def train(self, dsId=None):
if dsId:
self.args['dsId'] = dsId
self.net.train_flag = True
self.net.train()
if self.args["cuda"]:
self.net.cuda(self.args['device'])
if self.args["optim"] == "Adam":
optim = torch.optim.Adam(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "SGD":
optim = torch.optim.SGD(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "AdamW":
optim = torch.optim.AdamW(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "SparseAdam":
optim = torch.optim.SparseAdam(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "Adamax":
optim = torch.optim.Adamax(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "ASGD":
optim = torch.optim.ASGD(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "Rprop":
optim = torch.optim.Rprop(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "RMSprop":
optim = torch.optim.RMSprop(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "LBFGS":
optim = torch.optim.LBFGS(self.net.parameters(),lr=self.args['learning_rate'])
else:
print("undefined optimizer.")
return
crossEnt = torch.nn.BCELoss()
print('loading data in {}...'.format(self.args['dsId']))
trSet_path = os.path.join(self.args["dir"], "trainSet")
valSet_path = os.path.join(self.args["dir"], "valSet")
trSet = ngsimDataset(trSet_path, self.args["dir"], self.args["raw_dir"], 'train', self.args['dsId'], t_h=self.args['in_length'], t_f=self.args['out_length'])
valSet = ngsimDataset(valSet_path, self.args["dir"], self.args["raw_dir"], 'val', self.args['dsId'], t_h=self.args['in_length'], t_f=self.args['out_length'])
trDataloader = DataLoader(trSet,batch_size=self.args['batch_size'],shuffle=True,num_workers=8,collate_fn=trSet.collate_fn)
# trDataloader = DataLoader(valSet,batch_size=self.args['batch_size'],shuffle=True,num_workers=8,collate_fn=valSet.collate_fn)
valDataloader = DataLoader(valSet,batch_size=self.args['batch_size'],shuffle=True,num_workers=8,collate_fn=valSet.collate_fn)
print('start training {}...'.format(self.args["predAlgo"]))
if self.args["predAlgo"] == "Traphic":
engine = TraphicEngine(self.net, optim, trDataloader, valDataloader, self.args)
else:
engine = SocialEngine(self.net, optim, trDataloader, valDataloader, self.args)
engine.start()
def result_viz(self): #TODO
# This function is for visualizing the network output and ground truth trajectory
# print('loading data in {}...'.format(self.args['dsId']))
tstSet_path = os.path.join(self.args["dir"], "testSet")
testSet = ngsimDataset(valSet_path,
self.args["dir"],
self.args["raw_dir"],
'val',
self.args['dsId'],
t_h=self.args['in_length'],
t_f=self.args['out_length'])
testDataloader = DataLoader(testSet,
batch_size=self.args['batch_size'],
shuffle=True,
num_workers=4,
collate_fn=testSet.collate_fn)
if self.args["predAlgo"] == "Traphic":
engine = TraphicEngine(self.net, optim, trDataloader,
valDataloader, self.args)
else:
engine = SocialEngine(self.net, optim, trDataloader, valDataloader,
self.args)
def evaluate(self, dsId=None):
if dsId:
self.args['dsId'] = dsId
self.net.train_flag = False
self.net.eval()
d = os.path.join(self.args['modelLoc'], self.args['name'])
if os.path.exists(d):
self.net.load_state_dict(torch.load(d))
print("\n[INFO]: model {} loaded".format(d))
else:
print("\n[INFO]: can not find model at {} to evaluate, using existing net".format(d))
if self.args["cuda"]:
self.net.cuda(self.args['device'])
if self.args["optim"] == "Adam":
optim = torch.optim.Adam(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "SGD":
optim = torch.optim.SGD(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "AdamW":
optim = torch.optim.AdamW(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "SparseAdam":
optim = torch.optim.SparseAdam(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "Adamax":
optim = torch.optim.Adamax(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "ASGD":
optim = torch.optim.ASGD(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "Rprop":
optim = torch.optim.Rprop(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "RMSprop":
optim = torch.optim.RMSprop(self.net.parameters(),lr=self.args['learning_rate'])
elif self.args["optim"] == "LBFGS":
optim = torch.optim.LBFGS(self.net.parameters(),lr=self.args['learning_rate'])
else:
print("undefined optimizer.")
return
crossEnt = torch.nn.BCELoss()
print('loading data in {}...'.format(self.args['dsId']))
trSet_path = os.path.join(self.args["dir"], "trainSet")
valSet_path = os.path.join(self.args["dir"], "valSet")
tstSet_path = os.path.join(self.args["dir"], "testSet")
trSet = ngsimDataset(trSet_path, self.args["dir"], self.args["raw_dir"], 'train', self.args['dsId'], t_h=self.args['in_length'], t_f=self.args['out_length'])
trDataloader = DataLoader(trSet,batch_size=self.args['batch_size'],shuffle=True,num_workers=8,collate_fn=trSet.collate_fn)
testSet = ngsimDataset(valSet_path, self.args["dir"], self.args["raw_dir"], 'val', self.args['dsId'], t_h=self.args['in_length'], t_f=self.args['out_length'])
testDataloader = DataLoader(testSet,batch_size=self.args['batch_size'],shuffle=True,num_workers=8,collate_fn=testSet.collate_fn)
valSet = ngsimDataset(tstSet_path, self.args["dir"], self.args["raw_dir"], 'val', self.args['dsId'], t_h=self.args['in_length'], t_f=self.args['out_length'])
valDataloader = DataLoader(valSet,batch_size=self.args['batch_size'],shuffle=True,num_workers=8,collate_fn=valSet.collate_fn)
print('start testing {}...'.format(self.args["predAlgo"]))
if self.args["predAlgo"] == "Traphic":
engine = TraphicEngine(self.net, optim, trDataloader, valDataloader, self.args)
else:
engine = SocialEngine(self.net, optim, trDataloader, valDataloader, self.args)
engine.eval(testDataloader)
def evaluate(self, viewArgs, thread=None):
if thread:
thread.signalCanvas("\n[INFO] Evaluation started...")
args = self.getPredArgs(viewArgs)
args['eval'] = True
predAlgo = viewArgs["predAlgo"]
optimSelection = viewArgs["optim"]
if predAlgo == "Traphic":
if thread:
thread.signalCanvas(
"\n[INFO]: Using Traphic for the saved model")
args['train_flag'] = False
args["ours"] = True
net = traphicNet(args)
elif predAlgo == "Social GAN":
sganEval(sgan_args, thread)
return
elif predAlgo == "Social-LSTM":
print(predAlgo)
elif predAlgo == "Social Conv":
if thread:
thread.signalCanvas(
"\n[INFO]: Using Convolutional Social Pooling")
args['train_flag'] = False
net = highwayNet(args)
net.eval()
d = os.path.join(args['modelLoc'])
if thread:
thread.signalCanvas(d)
if os.path.exists(d):
net.load_state_dict(torch.load(d))
if thread:
thread.signalCanvas("\n[INFO]: model loaded")
else:
if thread:
thread.signalCanvas("\n[INFO]: can not find model to evaluate")
if args["cuda"]:
if thread:
thread.signalCanvas("\n[INFO]: Using CUDA")
net.cuda()
if optimSelection == "Adam":
optim = torch.optim.Adam(net.parameters(),
lr=args['learning_rate'])
if thread:
thread.signalCanvas("\n[INFO]: Optimizer: \n" + str(optim))
else:
if thread:
thread.signalCanvas("\n[INFO]: NOT YET IMPLEMENTED")
return
crossEnt = torch.nn.BCELoss()
if thread:
thread.signalCanvas("\n[INFO]: Loss: \n" + str(crossEnt))
# TODO: More hardcodes
dataset_name = viewArgs["dir"].split('/')[2]
prediction_data_path = 'model/Prediction/data/{}'.format(dataset_name)
trSet_path = os.path.join(prediction_data_path, "TrainSet.npy")
valSet_path = os.path.join(prediction_data_path, "ValSet.npy")
tstSet_path = os.path.join(prediction_data_path, "TestSet.npy")
trSet = ngsimDataset(trSet_path)
trDataloader = DataLoader(trSet,
batch_size=args['batch_size'],
shuffle=True,
num_workers=8,
collate_fn=trSet.collate_fn)
testSet = ngsimDataset(valSet_path)
testDataloader = DataLoader(testSet,
batch_size=args['batch_size'],
shuffle=True,
num_workers=8,
collate_fn=testSet.collate_fn)
valSet = ngsimDataset(tstSet_path)
valDataloader = DataLoader(valSet,
batch_size=args['batch_size'],
shuffle=True,
num_workers=8,
collate_fn=valSet.collate_fn)
if predAlgo == "Traphic":
args["ours"] = True
engine = TraphicEngine(net, optim, trDataloader, testDataloader,
args, thread)
elif predAlgo == "Social Conv":
engine = SocialEngine(net, optim, trDataloader, testDataloader,
args, thread)
else:
if thread:
thread.signalCanvas("\n[INFO]: NOT YET IMPLEMENTED")
if thread:
thread.signalCanvas(
"\n[INFO]: *** Evaluating Prediction Model ***")
engine.eval(testDataloader)
def train(self, viewArgs, thread=None):
if thread:
thread.signalCanvas("\n[INFO] Training started...")
args = self.getPredArgs(viewArgs)
args['eval'] = False
predAlgo = viewArgs["predAlgo"]
optimSelection = viewArgs["optim"]
if predAlgo == "Traphic":
if thread:
thread.signalCanvas("\n[INFO]: Using TRAPHIC model")
args["ours"] = True
args['train_flag'] = True
net = traphicNet(args)
elif predAlgo == "Social GAN":
sgan_args.num_epochs = int(args["pretrainEpochs"]) + int(
args["trainEpochs"])
sgan_args.batch_size = args["batch_size"]
sgan_args.dropout = args['dropout_prob']
sgan_args.g_learning_rate = args["learning_rate"]
sgan_args.g_learning_rate = args["learning_rate"]
if thread:
thread.signalCanvas("\n[INFO]: Using Sgan model")
thread.signalCanvas(
"\n[INFO]: *** Training Prediction Model ***")
sganTrain(sgan_args, thread)
return
elif predAlgo == "Social-LSTM":
print(predAlgo)
elif predAlgo == "Social Conv":
if thread:
thread.signalCanvas(
"\n[INFO]: Using Convolutional Social Pooling")
args['train_flag'] = True
net = highwayNet(args)
if args["cuda"]:
if thread:
thread.signalCanvas("\n[INFO]: Using CUDA")
net.cuda()
if optimSelection == "Adam":
optim = torch.optim.Adam(net.parameters(),
lr=args['learning_rate'])
if thread:
thread.signalCanvas("\n[INFO]: Optimizer: \n" + str(optim))
else:
if thread:
thread.signalCanvas("\n[INFO]: NOT YET IMPLEMENTED")
return
crossEnt = torch.nn.BCELoss()
if thread:
thread.signalCanvas("\n[INFO]: Loss: \n" + str(crossEnt))
# name of
dataset_name = viewArgs["dir"].split('/')[2]
prediction_data_path = 'model/Prediction/data/{}'.format(dataset_name)
trSet_path = os.path.join(prediction_data_path, "TrainSet.npy")
valSet_path = os.path.join(prediction_data_path, "ValSet.npy")
trSet = ngsimDataset(trSet_path)
valSet = ngsimDataset(valSet_path)
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 predAlgo == "Traphic":
args["ours"] = True
engine = TraphicEngine(net, optim, trDataloader, valDataloader,
args, thread)
elif predAlgo == "Social Conv":
engine = SocialEngine(net, optim, trDataloader, valDataloader,
args, thread)
else:
if thread:
thread.signalCanvas("\n[INFO]: NOT YET IMPLEMENTED")
if thread:
thread.signalCanvas("\n[INFO]: *** Training Prediction Model ***")
engine.start()