def train(model, train_set, optimizer, tsp_database_path):
model.train()
sum_duplicates = 0
sum_solution = 0
random.shuffle(train_set)
for data_file in train_set:
details = data_file.split('.')[0].split('_')
nb_cities, instance_id = int(details[1]), int(details[2])
ordered_path, total_weight = read_tsp_choco_solution_file(nb_cities, instance_id, tsp_database_path)
weight_matrix = ordered_path.get_weight_matrix()
input_data = Variable(torch.tensor(normalize_weight_matrix(weight_matrix).reshape((1, nb_cities * nb_cities)),
dtype=torch.float), requires_grad=True)
optimizer.zero_grad()
output = model(input_data) # calls the forward function
target = Variable(torch.tensor(ordered_path.to_ordered_path_binary_matrix().get_candidate(), dtype=torch.float),
requires_grad=True)
candidate = OrderedPath(np.array(torch.argmax(output.detach(), dim=1), dtype=int), weight_matrix)
sum_duplicates += candidate.get_nb_duplicates()
sum_solution += int(candidate.is_solution())
loss = model.loss_function(output, target)
loss.backward()
optimizer.step()
train_set_size = len(train_set)
print('Training indicators : Solution: {}, Nb duplicates: {}'.format(sum_solution / train_set_size,
sum_duplicates / train_set_size))
return model
def valid(model, valid_set, epoch, tsp_database_path):
model.eval()
valid_loss = 0
correct = 0
sum_duplicates = 0
sum_solution = 0
random.shuffle(valid_set)
for data_file in valid_set:
details = data_file.split('.')[0].split('_')
nb_cities, instance_id = int(details[1]), int(details[2])
ordered_path, total_weight = read_tsp_choco_solution_file(nb_cities, instance_id, tsp_database_path)
weight_matrix = ordered_path.get_weight_matrix()
input_data = Variable(torch.tensor(normalize_weight_matrix(weight_matrix).reshape((1, nb_cities * nb_cities)),
dtype=torch.float))
output = model(input_data)
target = Variable(torch.tensor(ordered_path.to_ordered_path_binary_matrix().get_candidate(), dtype=torch.float),
requires_grad=True)
candidate = OrderedPath(np.array(torch.argmax(output.detach(), dim=1), dtype=int).transpose(), weight_matrix)
valid_loss += model.loss_function(output, target) # sum up batch loss
sum_duplicates += candidate.get_nb_duplicates()
sum_solution += int(candidate.is_solution())
correct += int(candidate.is_solution())
valid_set_size = len(valid_set)
sum_solution /= valid_set_size
sum_duplicates /= valid_set_size
print('Valid phase indicators : Solution: {}, Nb duplicates: {}'.format(sum_solution,
sum_duplicates))
valid_loss /= valid_set_size
print('Epoch {} - valid set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
epoch, valid_loss, correct, valid_set_size, 100. * correct / valid_set_size))
return correct / valid_set_size, valid_loss, sum_duplicates
if trained_net == 0:
# Generator training
label = [0]
g_optimizer.zero_grad()
input_g = generator(wm)
can = np.array([
np.argmax(input_g.detach().numpy()[k * 10:(k + 1) * 10])
for k in range(10)
],
dtype=int)
op_can = OrderedPath(
can,
wm.detach().numpy().astype(int).reshape(10, 10))
nb_solutions += int(op_can.is_solution())
nb_duplicates += op_can.get_nb_duplicates()
output_g = torch.tensor(label,
dtype=torch.float,
requires_grad=False)
predicted_output_g = discriminator(input_g)
g_loss = F.binary_cross_entropy(predicted_output_g, output_g)
avg_g_loss += g_loss.item()
nb_g_loss += 1
if net_switch or first_g_loss is None:
nb_iterations = 1
first_g_loss = g_loss.item()
if last_g_loss is None:
last_g_loss = first_g_loss
else: