def create( self, arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum=0.9, weight_decay=5e-4, pretrained=False, size_input=388, num_classes=8, backbone='preactresnet', num_filters=32, breal='real', alpha=2, beta=2, ): """ Create -arch (string): architecture -loss (string): -lr (float): learning rate -optimizer (string) : -lrsch (string): scheduler learning rate -pretrained (bool) """ cfg_opt = {'momentum': momentum, 'weight_decay': weight_decay} #cfg_scheduler={ 'step_size':100, 'gamma':0.1 } cfg_scheduler = {'mode': 'min', 'patience': 10} cfg_model = {'num_filters': num_filters} self.num_classes = num_classes super(ClassNeuralNet, self).create( arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, pretrained, cfg_opt=cfg_opt, cfg_scheduler=cfg_scheduler, cfg_model=cfg_model, ) self.size_input = size_input self.backbone = backbone self.num_filters = num_filters self.topk = nloss.TopkAccuracy() self.logger_train = Logger('Train', ['loss', 'loss_bce'], ['topk'], self.plotter) self.logger_val = Logger('Val ', ['loss', 'loss_bce'], ['topk'], self.plotter) self.breal = breal
def create( self, arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum=0.9, weight_decay=5e-4, pretrained=False, topk=(1, ), size_input=128, ): """ Create Args: arch (string): architecture num_output_channels, num_input_channels, loss (string): lr (float): learning rate momentum, optimizer (string) : lrsch (string): scheduler learning rate pretrained (bool) """ cfg_opt = {'momentum': 0.9, 'weight_decay': 5e-4} cfg_scheduler = {'step_size': 100, 'gamma': 0.1} super(NeuralNetClassifier, self).create( arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, pretrained, cfg_opt=cfg_opt, cfg_scheduler=cfg_scheduler, ) self.size_input = size_input self.accuracy = nloss.TopkAccuracy(topk) self.cnf = nloss.ConfusionMeter(self.num_output_channels, normalized=True) self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject) # Set the graphic visualization self.metrics_name = ['top{}'.format(k) for k in topk] self.logger_train = Logger('Trn', ['loss'], self.metrics_name, self.plotter) self.logger_val = Logger('Val', ['loss'], self.metrics_name, self.plotter)
def create( self, arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum=0.9, weight_decay=5e-4, pretrained=False, th=0.0, size_input=32, ): """ Create Args: arch (string): architecture num_output_channels, num_input_channels, loss (string): lr (float): learning rate momentum, optimizer (string) : lrsch (string): scheduler learning rate pretrained (bool) """ cfg_opt = {'momentum': 0.9, 'weight_decay': 5e-4} cfg_scheduler = {'step_size': 50, 'gamma': 0.1} super(NeuralNetClassifier, self).create( arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, pretrained, cfg_opt=cfg_opt, cfg_scheduler=cfg_scheduler, ) self.size_input = size_input self.accuracy = nloss.MultAccuracyV1(th) self.f_score = nloss.F_score(threshold=th, beta=2) #self.cnf = nloss.ConfusionMeter( self.num_output_channels, normalized=True ) #self.visheatmap = gph.HeatMapVisdom( env_name=self.nameproject ) # Set the graphic visualization self.logger_train = Logger('Trn', ['loss'], ['acc', 'f1'], self.plotter) self.logger_val = Logger('Val', ['loss'], ['acc', 'f1'], self.plotter)
def create(self, arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum=0.9, weight_decay=5e-4, pretrained=False, size_input=388, num_classes=8, ): """ Create Args: -arch (string): architecture -num_output_channels, -num_input_channels, -loss (string): -lr (float): learning rate -optimizer (string) : -lrsch (string): scheduler learning rate -pretrained (bool) - """ super(AttentionGMMNeuralNet, self).create( arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum, weight_decay, pretrained, size_input, num_classes, ) self.logger_train = Logger( 'Train', ['loss', 'loss_gmm', 'loss_bce', 'loss_att' ], [ 'topk', 'gmm'], self.plotter ) self.logger_val = Logger( 'Val ', ['loss', 'loss_gmm', 'loss_bce', 'loss_att' ], [ 'topk', 'gmm'], self.plotter )
