def runPytorchModel(args):
# Initialize network
if args['model'] == "traphic":
net = traphicNet(args)
else:
net = highwayNet(args)
# net.load_state_dict(torch.load('trained_models/m_false/cslstm_b_pretrain2_NGSIM.tar'), strict=False)
if args['cuda']:
print("Using cuda")
net = net.cuda()
# net = net.to("cuda")
## Initialize optimizer
optim = torch.optim.Adam(net.parameters(), lr=args['learning_rate'])
crossEnt = torch.nn.BCELoss()
if verbose:
print("*" * 3, "Using model: ", net)
print("*" * 3, "Optim: ", optim)
print("*" * 3, "Creating dataset and dataloaders...")
## Initialize data loaders
trSet = ngsimDataset('model/Prediction/data/TRAF/TRAFTrainSet.npy')
valSet = ngsimDataset('model/Prediction/data/TRAF/TRAFValSet.npy')
# trSet = ngsimDataset('./Prediction/data/TrainSetTRAF.mat')
# valSet = ngsimDataset('./Prediction/data/ValSetTRAF.mat')
trDataloader = DataLoader(trSet,
batch_size=args['batch_size'],
shuffle=True,
num_workers=8,
collate_fn=trSet.collate_fn)
valDataloader = DataLoader(valSet,
batch_size=args['batch_size'],
shuffle=True,
num_workers=8,
collate_fn=valSet.collate_fn)
if verbose:
print("starting training...")
if args['model'] == "traphic":
if verbose:
print("Training TRAPHIC")
traphic = TraphicEngine(net, optim, trDataloader, valDataloader, args)
traphic.start()
else:
if verbose:
print("Training conv social pooling")
social = SocialEngine(net, optim, trDataloader, valDataloader, args)
social.start()
def evalPytorchModel(args):
# Initialize network
args['eval'] = True
if args['model'] == "traphic":
net = traphicNet(args)
else:
net = highwayNet(args)
# net.load_state_dict(torch.load('trained_models/m_false/cslstm_b_pretrain2_NGSIM.tar'), strict=False)
if args['use_cuda']:
print("Using cuda")
net = net.cuda()
# net = net.to("cuda")
## Initialize optimizer
optim = torch.optim.Adam(net.parameters(), lr=args['learning_rate'])
crossEnt = torch.nn.BCELoss()
if verbose:
print("*" * 3, "Using model: ", net)
print("*" * 3, "Optim: ", optim)
print("*" * 3, "Creating dataset and dataloaders...")
## Initialize data loaders
testSet = ngsimDataset('model/Prediction/data/TRAF/TRAFTestSet.npy')
testDataloader = DataLoader(testSet,
batch_size=args['batch_size'],
shuffle=True,
num_workers=8,
collate_fn=testSet.collate_fn)
if verbose:
print("evaluating...")
net.load_state_dict(
torch.load("model/Prediction/trained_models/{}_{}.tar".format(
args['model'], args['name'])))
if args['model'] == "traphic":
if verbose:
print("Evaluating TRAPHIC")
traphic = TraphicEngine(net, optim, None, testDataloader, args)
traphic.start()
else:
if verbose:
print("Evaluating SOCIAL")
social = SocialEngine(net, optim, None, testDataloader, args)
social.start()
start_time = datetime.datetime.now()
# Initialize network
net = highwayNet(args)
if args['use_cuda']:
net = net.cuda()
## Initialize optimizer
trainEpochs = 10
optimizer = torch.optim.Adam(net.parameters()) #lr = ...
batch_size = 128
crossEnt = torch.nn.BCELoss() # binary cross entropy
## Initialize data loaders
trSet = ngsimDataset('../../data/trajectory/TrainSet.mat')
valSet = ngsimDataset('../../data/trajectory/ValSet.mat')
trDataloader = DataLoader(trSet,
batch_size=batch_size,
shuffle=True,
num_workers=8,
collate_fn=trSet.collate_fn)
valDataloader = DataLoader(valSet,
batch_size=batch_size,
shuffle=True,
num_workers=8,
collate_fn=valSet.collate_fn)
## Variables holding train and validation loss values:
train_loss = []
val_loss = []
batch_size = 1024 # 1024 on AWS with V100, 768 on GTX 1080 TI
else:
batch_size = 128
## Initialize data loaders
logging.info("Loading the datasets...")
