc_data)
X1, X1_len, X2, X2_len, Y, F = to_device(X1,
X1_len,
X2,
X2_len,
Y,
F,
device=select_device)
X1, X2, X1_len, X2_len, X2_pos = select_net(X1, X2, X1_len, X2_len, X1_id,
X2_id)
X2, X2_len = utils.data_adaptor(X2, X2_len, select_net, rank_net,
letor_config)
output = rank_net(X1, X2, X1_len, X2_len, X2_pos)
# Update Rank Net
rank_loss = rank_net.pair_loss(output, Y)
print('rank loss:', rank_loss.item())
rank_loss_list.append(rank_loss.item())
rank_optimizer.zero_grad()
rank_loss.backward()
rank_optimizer.step()
end_t = time.time()
print('Time Cost: %s s' % (end_t - start_t))
torch.save(select_net, "qcentric.model")
torch.save(rank_net, "deeprank.model")