def create( self, arch, num_output_channels, num_input_channels, loss, lr, momentum, optimizer, lrsch, pretrained=False, topk=(1, ), ): """ Create Args: @arch (string): architecture @num_output_channels, @num_input_channels, @loss (string): @lr (float): learning rate @momentum, @optimizer (string) : @lrsch (string): scheduler learning rate @pretrained (bool) """ super(NeuralNetClassifier, self).create(arch, num_output_channels, num_input_channels, loss, lr, momentum, optimizer, lrsch, pretrained) self.accuracy = nloss.Accuracy(topk) self.cnf = nloss.ConfusionMeter(self.num_output_channels, normalized=True) self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject) # Set the graphic visualization self.metrics_name = ['top{}'.format(k) for k in topk] self.logger_train = Logger('Trn', ['loss'], self.metrics_name, self.plotter) self.logger_val = Logger('Val', ['loss'], self.metrics_name, self.plotter)
def create( self, arch, num_output_channels, num_input_channels, loss, lr, momentum, optimizer, lrsch, pretrained=False, size_input=388, ): """ Create -arch (string): architecture -loss (string): -lr (float): learning rate -optimizer (string) : -lrsch (string): scheduler learning rate -pretrained (bool) """ super(SegmentationNeuralNet, self).create(arch, num_output_channels, num_input_channels, loss, lr, momentum, optimizer, lrsch, pretrained) self.size_input = size_input self.accuracy = nloss.Accuracy() self.dice = nloss.Dice() # Set the graphic visualization self.logger_train = Logger('Train', ['loss'], ['accs', 'dices'], self.plotter) self.logger_val = Logger('Val ', ['loss'], ['accs', 'dices'], self.plotter) self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject, heatsize=(100, 100)) self.visimshow = gph.ImageVisdom(env_name=self.nameproject, imsize=(100, 100))
class NeuralNetClassifier(NeuralNetAbstract): r"""Convolutional Neural Net for classification Args: patchproject (str): path project nameproject (str): name project no_cuda (bool): system cuda (default is True) parallel (bool) seed (int) print_freq (int) gpu (int) """ def __init__(self, patchproject, nameproject, no_cuda=True, parallel=False, seed=1, print_freq=10, gpu=0): super(NeuralNetClassifier, self).__init__(patchproject, nameproject, no_cuda, parallel, seed, print_freq, gpu) def create( self, arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum=0.9, weight_decay=5e-4, pretrained=False, topk=(1, ), size_input=128, ): """ Create Args: arch (string): architecture num_output_channels, num_input_channels, loss (string): lr (float): learning rate momentum, optimizer (string) : lrsch (string): scheduler learning rate pretrained (bool) """ cfg_opt = {'momentum': 0.9, 'weight_decay': 5e-4} cfg_scheduler = {'step_size': 100, 'gamma': 0.1} super(NeuralNetClassifier, self).create( arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, pretrained, cfg_opt=cfg_opt, cfg_scheduler=cfg_scheduler, ) self.size_input = size_input self.accuracy = nloss.TopkAccuracy(topk) self.cnf = nloss.ConfusionMeter(self.num_output_channels, normalized=True) self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject) # Set the graphic visualization self.metrics_name = ['top{}'.format(k) for k in topk] self.logger_train = Logger('Trn', ['loss'], self.metrics_name, self.plotter) self.logger_val = Logger('Val', ['loss'], self.metrics_name, self.plotter) def training(self, data_loader, epoch=0): self.logger_train.reset() data_time = AverageMeter() batch_time = AverageMeter() # switch to evaluate mode self.net.train() end = time.time() for i, sample in enumerate(data_loader): # measure data loading time data_time.update(time.time() - end) # get data (image, label) x, y = sample['image'], sample['label'].argmax(1).long() batch_size = x.size(0) if self.cuda: x = x.cuda() y = y.cuda() # fit (forward) outputs = self.net(x) # measure accuracy and record loss loss = self.criterion(outputs, y.long()) pred = self.accuracy(outputs.data, y) # optimizer self.optimizer.zero_grad() loss.backward() self.optimizer.step() # update self.logger_train.update( {'loss': loss.item()}, dict( zip(self.metrics_name, [ pred[p].item() for p in range(len(self.metrics_name)) ])), batch_size, ) # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % self.print_freq == 0: self.logger_train.logger( epoch, epoch + float(i + 1) / len(data_loader), i, len(data_loader), batch_time, ) def evaluate(self, data_loader, epoch=0): self.logger_val.reset() self.cnf.reset() batch_time = AverageMeter() # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, sample in enumerate(data_loader): # get data (image, label) x, y = sample['image'], sample['label'].argmax(1).long() batch_size = x.size(0) if self.cuda: x = x.cuda() y = y.cuda() # fit (forward) outputs = self.net(x) # measure accuracy and record loss loss = self.criterion(outputs, y) pred = self.accuracy(outputs.data, y.data) self.cnf.add(outputs.argmax(1), y) # measure elapsed time batch_time.update(time.time() - end) end = time.time() # update self.logger_val.update( {'loss': loss.item()}, dict( zip(self.metrics_name, [ pred[p].item() for p in range(len(self.metrics_name)) ])), batch_size, ) if i % self.print_freq == 0: self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=False, bavg=True, bsummary=False, ) # save validation loss self.vallosses = self.logger_val.info['loss']['loss'].avg acc = self.logger_val.info['metrics']['top1'].avg self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=True, bavg=True, bsummary=True, ) print('Confusion Matriz') print(self.cnf.value(), flush=True) print('\n') self.visheatmap.show('Confusion Matriz', self.cnf.value()) return acc def test(self, data_loader): n = len(data_loader) * data_loader.batch_size Yhat = np.zeros((n, self.num_output_channels)) Y = np.zeros((n, 1)) k = 0 # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, sample in enumerate(tqdm(data_loader)): # get data (image, label) x, y = sample['image'], sample['label'].argmax(1).long() x = x.cuda() if self.cuda else x # fit (forward) yhat = self.net(x) yhat = F.softmax(yhat, dim=1) yhat = pytutils.to_np(yhat) for j in range(yhat.shape[0]): Y[k] = y[j] Yhat[k, :] = yhat[j] k += 1 Yhat = Yhat[:k, :] Y = Y[:k] return Yhat, Y def predict(self, data_loader): n = len(data_loader) * data_loader.batch_size Yhat = np.zeros((n, self.num_output_channels)) Ids = np.zeros((n, 1)) k = 0 # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, (Id, inputs) in enumerate(tqdm(data_loader)): x = inputs.cuda() if self.cuda else inputs # fit (forward) yhat = self.net(x) yhat = F.softmax(yhat, dim=1) yhat = pytutils.to_np(yhat) for j in range(yhat.shape[0]): Yhat[k, :] = yhat[j] Ids[k] = Id[j] k += 1 Yhat = Yhat[:k, :] Ids = Ids[:k] return Ids, Yhat def __call__(self, image): # switch to evaluate mode self.net.eval() with torch.no_grad(): x = image.cuda() if self.cuda else image msoft = nn.Softmax() yhat = msoft(self.net(x)) yhat = pytutils.to_np(yhat) return yhat def representation(self, data_loader): """" Representation -data_loader: simple data loader for image """ # switch to evaluate mode self.net.eval() n = len(data_loader) * data_loader.batch_size k = 0 # embebed features embX = np.zeros([n, self.net.dim]) embY = np.zeros([n, 1]) batch_time = AverageMeter() end = time.time() for i, sample in enumerate(data_loader): # get data (image, label) x, y = sample['image'], sample['label'].argmax(1).long() x = x.cuda() if self.cuda else x # representation emb = self.net.representation(x) emb = pytutils.to_np(emb) for j in range(emb.shape[0]): embX[k, :] = emb[j, :] embY[k] = y[j] k += 1 # measure elapsed time batch_time.update(time.time() - end) end = time.time() print( 'Representation: |{:06d}/{:06d}||{batch_time.val:.3f} ({batch_time.avg:.3f})|' .format(i, len(data_loader), batch_time=batch_time)) embX = embX[:k, :] embY = embY[:k] return embX, embY def _create_model(self, arch, num_output_channels, num_input_channels, pretrained): """ Create model @arch (string): select architecture @num_classes (int) @num_channels (int) @pretrained (bool) """ self.net = None self.size_input = 0 kw = { 'num_classes': num_output_channels, 'num_channels': num_input_channels, 'pretrained': pretrained } self.net = nnmodels.