newFeats = 0
behavFeats = 0
if 'X' in cmd_args.experiment: # new features experiments
newFeats = 3
behavFeats = 0
#trSet = ngsimDataset('data/TrainSetX.mat', grid_size=(19,3) )
#valSet = ngsimDataset('data/ValSetX.mat', grid_size=(19,3) )
trSet = ngsimDataset('data/TrainSetCVA.mat', newFeats=newFeats)
valSet = ngsimDataset('data/ValSetCVA.mat', newFeats=newFeats)
else:
trSet = ngsimDataset('data/TrainSet.mat')
valSet = ngsimDataset('data/ValSet.mat')
trDataloader = DataLoader(trSet,
batch_size=batch_size,
shuffle=True,
num_workers=8,
collate_fn=trSet.collate_fn)
valDataloader = DataLoader(valSet,
batch_size=batch_size,
shuffle=True,
num_workers=8,
collate_fn=valSet.collate_fn)
args['device'] = torch.device('cuda:' + str(args['gpu']))
# Evaluation metric:
# metric = 'nll' #or rmse
metric = 'rmse'
# Initialize network
net = SocialSTGCNModel(args)
net.load_state_dict(torch.load('../saved_models/student_stgcn.tar'))
state = torch.load('../saved_models/teacher_gmm_set.tar')
gmm_dict = state['gmm_dict']
tsSet = None
if args['dset_name'] == 'ngsim':
tsSet = ngsimDataset('../../data/TestSetNGSIM.mat', gmm_dict, args['Mq'],
args['z_size'])
elif args['dset_name'] == 'gta':
tsSet = gtaDataset('../../data/TestSetGTA', args['dset_tag'], gmm_dict,
args['Mq'], args['z_size'])
assert tsSet != None
tsDataloader = DataLoader(tsSet,
batch_size=64,
shuffle=True,
num_workers=8,
collate_fn=tsSet.collate_fn)
lossVals = torch.zeros(25).cuda()
counts = torch.zeros(25).cuda()
for i, data in enumerate(tsDataloader):
# Initialize network
net_vae = GCNVAE(args)
net_gmm = GMM(args)
## Initialize optimizer
pretrainEpochs = 0
trainEpochs = 1
optim_vae = torch.optim.Adam(net_vae.parameters(), lr=args['lr_vae'])
optim_gmm = torch.optim.Adam(net_gmm.parameters(), lr=args['lr_gmm'])
batch_size = 32
## Initialize data loaders
trSet = None
if args['dset_name'] == 'ngsim':
trSet = ngsimDataset('../../data/TeacherSetNGSIM.mat')
elif args['dset_name'] == 'gta':
trSet = gtaDataset('../../data/TeacherSetGTA', args['dset_tag'])
assert trSet != None
trDataloader = DataLoader(trSet,
batch_size=batch_size,
shuffle=True,
num_workers=2,
collate_fn=trSet.collate_fn)
## Variables holding train and validation loss values:
vae_losses = []
gmm_losses = []
for epoch_num in range(pretrainEpochs + trainEpochs):
# params.requires_grad = False
if args['use_cuda']:
net = net.cuda()
## Initialize optimizer
# pretrainEpochs = 80
# trainEpochs = 200
# batch_size = 128
optimizer = torch.optim.Adam(net.parameters(),lr=lr)
crossEnt = torch.nn.BCELoss()
## Initialize data loaders
trSet = ngsimDataset('data/TrainSet.mat')
valSet = ngsimDataset('data/ValSet.mat')
trDataloader = DataLoader(trSet,batch_size=batch_size,shuffle=True,num_workers=8,collate_fn=trSet.collate_fn)
valDataloader = DataLoader(valSet,batch_size=batch_size,shuffle=True,num_workers=8,collate_fn=valSet.collate_fn)
## Variables holding train and validation loss values:
train_loss = []
val_loss = []
prev_val_loss = math.inf
for epoch_num in range(pretrainEpochs+trainEpochs):
if epoch_num == 0:
print('Pre-training with MSE loss')
elif epoch_num == pretrainEpochs:
print('Training with NLL loss')
args['conv_3x1_depth'] = 16
args['dyn_embedding_size'] = 32
args['input_embedding_size'] = 32
args['num_lat_classes'] = 3
args['num_lon_classes'] = 2
args['use_maneuvers'] = True # visualize multi-model predictions
args['train_flag'] = False
# Initialize network
net = highwayNet(args)
net.load_state_dict(torch.load('trained_models/cslstm_m.tar'))
#net.load_state_dict(torch.load('trained_models/cslstm.tar'))
if args['use_cuda']:
net = net.cuda()
data_set = ngsimDataset('data/TestSet.mat')
skip = 10
lossVals = torch.zeros(25).cuda()
counts = torch.zeros(25).cuda()
delta = 2.0
x_min = -120
x_max = 120
range_y = 20
scale = 0.4
v_w = 5
v_l = 15
else:
metricFCT = maskedBIGERRTest
# Initialize network
batch_size = 1024
batch_size = 16
#batch_size=6
logging.info("Loading the datasets...")
