parser.add_argument('--max_speed', type=float, default=10., help='max speed')
parser.add_argument('--max_t', type=float, default=3., help='max time')
opt = parser.parse_args()
if opt.test_mode: opt.batch_size = 1
description = 'wgan-gp mirror v0/opt.max_speed'
log_path = 'result/log/' + opt.dataset_name + '/'
os.makedirs('result/saved_models/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output2/%s' % opt.dataset_name, exist_ok=True)
if not opt.test_mode:
logger = SummaryWriter(log_dir=log_path)
write_params(log_path, parser, description)
generator = Generator(opt.vector_dim + 2).to(device)
discriminator = Discriminator(opt.points_num * 2 + 1).to(device)
#discriminator.load_state_dict(torch.load('result/saved_models/wgan-gp-10/discriminator_40000.pth'))
#generator.load_state_dict(torch.load('result/saved_models/wgan-gp-10/generator_40000.pth'))
start_point_criterion = torch.nn.MSELoss()
criterion = torch.nn.BCELoss() #.to(device)
trajectory_criterion = torch.nn.MSELoss()
g_optimizer = torch.optim.RMSprop(generator.parameters(),
lr=opt.lr,
weight_decay=opt.weight_decay)
#g_optimizer = torch.optim.Adam(generator.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
d_optimizer = torch.optim.RMSprop(discriminator.parameters(),
lr=opt.lr,
weight_decay=opt.weight_decay)
#d_optimizer = torch.optim.Adam(discriminator.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
description = 'dropout'
log_path = 'result/log/' + opt.dataset_name + '/'
os.makedirs('result/saved_models/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output2/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output3/%s' % opt.dataset_name, exist_ok=True)
if not opt.test_mode:
logger = SummaryWriter(log_dir=log_path)
write_params(log_path, parser, description)
# generator = Generator(input_dim=128+32+1+1, output=2).to(device)
# discriminator = Discriminator(opt.points_num*2+32+1).to(device)
generator = Generator(input_dim=2 + 2 + 1 + 1, output=2).to(device)
discriminator = Discriminator(opt.points_num * 2 + 2 + 1).to(device)
# encoder = CNN(input_dim=1, out_dim=32).to(device)
encoder = CNNNorm(input_dim=1, out_dim=2).to(device)
encoder.load_state_dict(
torch.load('result/saved_models/il-uncertainty-02/encoder_119000.pth'))
# DO NOT TRAIN ENCODER
encoder.eval()
# discriminator.load_state_dict(torch.load('result/saved_models/train-gan-costmap-01/discriminator_120000.pth'))
generator.load_state_dict(
torch.load(
'result/saved_models/train-gan-costmap-01/generator_120000.pth'))
start_point_criterion = torch.nn.MSELoss()
criterion = torch.nn.BCELoss() #.to(device)
trajectory_criterion = torch.nn.MSELoss()
# e_optimizer = torch.optim.RMSprop(encoder.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
parser.add_argument('--max_t', type=float, default=3., help='max time')
opt = parser.parse_args()
if opt.test_mode: opt.batch_size = 1
description = 'wgan-gp mirror v0/opt.max_speed'
log_path = 'result/log/'+opt.dataset_name+'/'
os.makedirs('result/saved_models/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output2/%s' % opt.dataset_name, exist_ok=True)
if not opt.test_mode:
logger = SummaryWriter(log_dir=log_path)
write_params(log_path, parser, description)
generator = Generatorv2(opt.vector_dim+2).to(device)
discriminator = Discriminator(opt.points_num*2+1).to(device)
generator.load_state_dict(torch.load('result/saved_models/pretrain-gan-01/generator_50000.pth'))
discriminator.load_state_dict(torch.load('result/saved_models/pretrain-gan-01/discriminator_50000.pth'))
#discriminator.load_state_dict(torch.load('result/saved_models/wgan-gp-10/discriminator_40000.pth'))
#generator.load_state_dict(torch.load('result/saved_models/wgan-gp-10/generator_40000.pth'))
start_point_criterion = torch.nn.MSELoss()
criterion = torch.nn.BCELoss()#.to(device)
trajectory_criterion = torch.nn.MSELoss()
g_optimizer = torch.optim.RMSprop(generator.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
#g_optimizer = torch.optim.Adam(generator.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
d_optimizer = torch.optim.RMSprop(discriminator.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
#d_optimizer = torch.optim.Adam(discriminator.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
# train_loader = DataLoader(CostMapDataset(data_index=[item for item in range(1,10)], opt=opt, dataset_path='/media/wang/DATASET/CARLA_HUMAN/town01/'), batch_size=opt.batch_size, shuffle=False, num_workers=opt.n_cpu)
# discriminator = Discriminator(opt.points_num*2+1).to(device)
# encoder = Encoder(input_dim=3, out_dim=opt.vector_dim).to(device)
encoder = MobileNetV2(num_classes=opt.vector_dim, in_channels=3).to(device)
# generator.load_state_dict(torch.load('result/saved_models/train-gan-costmap-03/generator_10000.pth'))
generator.load_state_dict(torch.load('result/saved_models/pretrain-gan-01/generator_50000.pth'))
trajectory_criterion = torch.nn.MSELoss().to(device)
latent_criterion = torch.nn.MSELoss().to(device)
e_optimizer = torch.optim.Adam(encoder.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
discriminator = Discriminator(input_dim=opt.vector_dim*2, output=1).to(device)
discriminator_criterion = nn.BCEWithLogitsLoss().to(device)
d_optimizer = torch.optim.Adam(discriminator.parameters(), lr=3e-4, weight_decay=opt.weight_decay)
param = parse_yaml_file_unsafe('./param_oxford.yaml')
train_loader = DataLoader(DIVADataset(param, mode='train', opt=opt), batch_size=opt.batch_size, shuffle=False, num_workers=opt.n_cpu)
train_samples = iter(train_loader)
param = parse_yaml_file_unsafe('./param_kitti.yaml')
eval_loader = DataLoader(KittiDataset(param, mode='eval', opt=opt), batch_size=1, shuffle=False, num_workers=1)
eval_samples = iter(eval_loader)
def show_traj(fake_traj, real_traj, t, step, img=None):
fake_xy = fake_traj
x = fake_xy[:,0]#*opt.max_dist
description = 'dropout'
log_path = 'result/log/' + opt.dataset_name + '/'
os.makedirs('result/saved_models/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output2/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output3/%s' % opt.dataset_name, exist_ok=True)
if not opt.test_mode:
logger = SummaryWriter(log_dir=log_path)
write_params(log_path, parser, description)
# generator = Generator(input_dim=128+32+1+1, output=2).to(device)
# discriminator = Discriminator(opt.points_num*2+32+1).to(device)
generator = Generator(input_dim=2 + 2 + 1 + 1, output=2).to(device)
discriminator = Discriminator(opt.points_num * 2 + 2 + 1).to(device)
# encoder = CNN(input_dim=1, out_dim=32).to(device)
encoder = CNNNorm(input_dim=1, out_dim=2).to(device)
encoder.load_state_dict(
torch.load('result/saved_models/il-uncertainty-02/encoder_119000.pth'))
# DO NOT TRAIN ENCODER
encoder.eval()
generator.eval()
discriminator.eval()
# discriminator.load_state_dict(torch.load('result/saved_models/train-gan-costmap-01/discriminator_120000.pth'))
# generator.load_state_dict(torch.load('result/saved_models/train-gan-costmap-01/generator_120000.pth'))
discriminator.load_state_dict(
torch.load(
'result/saved_models/train-gan-costmap-dropout-01/discriminator_35000.pth'
os.makedirs('result/saved_models/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output/%s' % opt.dataset_name, exist_ok=True)
if not opt.test_mode:
logger = SummaryWriter(log_dir=log_path)
write_params(log_path, parser, description)
# generator = Generatorv2(opt.vector_dim+2).to(device)
generator = Generator(input_dim=1 + 1 + opt.vector_dim, output=2).to(device)
# discriminator = Discriminator(opt.points_num*2+1).to(device)
# encoder = Encoder(input_dim=6, out_dim=opt.vector_dim).to(device)
encoder = MobileNetV2(num_classes=opt.vector_dim, in_channels=6).to(device)
# encoder = RNNEncoder(input_dim=3, out_dim=opt.vector_dim).to(device)
similarity_net = Discriminator(input_dim=4 * opt.points_num + 2).to(device)
generator.load_state_dict(
torch.load('result/saved_models/train-gan-costmap-03/generator_10000.pth'))
# generator.load_state_dict(torch.load('result/saved_models/train-gan-costmap-03/generator_10000.pth'))
# generator.load_state_dict(torch.load('result/saved_models/pretrain-gan-01/generator_50000.pth'))
# encoder.load_state_dict(torch.load('result/saved_models/train-GAN-gen-03/encoder_0.pth'))
# encoder.load_state_dict(torch.load('result/saved_models/train-GAN-gen-06/encoder_160000.pth'))
# encoder.load_state_dict(torch.load('result/saved_models/train-speed-divide-01/encoder_16000.pth'))
# similarity_net.load_state_dict(torch.load('result/saved_models/train-speed-divide-01/similarity_net_16000.pth'))
trajectory_criterion = torch.nn.MSELoss().to(device)
latent_criterion = torch.nn.MSELoss().to(device)
mse_criterion = torch.nn.MSELoss().to(device)
e_optimizer = torch.optim.SGD(encoder.parameters(),
lr=opt.lr,
opt = parser.parse_args()
if opt.test_mode: opt.batch_size = 1
description = 'dropout'
log_path = 'result/log/'+opt.dataset_name+'/'
os.makedirs('result/saved_models/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output2/%s' % opt.dataset_name, exist_ok=True)
os.makedirs('result/output3/%s' % opt.dataset_name, exist_ok=True)
if not opt.test_mode:
logger = SummaryWriter(log_dir=log_path)
write_params(log_path, parser, description)
generator = Generator(input_dim=1+1+opt.vector_dim, output=2).to(device)
discriminator = Discriminator(opt.points_num*2+1).to(device)
# generator.load_state_dict(torch.load('result/saved_models/train-gan-costmap-03/generator_10000.pth'))
# discriminator.load_state_dict(torch.load('result/saved_models/pretrain-gan-01/discriminator_50000.pth'))
# generator.load_state_dict(torch.load('result/saved_models/train-gan-costmap-vector2-01/generator_5000.pth'))
# discriminator.load_state_dict(torch.load('result/saved_models/train-gan-costmap-02/discriminator_46000.pth'))
# generator.load_state_dict(torch.load('result/saved_models/train-gan-costmap-02/generator_46000.pth'))
# discriminator.load_state_dict(torch.load('result/saved_models/train-gan-costmap-dropout-01/discriminator_35000.pth'))
# generator.load_state_dict(torch.load('result/saved_models/train-gan-costmap-dropout-01/generator_35000.pth'))
start_point_criterion = torch.nn.MSELoss()
criterion = torch.nn.BCELoss()#.to(device)
trajectory_criterion = torch.nn.MSELoss()
# e_optimizer = torch.optim.RMSprop(encoder.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
g_optimizer = torch.optim.RMSprop(generator.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)