def main(): model = PSMNet(args.maxdisp, args.mindisp).cuda() if args.load_model is not None: if args.load is not None: warn('args.load is not None. load_model will be covered by load.') ckpt = torch.load(args.load_model, 'cpu') if 'model' in ckpt.keys(): pretrained = ckpt['model'] elif 'state_dict' in ckpt.keys(): pretrained = ckpt['state_dict'] else: raise RuntimeError() pretrained = { k.replace('module.', ''): v for k, v in pretrained.items() } model.load_state_dict(pretrained) train_dl = DataLoader(KITTIRoiDataset(args.data_dir, 'train', args.resolution, args.maxdisp, args.mindisp), batch_size=args.batch_size, shuffle=True, num_workers=args.workers) val_dl = DataLoader(KITTIRoiDataset(args.data_dir, 'val', args.resolution, args.maxdisp, args.mindisp), batch_size=args.batch_size, num_workers=args.workers) loss_fn = PSMLoss() databunch = DataBunch(train_dl, val_dl, device='cuda') learner = Learner(databunch, model, loss_func=loss_fn, model_dir=args.model_dir) learner.callbacks = [ DistributedSaveModelCallback(learner), TensorBoardCallback(learner) ] if num_gpus > 1: learner.to_distributed(get_rank()) if args.load is not None: learner.load(args.load) if args.mode == 'train': learner.fit(args.epochs, args.maxlr) elif args.mode == 'train_oc': fit_one_cycle(learner, args.epochs, args.maxlr) else: raise ValueError('args.mode not supported.')
def build_learner( databunch, model, loss_func=nn.MSELoss(), pretrained_learner_fname=pretrained_learner_fname, ): learn = Learner(data=databunch, model=model, loss_func=loss_func) if pretrained_learner_fname: learn = learn.load(pretrained_learner_fname) return learn
siamese.to(device) # In[11]: learn = Learner( data, siamese, #enable_validate=True, path=learn_path, #loss_func=BCEWithLogitsFlat(), loss_func=ContrastiveLoss(margin=contrastive_neg_margin), metrics=[avg_pos_dist, avg_neg_dist] #metrics=[lambda preds, targs: accuracy_thresh(preds.squeeze(), targs, sigmoid=False)] ) learn.load(f'{name}_80') dist_mat, val_target, _ = cal_mat(learn.model, data.valid_dl, data.fix_dl, ds_with_target1=True, ds_with_target2=True) for threshold in np.linspace(1, 9, 30): top5_matrix, map5 = cal_mapk(dist_mat, val_target, k=5, threshold=threshold, ref_idx2class=learn.data.fix_dl.y, target_idx2class=learn.data.valid_dl.y) print(threshold, map5)
def main(args): if args.deterministic: set_seed(42) # Set device if args.device is None: if torch.cuda.is_available(): args.device = 'cuda:0' else: args.device = 'cpu' defaults.device = torch.device(args.device) # Aggregate path and labels into list for fastai ImageDataBunch fnames, labels, is_valid = [], [], [] dataset = OpenFire(root=args.data_path, train=True, download=True, img_folder=args.img_folder) for sample in dataset.data: fnames.append( dataset._images.joinpath(sample['name']).relative_to(dataset.root)) labels.append(sample['target']) is_valid.append(False) dataset = OpenFire(root=args.data_path, train=False, download=True) for sample in dataset.data: fnames.append( dataset._images.joinpath(sample['name']).relative_to(dataset.root)) labels.append(sample['target']) is_valid.append(True) df = pd.DataFrame.from_dict( dict(name=fnames, label=labels, is_valid=is_valid)) # Split train and valid sets il = vision.ImageList.from_df( df, path=args.data_path).split_from_df('is_valid') # Encode labels il = il.label_from_df(cols='label', label_cls=FloatList if args.binary else CategoryList) # Set transformations il = il.transform(vision.get_transforms(), size=args.resize) # Create the Databunch data = il.databunch(bs=args.batch_size, num_workers=args.workers).normalize( vision.imagenet_stats) # Metric metric = partial(vision.accuracy_thresh, thresh=0.5) if args.binary else vision.error_rate # Create model model = models.__dict__[args.model](imagenet_pretrained=args.pretrained, num_classes=data.c, lin_features=args.lin_feats, concat_pool=args.concat_pool, bn_final=args.bn_final, dropout_prob=args.dropout_prob) # Create learner learner = Learner(data, model, wd=args.weight_decay, loss_func=CustomBCELogitsLoss() if args.binary else nn.CrossEntropyLoss(), metrics=metric) # Form layer group for optimization meta = model_meta.get(args.model, _default_meta) learner.split(meta['split']) # Freeze model's head if args.pretrained: learner.freeze() if args.resume: learner.load(args.resume) if args.unfreeze: learner.unfreeze() learner.fit_one_cycle(args.epochs, max_lr=slice(None, args.lr, None), div_factor=args.div_factor, final_div=args.final_div_factor) learner.save(args.checkpoint)
def train(self, tmp_dir): """Train a model.""" self.print_options() # Sync output of previous training run from cloud. train_uri = self.backend_opts.train_uri train_dir = get_local_path(train_uri, tmp_dir) make_dir(train_dir) sync_from_dir(train_uri, train_dir) # Get zip file for each group, and unzip them into chip_dir. chip_dir = join(tmp_dir, 'chips') make_dir(chip_dir) for zip_uri in list_paths(self.backend_opts.chip_uri, 'zip'): zip_path = download_if_needed(zip_uri, tmp_dir) with zipfile.ZipFile(zip_path, 'r') as zipf: zipf.extractall(chip_dir) # Setup data loader. train_images = [] train_lbl_bbox = [] for annotation_path in glob.glob(join(chip_dir, 'train/*.json')): images, lbl_bbox = get_annotations(annotation_path) train_images += images train_lbl_bbox += lbl_bbox val_images = [] val_lbl_bbox = [] for annotation_path in glob.glob(join(chip_dir, 'valid/*.json')): images, lbl_bbox = get_annotations(annotation_path) val_images += images val_lbl_bbox += lbl_bbox images = train_images + val_images lbl_bbox = train_lbl_bbox + val_lbl_bbox img2bbox = dict(zip(images, lbl_bbox)) get_y_func = lambda o: img2bbox[o.name] num_workers = 0 if self.train_opts.debug else 4 data = ObjectItemList.from_folder(chip_dir) data = data.split_by_folder() data = data.label_from_func(get_y_func) data = data.transform( get_transforms(), size=self.task_config.chip_size, tfm_y=True) data = data.databunch( bs=self.train_opts.batch_sz, collate_fn=bb_pad_collate, num_workers=num_workers) print(data) if self.train_opts.debug: make_debug_chips( data, self.task_config.class_map, tmp_dir, train_uri) # Setup callbacks and train model. ratios = [1/2, 1, 2] scales = [1, 2**(-1/3), 2**(-2/3)] model_arch = getattr(models, self.train_opts.model_arch) encoder = create_body(model_arch, cut=-2) model = RetinaNet(encoder, data.c, final_bias=-4) crit = RetinaNetFocalLoss(scales=scales, ratios=ratios) learn = Learner(data, model, loss_func=crit, path=train_dir) learn = learn.split(retina_net_split) model_path = get_local_path(self.backend_opts.model_uri, tmp_dir) pretrained_uri = self.backend_opts.pretrained_uri if pretrained_uri: print('Loading weights from pretrained_uri: {}'.format( pretrained_uri)) pretrained_path = download_if_needed(pretrained_uri, tmp_dir) learn.load(pretrained_path[:-4]) callbacks = [ TrackEpochCallback(learn), SaveModelCallback(learn, every='epoch'), MyCSVLogger(learn, filename='log'), ExportCallback(learn, model_path), SyncCallback(train_dir, self.backend_opts.train_uri, self.train_opts.sync_interval) ] learn.unfreeze() learn.fit(self.train_opts.num_epochs, self.train_opts.lr, callbacks=callbacks) # Since model is exported every epoch, we need some other way to # show that training is finished. str_to_file('done!', self.backend_opts.train_done_uri) # Sync output to cloud. sync_to_dir(train_dir, self.backend_opts.train_uri)
def run_ner( lang: str = 'eng', log_dir: str = 'logs', task: str = NER, batch_size: int = 1, lr: float = 5e-5, epochs: int = 1, dataset: str = 'data/conll-2003/', loss: str = 'cross', max_seq_len: int = 128, do_lower_case: bool = False, warmup_proportion: float = 0.1, grad_acc_steps: int = 1, rand_seed: int = None, fp16: bool = False, loss_scale: float = None, ds_size: int = None, data_bunch_path: str = 'data/conll-2003/db', bertAdam: bool = False, freez: bool = False, one_cycle: bool = False, discr: bool = False, lrm: int = 2.6, div: int = None, tuned_learner: str = None, do_train: str = False, do_eval: str = False, save: bool = False, name: str = 'ner', mask: tuple = ('s', 's'), ): name = "_".join( map(str, [ name, task, lang, mask[0], mask[1], loss, batch_size, lr, max_seq_len, do_train, do_eval ])) log_dir = Path(log_dir) log_dir.mkdir(parents=True, exist_ok=True) init_logger(log_dir, name) if rand_seed: random.seed(rand_seed) np.random.seed(rand_seed) torch.manual_seed(rand_seed) if torch.cuda.is_available(): torch.cuda.manual_seed_all(rand_seed) trainset = dataset + lang + '/train.txt' devset = dataset + lang + '/dev.txt' testset = dataset + lang + '/test.txt' bert_model = 'bert-base-cased' if lang == 'eng' else 'bert-base-multilingual-cased' print(f'Lang: {lang}\nModel: {bert_model}\nRun: {name}') model = BertForTokenClassification.from_pretrained(bert_model, num_labels=len(VOCAB), cache_dir='bertm') model = torch.nn.DataParallel(model) model_lr_group = bert_layer_list(model) layers = len(model_lr_group) kwargs = {'max_seq_len': max_seq_len, 'ds_size': ds_size, 'mask': mask} train_dl = DataLoader(dataset=NerDataset(trainset, bert_model, train=True, **kwargs), batch_size=batch_size, shuffle=True, collate_fn=partial(pad, train=True)) dev_dl = DataLoader(dataset=NerDataset(devset, bert_model, **kwargs), batch_size=batch_size, shuffle=False, collate_fn=pad) test_dl = DataLoader(dataset=NerDataset(testset, bert_model, **kwargs), batch_size=batch_size, shuffle=False, collate_fn=pad) data = DataBunch(train_dl=train_dl, valid_dl=dev_dl, test_dl=test_dl, collate_fn=pad, path=Path(data_bunch_path)) loss_fun = ner_loss_func if loss == 'cross' else partial(ner_loss_func, zero=True) metrics = [Conll_F1()] learn = Learner( data, model, BertAdam, loss_func=loss_fun, metrics=metrics, true_wd=False, layer_groups=None if not freez else model_lr_group, path='learn', ) # initialise bert adam optimiser train_opt_steps = int(len(train_dl.dataset) / batch_size) * epochs optim = BertAdam(model.parameters(), lr=lr, warmup=warmup_proportion, t_total=train_opt_steps) if bertAdam: learn.opt = OptimWrapper(optim) else: print("No Bert Adam") # load fine-tuned learner if tuned_learner: print('Loading pretrained learner: ', tuned_learner) learn.load(tuned_learner) # Uncomment to graph learning rate plot # learn.lr_find() # learn.recorder.plot(skip_end=15) # set lr (discriminative learning rates) if div: layers = div lrs = lr if not discr else learn.lr_range(slice(lr / lrm**(layers), lr)) results = [['epoch', 'lr', 'f1', 'val_loss', 'train_loss', 'train_losses']] if do_train: for epoch in range(epochs): if freez: lay = (layers // (epochs - 1)) * epoch * -1 if lay == 0: print('Freeze') learn.freeze() elif lay == layers: print('unfreeze') learn.unfreeze() else: print('freeze2') learn.freeze_to(lay) print('Freezing layers ', lay, ' off ', layers) # Fit Learner - eg train model if one_cycle: learn.fit_one_cycle(1, lrs, moms=(0.8, 0.7)) else: learn.fit(1, lrs) results.append([ epoch, lrs, learn.recorder.metrics[0][0], learn.recorder.val_losses[0], np.array(learn.recorder.losses).mean(), learn.recorder.losses, ]) if save: m_path = learn.save(f"{lang}_{epoch}_model", return_path=True) print(f'Saved model to {m_path}') if save: learn.export(f'{lang}.pkl') if do_eval: res = learn.validate(test_dl, metrics=metrics) met_res = [f'{m.__name__}: {r}' for m, r in zip(metrics, res[1:])] print(f'Validation on TEST SET:\nloss {res[0]}, {met_res}') results.append(['val', '-', res[1], res[0], '-', '-']) with open(log_dir / (name + '.csv'), 'a') as resultFile: wr = csv.writer(resultFile) wr.writerows(results)
name=model_se_str), partial(WarmRestartsLRScheduler, n_cycles=args.epochs, lr=(args.lr, args.lr / args.lr_div), mom=(args.mom, 0.95), cycle_len=args.cycle_len, cycle_mult=args.cycle_mult, start_epoch=args.start_epoch) ] learn = Learner(db, model, metrics=[rmse, mae], callback_fns=callback_fns, wd=args.wd, loss_func=contribs_rmse_loss) if args.start_epoch > 0: learn.load(model_se_str + f'_{args.start_epoch-1}') else: learn.load(model_str) torch.cuda.empty_cache() if distributed_train: learn = learn.to_distributed(args.local_rank) learn.fit(args.epochs) # make predictions n_val = len(train_df[train_df['molecule_id'].isin(val_mol_ids)]) val_preds = np.zeros((n_val, args.epochs)) test_preds = np.zeros((len(test_df), args.epochs)) for m in range(args.epochs): print(f'Predicting for model {m}') learn.load(model_se_str + f'_{m}') val_contrib_preds = learn.get_preds(DatasetType.Valid) test_contrib_preds = learn.get_preds(DatasetType.Test)
callback_fns = [ partial(GradientClipping, clip=10), GroupMeanLogMAE, partial(SaveModelCallback, every='improvement', mode='min', monitor='group_mean_log_mae', name=model_str) ] learn = Learner(db, model, metrics=[rmse, mae], callback_fns=callback_fns, wd=args.wd, loss_func=contribs_rmse_loss) if args.start_epoch > 0: learn.load(model_str) torch.cuda.empty_cache() if distributed_train: learn = learn.to_distributed(args.local_rank) learn.fit_one_cycle(args.epochs, max_lr=args.lr, start_epoch=args.start_epoch) # make predictions val_contrib_preds = learn.get_preds(DatasetType.Valid) test_contrib_preds = learn.get_preds(DatasetType.Test) val_preds = val_contrib_preds[0][:, -1].detach().numpy() * C.SC_STD + C.SC_MEAN test_preds = test_contrib_preds[0][:, -1].detach().numpy() * C.SC_STD + C.SC_MEAN # store results store_submit(test_preds, model_str, print_head=True) store_oof(val_preds, model_str, print_head=True)