newFeats = 0
behavFeats = 0
if 'X' in cmd_args.experiment:
newFeats = 3
behavFeats = 0
#tsSet = ngsimDataset('data/TestSetV.mat', newFeats=newFeats)
tsSet = ngsimDataset('data/TestSetCVA.mat', newFeats=newFeats)
else:
tsSet = ngsimDataset('data/TestSet.mat')
tsDataloader = DataLoader(tsSet,
batch_size=batch_size,
shuffle=False,
num_workers=0,
collate_fn=tsSet.collate_fn)
net = highwayNet(params, newFeats=newFeats, behavFeats=behavFeats)
net_path = os.path.join(args['model_dir'], 'best.pth.tar')
assert os.path.isfile(net_path), "No net file found at {}".format(net_path)
#net.load_state_dict(torch.load('trained_models/cslstm_m.tar'))
utils.load_checkpoint(net_path, net)
args['model_path'] = 'trained_models/m_false/cslstm_b_pretrain2_NGSIM.tar'
args['dataset_path'] = 'data/TestSetTRAF.mat'
args['ours'] = True
# Evaluation metric:
metric = 'rmse' #nll or rmse
# Initialize network
net = highwayNet(args)
batch_size = 128
net.load_state_dict(torch.load(args['model_path']),strict=False)
if args['use_cuda']:
net = net.cuda()
tsSet = ngsimDataset(args['dataset_path'])
tsDataloader = DataLoader(tsSet,batch_size=batch_size,shuffle=True,num_workers=8,collate_fn=tsSet.collate_fn)
lossVals = torch.zeros(args['out_length']).cuda()
counts = torch.zeros(args['out_length']).cuda()
for i, data in enumerate(tsDataloader):
st_time = time.time()
hist, upp_nbrs, nbrs, upp_mask, mask, lat_enc, lon_enc, fut, op_mask = data
# Initialize Variables
if args['use_cuda']:
hist = hist.cuda()
nbrs = nbrs.cuda()
upp_nbrs = upp_nbrs.cuda()
args['input_embedding_size'] = 32
args['train_flag'] = False
# Evaluation metric:
metric = 'rmse'
# Initialize network
net = highwayNet(args)
net.load_state_dict(torch.load('trained_models/sta_lstm_10272020.tar'))
if args['use_cuda']:
net = net.cuda()
tsSet = ngsimDataset('../../data/trajectory/TestSet.mat')
tsDataloader = DataLoader(tsSet,
batch_size=128,
shuffle=True,
num_workers=8,
collate_fn=tsSet.collate_fn) #
lossVals = torch.zeros(5).cuda()
counts = torch.zeros(5).cuda()
lossVal = 0 # revised by Lei
count = 0
vehid = []
pred_x = []
pred_y = []
T = []
assert args['h_dim2'] == args['z_dim']
args['device'] = torch.device('cpu')
if args['use_cuda']:
args['device'] = torch.device('cuda:' + str(args['gpu']))
# Initialize network
net_vae = GCNVAE(args)
net_vae.load_state_dict(torch.load('../saved_models/teacher_vae.tar'))
net_gmm = GMM(args)
net_gmm.load_state_dict(torch.load('../saved_models/teacher_gmm.tar'))
## Initialize data loaders
batch_size = 32
trSet = ngsimDataset('../../data/TrainSet.mat')
trDataloader = DataLoader(trSet,
batch_size=batch_size,
shuffle=False,
num_workers=2,
collate_fn=trSet.collate_fn)
## Infer mu_q, sigma_q and dump the dict
gmm_count = 0
gmm_dict = {}
avg_time = 0
net_vae.training = False
for i, data in enumerate(trDataloader):
st_time = time.time()
data_list, graph_list = data
# Initialize network
net = highwayNet(args)
if args['use_cuda']:
print('cuda success!')
net = net.cuda()
## Initialize optimizer
trainEpochs = 10
optimizer = torch.optim.Adam(net.parameters()) #lr = ...
batch_size = 128
crossEnt = torch.nn.BCELoss() # binary cross entropy
cav_ratio = 0.4 # change CAV percentage here -1 means CAV market place = 0, no neighbor CAV
t_h = 30 # historical step 30
trSet = ngsimDataset('../../data/trajectory/TrainSet_us101.mat',
t_h=t_h,
CAV_ratio=cav_ratio)
valSet = ngsimDataset('../../data/trajectory/ValSet_us101.mat',
t_h=t_h,
CAV_ratio=cav_ratio)
trDataloader = DataLoader(trSet,
batch_size=batch_size,
shuffle=True,
num_workers=8,
collate_fn=trSet.collate_fn)
valDataloader = DataLoader(valSet,
batch_size=batch_size,
shuffle=True,
num_workers=8,
collate_fn=valSet.collate_fn)