def makeTabularTrainer(**config):
cfg = {'dataset':HEPMASS,'network':SmallNN,'net_config': {},
'loader_config': {'amnt_labeled':20+5000,'amnt_dev':5000,'lab_BS':20},
'opt_config': {'lr':1e-4},#{'lr':.1, 'momentum':.9, 'weight_decay':1e-4, 'nesterov':True},
'num_epochs':200,
'unlab_loader_config':{'batch_size':2000,'num_workers':4,'pin_memory':True},
'trainer_config':{'log_dir':os.path.expanduser('~/tb-experiments/UCI/'),'log_args':{'minPeriod':.1, 'timeFrac':3/10}},
}
recursively_update(cfg,config)
trainset = cfg['dataset'](train=True)
testset = cfg['dataset'](train=False)
print(f"Trainset: {len(trainset)}, Testset: {len(testset)}")
device = torch.device('cuda')
model = cfg['network'](num_classes=trainset.num_classes,dim_in=trainset.dim,**cfg['net_config']).to(device)
dataloaders = {}
dataloaders['lab'], dataloaders['dev'] = getLabLoader(trainset,**cfg['loader_config'])
dataloaders['train'] = dataloaders['Train'] = dataloaders['lab']
full_data_loader = DataLoader(trainset,shuffle=True,**cfg['unlab_loader_config'])
dataloaders['_unlab'] = imap(lambda z: z[0], full_data_loader)
dataloaders['test'] = DataLoader(testset,batch_size=cfg['loader_config']['lab_BS'],shuffle=False)
dataloaders = {k:LoaderTo(v,device) for k,v in dataloaders.items()}
opt_constr = lambda params: torch.optim.Adam(params, **cfg['opt_config'])
lr_sched = lambda e: 1.#cosLr(cfg['num_epochs'])
return cfg['trainer'](model,dataloaders,opt_constr,lr_sched,**cfg['trainer_config'])
def makeTrainer(*, task='h**o', device='cuda', lr=3e-3, bs=75, num_epochs=500,network=MolecLieResNet,
net_config={'k':1536,'nbhd':100,'act':'swish','group':lieGroups.T(3),
'bn':True,'aug':True,'mean':True,'num_layers':6}, recenter=False,
subsample=False, trainer_config={'log_dir':None,'log_suffix':''}):#,'log_args':{'timeFrac':1/4,'minPeriod':0}}):
# Create Training set and model
device = torch.device(device)
with FixedNumpySeed(0):
datasets, num_species, charge_scale = QM9datasets()
if subsample: datasets.update(split_dataset(datasets['train'],{'train':subsample}))
ds_stats = datasets['train'].stats[task]
if recenter:
m = datasets['train'].data['charges']>0
pos = datasets['train'].data['positions'][m]
mean,std = pos.mean(dim=0),1#pos.std()
for ds in datasets.values():
ds.data['positions'] = (ds.data['positions']-mean[None,None,:])/std
model = network(num_species,charge_scale,**net_config).to(device)
# Create train and Val(Test) dataloaders and move elems to gpu
dataloaders = {key:LoaderTo(DataLoader(dataset,batch_size=bs,num_workers=0,
shuffle=(key=='train'),pin_memory=False,collate_fn=collate_fn,drop_last=True),
device) for key,dataset in datasets.items()}
# subsampled training dataloader for faster logging of training performance
dataloaders['Train'] = islice(dataloaders['train'],len(dataloaders['test']))#islice(dataloaders['train'],len(dataloaders['train'])//10)
# Initialize optimizer and learning rate schedule
opt_constr = functools.partial(Adam, lr=lr)
cos = cosLr(num_epochs)
lr_sched = lambda e: min(e / (.01 * num_epochs), 1) * cos(e)
return MoleculeTrainer(model,dataloaders,opt_constr,lr_sched,
task=task,ds_stats=ds_stats,**trainer_config)
def makeTrainer(*,
network=CHNN,
net_cfg={},
lr=3e-3,
n_train=800,
regen=False,
dataset=RigidBodyDataset,
body=ChainPendulum(3),
C=5,
dtype=torch.float32,
device=torch.device("cuda"),
bs=200,
num_epochs=100,
trainer_config={},
opt_cfg={'weight_decay': 1e-5}):
# Create Training set and model
angular = not issubclass(network, (CH, CL))
splits = {"train": n_train, "test": 200}
with FixedNumpySeed(0):
dataset = dataset(n_systems=n_train + 200,
regen=regen,
chunk_len=C,
body=body,
angular_coords=angular)
datasets = split_dataset(dataset, splits)
dof_ndim = dataset.body.D if angular else dataset.body.d
model = network(dataset.body.body_graph,
dof_ndim=dof_ndim,
angular_dims=dataset.body.angular_dims,
**net_cfg)
model = model.to(device=device, dtype=dtype)
# Create train and Dev(Test) dataloaders and move elems to gpu
dataloaders = {
k: LoaderTo(DataLoader(v,
batch_size=min(bs, splits[k]),
num_workers=0,
shuffle=(k == "train")),
device=device,
dtype=dtype)
for k, v in datasets.items()
}
dataloaders["Train"] = dataloaders["train"]
# Initialize optimizer and learning rate schedule
opt_constr = lambda params: AdamW(params, lr=lr, **opt_cfg)
lr_sched = cosLr(num_epochs)
return IntegratedDynamicsTrainer(model,
dataloaders,
opt_constr,
lr_sched,
log_args={
"timeFrac": 1 / 4,
"minPeriod": 0.0
},
**trainer_config)
def makeTrainer(config):
cfg = {
'dataset': CIFAR10,
'network': iCNN,
'net_config': {},
'loader_config': {
'amnt_dev': 5000,
'lab_BS': 32,
'pin_memory': True,
'num_workers': 3
},
'opt_config': {
'lr': .0003,
}, # 'momentum':.9, 'weight_decay':1e-4,'nesterov':True},
'num_epochs': 100,
'trainer_config': {},
'parallel': False,
}
recursively_update(cfg, config)
train_transforms = transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip()])
trainset = cfg['dataset'](
'~/datasets/{}/'.format(cfg['dataset']),
flow=True,
)
device = torch.device('cuda')
fullCNN = cfg['network'](num_classes=trainset.num_classes,
**cfg['net_config']).to(device)
if cfg['parallel']: fullCNN = multigpu_parallelize(fullCNN, cfg)
dataloaders = {}
dataloaders['train'], dataloaders['dev'] = getLabLoader(
trainset, **cfg['loader_config'])
dataloaders['Train'] = islice(dataloaders['train'],
10000 // cfg['loader_config']['lab_BS'])
if len(dataloaders['dev']) == 0:
testset = cfg['dataset']('~/datasets/{}/'.format(cfg['dataset']),
train=False,
flow=True)
dataloaders['test'] = DataLoader(
testset,
batch_size=cfg['loader_config']['lab_BS'],
shuffle=False)
dataloaders = {k: LoaderTo(v, device)
for k, v in dataloaders.items()} #LoaderTo(v,device)
opt_constr = lambda params: torch.optim.Adam(params, **cfg['opt_config'
])
lr_sched = cosLr(cfg['num_epochs'])
return Flow(fullCNN, dataloaders, opt_constr, lr_sched,
**cfg['trainer_config'])
def makeTrainer(*,
dataset=YAHOO,
network=SmallNN,
num_epochs=15,
bs=5000,
lr=1e-3,
optim=AdamW,
device='cuda',
trainer=Classifier,
split={
'train': 20,
'val': 5000
},
net_config={},
opt_config={'weight_decay': 1e-5},
trainer_config={
'log_dir': os.path.expanduser('~/tb-experiments/UCI/'),
'log_args': {
'minPeriod': .1,
'timeFrac': 3 / 10
}
},
save=False):
# Prep the datasets splits, model, and dataloaders
with FixedNumpySeed(0):
datasets = split_dataset(dataset(), splits=split)
datasets['_unlab'] = dmap(lambda mb: mb[0], dataset())
datasets['test'] = dataset(train=False)
#print(datasets['test'][0])
device = torch.device(device)
model = network(num_classes=datasets['train'].num_classes,
dim_in=datasets['train'].dim,
**net_config).to(device)
dataloaders = {
k: LoaderTo(
DataLoader(v,
batch_size=min(bs, len(datasets[k])),
shuffle=(k == 'train'),
num_workers=0,
pin_memory=False), device)
for k, v in datasets.items()
}
dataloaders['Train'] = dataloaders['train']
opt_constr = partial(optim, lr=lr, **opt_config)
lr_sched = cosLr(num_epochs) #lambda e:1#
return trainer(model, dataloaders, opt_constr, lr_sched, **trainer_config)
def makeTrainer(*,
dataset=MnistRotDataset,
network=ImgLieResnet,
num_epochs=100,
bs=50,
lr=3e-3,
aug=True,
optim=Adam,
device='cuda',
trainer=Classifier,
split={'train': 12000},
small_test=False,
net_config={},
opt_config={},
trainer_config={'log_dir': None}):
# Prep the datasets splits, model, and dataloaders
datasets = split_dataset(dataset(f'~/datasets/{dataset}/'), splits=split)
datasets['test'] = dataset(f'~/datasets/{dataset}/', train=False)
device = torch.device(device)
model = network(num_targets=datasets['train'].num_targets,
**net_config).to(device)
if aug:
model = torch.nn.Sequential(datasets['train'].default_aug_layers(),
model)
model, bs = try_multigpu_parallelize(model, bs)
dataloaders = {
k: LoaderTo(
DataLoader(v,
batch_size=bs,
shuffle=(k == 'train'),
num_workers=0,
pin_memory=False), device)
for k, v in datasets.items()
}
dataloaders['Train'] = islice(dataloaders['train'],
1 + len(dataloaders['train']) // 10)
if small_test:
dataloaders['test'] = islice(dataloaders['test'],
1 + len(dataloaders['train']) // 10)
# Add some extra defaults if SGD is chosen
opt_constr = partial(optim, lr=lr, **opt_config)
lr_sched = cosLr(num_epochs)
return trainer(model, dataloaders, opt_constr, lr_sched, **trainer_config)
def makeTrainer(*,network,net_cfg,lr=1e-2,n_train=5000,regen=False,
dtype=torch.float32,device=torch.device('cuda'),bs=200,num_epochs=2,
trainer_config={'log_dir':'data_scaling_study_final'}):
# Create Training set and model
splits = {'train':n_train,'val':min(n_train,2000),'test':2000}
dataset = SpringDynamics(n_systems=100000, regen=regen)
with FixedNumpySeed(0):
datasets = split_dataset(dataset,splits)
model = network(**net_cfg).to(device=device,dtype=dtype)
# Create train and Dev(Test) dataloaders and move elems to gpu
dataloaders = {k:LoaderTo(DataLoader(v,batch_size=min(bs,n_train),num_workers=0,shuffle=(k=='train')),
device=device,dtype=dtype) for k,v in datasets.items()}
dataloaders['Train'] = islice(dataloaders['train'],len(dataloaders['val']))
# Initialize optimizer and learning rate schedule
opt_constr = lambda params: Adam(params, lr=lr)
lr_sched = cosLr(num_epochs)
return IntegratedDynamicsTrainer2(model,dataloaders,opt_constr,lr_sched,
log_args={'timeFrac':1/4,'minPeriod':0.0},**trainer_config)
def makeTrainer(config):
cfg = {
'dataset': CIFAR10,
'network': layer13s,
'net_config': {},
'loader_config': {
'amnt_dev': 5000,
'lab_BS': 20,
'pin_memory': True,
'num_workers': 2
},
'opt_config': {
'lr': .3e-4
}, #, 'momentum':.9, 'weight_decay':1e-4,'nesterov':True},
'num_epochs': 100,
'trainer_config': {},
}
recursively_update(cfg, config)
trainset = cfg['dataset']('~/datasets/{}/'.format(cfg['dataset']),
flow=True)
device = torch.device('cuda')
fullCNN = torch.nn.Sequential(
trainset.default_aug_layers(),
cfg['network'](num_classes=trainset.num_classes,
**cfg['net_config']).to(device))
dataloaders = {}
dataloaders['train'], dataloaders['dev'] = getLabLoader(
trainset, **cfg['loader_config'])
dataloaders['Train'] = islice(dataloaders['train'],
10000 // cfg['loader_config']['lab_BS'])
if len(dataloaders['dev']) == 0:
testset = cfg['dataset']('~/datasets/{}/'.format(cfg['dataset']),
train=False)
dataloaders['test'] = DataLoader(
testset,
batch_size=cfg['loader_config']['lab_BS'],
shuffle=False)
dataloaders = {k: LoaderTo(v, device) for k, v in dataloaders.items()}
opt_constr = lambda params: torch.optim.Adam(params, **cfg[
'opt_config']) #torch.optim.SGD(params, **cfg['opt_config'])
lr_sched = cosLr(cfg['num_epochs'])
return iClassifier(fullCNN, dataloaders, opt_constr, lr_sched,
**cfg['trainer_config'])
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
device = torch.device('cuda')
datasets, num_species, charge_scale = QM9datasets()
dataloaders = {
key: LoaderTo(
DataLoader(dataset,
batch_size=5,
num_workers=0,
shuffle=(key == 'train'),
pin_memory=True,
collate_fn=collate_fn), device)
for key, dataset in datasets.items()
}
for mb in dataloaders['train']:
self.mb = mb
break
#meanstd = datasets['train'].stats['h**o']
self.model = MolecLieResNet(num_species,
charge_scale,
nbhd=10,
mean=True,
radius=1.5,
liftsamples=6).to(device)
def make_trainer(
chunk_len: int,
angular: Union[Tuple, bool],
body,
bs: int,
dataset,
dt: float,
lr: float,
n_train: int,
n_val: int,
n_test: int,
net_cfg: dict,
network,
num_epochs: int,
regen: bool,
seed: int = 0,
device=torch.device("cuda"),
dtype=torch.float32,
trainer_config={},
):
# Create Training set and model
splits = {"train": n_train, "val": n_val, "test": n_test}
dataset = dataset(
n_systems=n_train + n_val + n_test,
regen=regen,
chunk_len=chunk_len,
body=body,
dt=dt,
integration_time=10,
angular_coords=angular,
)
# dataset=CartpoleDataset(batch_size=500,regen=regen)
with FixedNumpySeed(seed):
datasets = split_dataset(dataset, splits)
model = network(G=dataset.body.body_graph, **net_cfg).to(device=device,
dtype=dtype)
# Create train and Dev(Test) dataloaders and move elems to gpu
dataloaders = {
k: LoaderTo(
DataLoader(v,
batch_size=min(bs, splits[k]),
num_workers=0,
shuffle=(k == "train")),
device=device,
dtype=dtype,
)
for k, v in datasets.items()
}
dataloaders["Train"] = dataloaders["train"]
# Initialize optimizer and learning rate schedule
opt_constr = lambda params: Adam(params, lr=lr)
lr_sched = cosLr(num_epochs)
return IntegratedDynamicsTrainer(model,
dataloaders,
opt_constr,
lr_sched,
log_args={
"timeFrac": 1 / 4,
"minPeriod": 0.0
},
**trainer_config)
avg_logits = logits2average_depth(filtered_logits,labels[None,:,None,None])
return avg_logits
if __name__=='__main__':
eps_start=1e-3
r = 5
ds=16
niters=1
device = torch.device('cuda')
model = CRFdepthUpsampler(r=r,eps=eps_start,niters=niters).to(device)
trainset= StereoUpsampling05(downsize=ds,val=False,use_vgg=False)
valset = StereoUpsampling05(downsize=ds,val=True,use_vgg=False)
train_loader = DataLoader(trainset,batch_size=1,shuffle=True)
val_loader = DataLoader(valset,batch_size=1,shuffle=True)
dataloaders = {'train':train_loader,'train_':train_loader,'val':val_loader}
dataloaders = {k:LoaderTo(v,device) for k,v in dataloaders.items()}
opt_constr = lambda params: torch.optim.Adam(params, lr=3e-3,betas=(.9,.9))
trialname = 'upsampling/r_{}_{}_niters{}'.format(r,ds,niters)
trainer = Dupsampling(model,dataloaders,opt_constr,log_suffix=trialname,log_args={'minPeriod':.1})
trainer.train(100)
# if __name__=='__main__':
# device = torch.device('cuda')
# eps_start=1e-2
# r = 50
# model = CRFdepthRefiner(r=r,eps=eps_start).to(device)
# trainset=MBStereo14Unary(downsize=8)
# train_loader = DataLoader(trainset,batch_size=1,shuffle=True)
# dataloaders = {'train':train_loader,'train_':train_loader}
# dataloaders = {k:LoaderTo(v,device) for k,v in dataloaders.items()}
# opt_constr = lambda params: torch.optim.Adam(params, lr=3e-3,betas=(.9,.9))
def make_trainer(
train_data,
test_data,
bs=5000,
split={
'train': 200,
'val': 5000
},
network=RealNVPTabularWPrior,
net_config={},
num_epochs=15,
optim=AdamW,
lr=1e-3,
opt_config={'weight_decay': 1e-5},
swag=False,
swa_config={
'swa_dec_pct': .5,
'swa_start_pct': .75,
'swa_freq_pct': .05,
'swa_lr_factor': .1
},
swag_config={
'subspace': 'covariance',
'max_num_models': 20
},
# subspace='covariance', max_num_models=20,
trainer=SemiFlow,
trainer_config={
'log_dir': os.path.expanduser('~/tb-experiments/UCI/'),
'log_args': {
'minPeriod': .1,
'timeFrac': 3 / 10
}
},
dev='cuda',
save=False):
with FixedNumpySeed(0):
datasets = split_dataset(train_data, splits=split)
datasets['_unlab'] = dmap(lambda mb: mb[0], train_data)
datasets['test'] = test_data
device = torch.device(dev)
dataloaders = {
k: LoaderTo(
DataLoader(v,
batch_size=min(bs, len(datasets[k])),
shuffle=(k == 'train'),
num_workers=0,
pin_memory=False), device)
for k, v in datasets.items()
}
dataloaders['Train'] = dataloaders['train']
# model = network(num_classes=train_data.num_classes, dim_in=train_data.dim, **net_config).to(device)
# swag_model = SWAG(model_cfg.base,
# subspace_type=args.subspace, subspace_kwargs={'max_rank': args.max_num_models},
# *model_cfg.args, num_classes=num_classes, **model_cfg.kwargs)
# swag_model.to(args.device)
opt_constr = partial(optim, lr=lr, **opt_config)
model = network(num_classes=train_data.num_classes,
dim_in=train_data.dim,
**net_config).to(device)
if swag:
swag_model = RealNVPTabularSWAG(dim_in=train_data.dim,
**net_config,
**swag_config)
# swag_model = SWAG(RealNVPTabular,
# subspace_type=subspace, subspace_kwargs={'max_rank': max_num_models},
# num_classes=train_data.num_classes, dim_in=train_data.dim,
# num_coupling_layers=coupling_layers,in_dim=dim_in,**net_config)
# swag_model.to(device)
# swag_model = SWAG(RealNVPTabular, num_classes=train_data.num_classes, dim_in=train_data.dim,
# swag=True, **swag_config, **net_config)
# model.to(device)
swag_model.to(device)
swa_config['steps_per_epoch'] = len(dataloaders['_unlab'])
swa_config['num_epochs'] = num_epochs
lr_sched = swa_learning_rate(**swa_config)
# lr_sched = cosLr(num_epochs)
return trainer(model,
dataloaders,
opt_constr,
lr_sched,
swag_model=swag_model,
**swa_config,
**trainer_config)
else:
# model = network(num_classes=train_data.num_classes, dim_in=train_data.dim, **net_config).to(device)
lr_sched = cosLr(num_epochs)
# lr_sched = lambda e:1
return trainer(model, dataloaders, opt_constr, lr_sched,
**trainer_config)