__dict__[arch](**kw) self.s_arch = arch self.size_input = self.net.size_input self.num_output_channels = num_output_channels self.num_input_channels = num_input_channels if self.cuda == True: self.net.cuda() if self.parallel == True and self.cuda == True: self.net = nn.DataParallel(self.net, device_ids=range( torch.cuda.device_count())) def _create_loss(self, loss): # create loss if loss == 'cross': self.criterion = nn.CrossEntropyLoss().cuda() elif loss == 'mse': self.criterion = nn.MSELoss(size_average=True).cuda() elif loss == 'l1': self.criterion = nn.L1Loss(size_average=True).cuda() else: assert (False) self.s_loss = loss
class ClassNeuralNet(NeuralNetAbstract): """ classification Neural Net like preactresnet """ def __init__(self, patchproject, nameproject, no_cuda=True, parallel=False, seed=1, print_freq=10, gpu=0, view_freq=1): """ Initialization -patchproject (str): path project -nameproject (str): name project -no_cuda (bool): system cuda (default is True) -parallel (bool) -seed (int) -print_freq (int) -gpu (int) -view_freq (in epochs) """ super(ClassNeuralNet, self).__init__(patchproject, nameproject, no_cuda, parallel, seed, print_freq, gpu) self.view_freq = view_freq def create( self, arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum=0.9, weight_decay=5e-4, pretrained=False, size_input=388, num_classes=8, backbone='preactresnet', num_filters=32, breal='real', alpha=2, beta=2, ): """ Create -arch (string): architecture -loss (string): -lr (float): learning rate -optimizer (string) : -lrsch (string): scheduler learning rate -pretrained (bool) """ cfg_opt = {'momentum': momentum, 'weight_decay': weight_decay} #cfg_scheduler={ 'step_size':100, 'gamma':0.1 } cfg_scheduler = {'mode': 'min', 'patience': 10} cfg_model = {'num_filters': num_filters} self.num_classes = num_classes super(ClassNeuralNet, self).create( arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, pretrained, cfg_opt=cfg_opt, cfg_scheduler=cfg_scheduler, cfg_model=cfg_model, ) self.size_input = size_input self.backbone = backbone self.num_filters = num_filters self.topk = nloss.TopkAccuracy() self.logger_train = Logger('Train', ['loss', 'loss_bce'], ['topk'], self.plotter) self.logger_val = Logger('Val ', ['loss', 'loss_bce'], ['topk'], self.plotter) self.breal = breal # Set the graphic visualization # self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject, heatsize=(100,100) ) def _create_model(self, arch, num_output_channels, num_input_channels, pretrained, **kwargs): """ Create model -arch (string): select architecture -num_classes (int) -num_channels (int) -pretrained (bool) """ self.net = None #-------------------------------------------------------------------------------------------- # select architecture #-------------------------------------------------------------------------------------------- num_filters = kwargs.get("num_filters", 32) # num_classes=1000, num_channels=3, initial_channels=64 for preactresnet kw = { 'num_classes': self.num_classes, 'num_channels': num_input_channels, 'pretrained': pretrained } print("kw", kw) self.net = nnmodels.__dict__[arch](**kw) self.s_arch = arch self.num_output_channels = num_output_channels self.num_input_channels = num_input_channels self.num_filters = num_filters if self.cuda: self.net.cuda() if self.parallel and self.cuda: self.net = nn.DataParallel(self.net, device_ids=range( torch.cuda.device_count())) def save(self, epoch, prec, is_best=False, filename='checkpoint.pth.tar'): """ Save model """ print('>> save model epoch {} ({}) in {}'.format( epoch, prec, filename)) net = self.net.module if self.parallel else self.net pytutils.save_checkpoint( { 'epoch': epoch + 1, 'arch': self.s_arch, 'imsize': self.size_input, 'num_output_channels': self.num_output_channels, 'num_input_channels': self.num_input_channels, 