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)