def eval(opt):
img_logs, weight_logs, tensor_writer = init_logs(opt)
model = CycleGANModel(opt)
model.parallel_init(device_ids=opt.device_ids)
weight_logs = weight_logs.replace('weights', 'weights_bak3')
print(weight_logs)
model.load(weight_logs)
# model.netG_B.eval()
model.img_policy.online_test(model.netG_B, 5, img_logs)
def eval(opt):
img_logs, weight_logs, tensor_writer = init_logs(opt)
model = CycleGANModel(opt)
model.parallel_init(device_ids=opt.device_ids)
print(weight_logs)
dataset = Robotdata.get_loader(opt)
# fdataset = RobotStackFdata.get_loader(opt)
# model.train_forward_state(fdataset, True)
weight_list = list(
filter(lambda x: 'weights_' in x,
os.listdir(weight_logs.replace('weights', ''))))
weight_list = sorted(weight_list, key=lambda x: int(x.split('_')[1]))
for weight_path in weight_list:
weight_id = int(weight_path.split('_')[1])
weight_path = weight_logs.replace('weights', weight_path)
print(weight_path)
model.load(weight_path)
# model.img_policy.online_test(model.netG_B,5)
dataset = Robotdata.get_loader(opt)
ave_loss = {}
count_n = 10
gt, pred = [], []
for batch_id, data in enumerate(dataset):
model.set_input(data)
model.test()
errors = model.get_current_errors()
display = '===> Batch({}/{})'.format(batch_id, len(dataset))
for key, value in errors.items():
display += '{}:{:.4f} '.format(key, value)
try:
ave_loss[key] = ave_loss[key] + value
except:
ave_loss[key] = value
gt.append(model.gt0.cpu().data.numpy())
pred.append(model.fake_At0.cpu().data.numpy())
if batch_id >= count_n - 1:
path = os.path.join(img_logs, 'imgA_{}.jpg'.format(weight_id))
model.visual(path)
break
gt = np.vstack(gt)
pred = np.vstack(pred)
np.save(os.path.join(img_logs, 'gt_{}.npy'.format(weight_id)), gt)
np.save(os.path.join(img_logs, 'pred_{}.npy'.format(weight_id)), pred)
display = 'average loss: '
for key, value in ave_loss.items():
display += '{}:{:.4f} '.format(key, value / (count_n))
print(display)
def train(opt):
img_logs, weight_logs, tensor_writer = init_logs(opt)
model = CycleGANModel(opt)
dataset = Robotdata.get_loader(opt)
fdataset = RobotStackFdata.get_loader(opt)
model.train_forward_state(fdataset, opt.pretrain_f)
model.parallel_init(device_ids=opt.device_ids)
model.img_policy.online_test(model.netG_B, 1)
best_reward = 0
niter = 0
for epoch_id in range(opt.epoch_size):
for batch_id, data in enumerate(dataset):
model.set_input(data)
model.optimize_parameters()
niter += 1
if (batch_id) % opt.display_gap == 0:
errors = model.get_current_errors()
display = '===> Epoch[{}]({}/{})'.format(
epoch_id, batch_id, len(dataset))
for key, value in errors.items():
display += '{}:{:.4f} '.format(key, value)
tensor_writer.add_scalar(key, value, niter)
print(display)
if (batch_id) % opt.save_weight_gap == 0:
path = os.path.join(
img_logs, 'imgA_{}_{}.jpg'.format(epoch_id, batch_id + 1))
model.visual(path)
reward = model.img_policy.online_test(model.netG_B, 1)
reward = reward.mean()
if reward > best_reward:
best_reward = reward
model.save(weight_logs)
print('best_reward:{} cur_reward:{}'.format(
best_reward, reward))
def eval(opt):
img_logs, weight_logs, tensor_writer = init_logs(opt)
model = CycleGANModel(opt)
model.parallel_init(device_ids=opt.device_ids)
print(weight_logs)
model.load(weight_logs)
# model.img_policy.online_test(model.netG_B,5)
dataset = Robotdata.get_loader(opt)
ave_loss = {}
for batch_id, data in enumerate(dataset):
model.set_input(data)
model.test()
errors = model.get_current_errors()
display = '===> Batch({}/{})'.format(batch_id, len(dataset))
for key, value in errors.items():
display += '{}:{:.4f} '.format(key, value)
try:
ave_loss[key] = ave_loss[key] + value
except:
ave_loss[key] = value
print(display)
if (batch_id + 1) % opt.display_gap == 0:
path = os.path.join(img_logs, 'imgA_{}.jpg'.format(batch_id + 1))
model.visual(path)