'num_classes': self.num_classes, 'state_dict': net.state_dict(), 'prec': prec, 'optimizer': self.optimizer.state_dict(), }, is_best, self.pathmodels, filename) def load(self, pathnamemodel): """ load model from pretrained model :param pathnamemodel: model path :return: """ bload = False if pathnamemodel: if os.path.isfile(pathnamemodel): print("=> loading checkpoint '{}'".format(pathnamemodel)) checkpoint = torch.load( pathnamemodel) if self.cuda else torch.load( pathnamemodel, map_location=lambda storage, loc: storage) self.num_classes = checkpoint['num_classes'] self._create_model(checkpoint['arch'], checkpoint['num_output_channels'], checkpoint['num_input_channels'], False) self.size_input = checkpoint['imsize'] self.net.load_state_dict(checkpoint['state_dict']) print("=> loaded checkpoint for {} arch!".format( checkpoint['arch'])) bload = True else: print("=> no checkpoint found at '{}'".format(pathnamemodel)) return bload def training(self, data_loader, epoch=0, *args): #reset logger self.logger_train.reset() data_time = AverageMeter() batch_time = AverageMeter() # switch to evaluate mode self.net.train() end = time.time() for i, sample in enumerate(data_loader): # measure data loading time data_time.update(time.time() - end) # if dataset is real if self.breal == 'real': x_img, y_lab = sample['image'], sample['label'] y_lab = y_lab.argmax(dim=1) if self.cuda: x_img = x_img.cuda() y_lab = y_lab.cuda() x_org = x_img.clone().detach() else: # if dataset is synthetic x_org, x_img, y_mask, meta = sample y_lab = meta[:, 0] y_theta = meta[:, 1:].view(-1, 2, 3) if self.cuda: x_org = x_org.cuda() x_img = x_img.cuda() y_mask = y_mask.cuda() y_lab = y_lab.cuda() y_theta = y_theta.cuda() # fit (forward) y_lab_hat = self.net(x_img) # calculate classification loss loss_bce = self.criterion_bce(y_lab_hat, y_lab.long()) loss = loss_bce # accuracy of choosing top k predicted classes topk = self.topk(y_lab_hat, y_lab.long()) batch_size = x_img.shape[0] # optimizer self.optimizer.zero_grad() (loss * batch_size).backward() self.optimizer.step() # update self.logger_train.update( { 'loss': loss.cpu().item(), 'loss_bce': loss_bce.cpu().item() }, {'topk': topk[0][0].cpu()}, batch_size, ) # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % self.print_freq == 0: self.logger_train.logger( epoch, epoch + float(i + 1) / len(data_loader), i, len(data_loader), batch_time, ) def evaluate(self, data_loader, epoch=0, *args): """ evaluate on validation dataset :param data_loader: which data_loader to use :param epoch: current epoch :return: acc: average accuracy on data_loader """ # reset loader self.logger_val.reset() batch_time = AverageMeter() # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, sample in enumerate(data_loader): # get data (image, label) if self.breal == 'real': x_img, y_lab = sample["image"], sample["label"] y_lab = y_lab.argmax(dim=1) if self.cuda: x_img = x_img.cuda() y_lab = y_lab.cuda() x_org = x_img.clone().detach() else: x_org, x_img, y_mask, meta = sample y_lab = meta[:, 0] y_theta = meta[:, 1:].view(-1, 2, 3) if self.cuda: x_org = x_org.cuda() x_img = x_img.cuda() y_mask = y_mask.cuda() y_lab = y_lab.cuda() y_theta = y_theta.cuda() # fit (forward) # print("x_img size", x_img.size()) y_lab_hat = self.net(x_img) # measure accuracy and record loss loss_bce = self.criterion_bce(y_lab_hat, y_lab.long()) loss = loss_bce topk = self.topk(y_lab_hat, y_lab.long()) batch_size = x_img.shape[0] # measure elapsed time batch_time.update(time.time() - end) end = time.time() # update self.logger_val.update( { 'loss': loss.cpu().item(), 'loss_bce': loss_bce.cpu().item() }, {'topk': topk[0][0].cpu()}, batch_size, ) # print the result in certain print frequency when iterating batches if i % self.print_freq == 0: self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=False, bavg=True, bsummary=False, ) #save validation loss and accuracy self.vallosses = self.logger_val.info['loss']['loss'].avg acc = self.logger_val.info['metrics']['topk'].avg # print the average loss and accuracy after one iteration self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=True, bavg=True, bsummary=True, ) return acc def representation(self, dataloader, breal='real'): """ :param dataloader: :param breal:'real' or 'synthetic' :return: Y_labs: true labels Y_lab_hats: predicted labels """ Y_labs = [] Y_lab_hats = [] self.net.eval() with torch.no_grad(): for i_batch, sample in enumerate(tqdm(dataloader)): if breal == 'real': x_img, y_lab = sample['image'], sample['label'] y_lab = y_lab.argmax(dim=1) else: x_org, x_img, y_mask, y_lab = sample y_lab = y_lab[:, 0] if self.cuda: x_img = x_img.cuda() y_lab_hat = self.net(x_img) Y_labs.append(y_lab) Y_lab_hats.append(y_lab_hat.data.cpu()) Y_labs = np.concatenate(Y_labs, axis=0) Y_lab_hats = np.concatenate(Y_lab_hats, axis=0) return Y_labs, Y_lab_hats def __call__(self, image): # when calling the class, switch to evaluate mode self.net.eval() with torch.no_grad(): x = image.cuda() if self.cuda else image y_lab_hat, att, fmap, srf = self.net(x) y_lab_hat = F.softmax(y_lab_hat, dim=1) return y_lab_hat, att, fmap, srf def _create_loss(self, loss, alpha=None, beta=None): # private method # create cross entropy loss self.criterion_bce = nn.CrossEntropyLoss().cuda() self.s_loss = loss
class NeuralNetClassifier(NeuralNetAbstract): r"""Convolutional Neural Net for classification Args: patchproject (str): path project nameproject (str): name project no_cuda (bool): system cuda (default is True) parallel (bool) seed (int) print_freq (int) gpu (int) """ def __init__(self, patchproject, nameproject, no_cuda=True, parallel=False, seed=1, print_freq=10, gpu=0): super(NeuralNetClassifier, self).__init__(patchproject, nameproject, no_cuda, parallel, seed, print_freq, gpu) def create( self, arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum=0.9, weight_decay=5e-4, pretrained=False, th=0.0, size_input=32, ): """ Create Args: arch (string): architecture num_output_channels, num_input_channels, loss (string): lr (float): learning rate momentum, optimizer (string) : lrsch (string): scheduler learning rate pretrained (bool) """ cfg_opt = {'momentum': 0.9, 'weight_decay': 5e-4} cfg_scheduler = {'step_size': 50, 'gamma': 0.1} super(NeuralNetClassifier, self).create( arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, pretrained, cfg_opt=cfg_opt, cfg_scheduler=cfg_scheduler, ) self.size_input = size_input self.accuracy = nloss.MultAccuracyV1(th) self.f_score = nloss.F_score(threshold=th, beta=2) #self.cnf = nloss.ConfusionMeter( self.num_output_channels, normalized=True ) #self.visheatmap = gph.HeatMapVisdom( env_name=self.nameproject ) # Set the graphic visualization self.logger_train = Logger('Trn', ['loss'], ['acc', 'f1'], self.plotter) self.logger_val = Logger('Val', ['loss'], ['acc', 'f1'], self.plotter) def training(self, data_loader, epoch=0): self.logger_train.reset() data_time = AverageMeter() batch_time = AverageMeter() # switch to evaluate mode self.net.train() end = time.time() for i, (iD, image, prob) in enumerate(data_loader): # measure data loading time data_time.update(time.time() - end) x, y = image, prob batch_size = x.size(0) if self.cuda: x = x.cuda() y = y.cuda() # fit (forward) yhat = self.net(x) # measure accuracy and record loss loss = self.criterion(yhat, y.float()) pred = self.accuracy(yhat.data, y.data) f1 = self.f_score(yhat.data, y.data) # optimizer self.optimizer.zero_grad() (loss).backward() self.optimizer.step() # update self.logger_train.update( {'loss': loss.data[0]}, { 'acc': pred.data[0], 'f1': f1.data[0] }, batch_size, ) # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % self.print_freq == 0: self.logger_train.logger( epoch, epoch + float(i + 1) / len(data_loader), i, len(data_loader), batch_time, ) def evaluate(self, data_loader, epoch=0): self.logger_val.reset() #self.cnf.reset() batch_time = AverageMeter() # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, (iD, image, prob) in enumerate(data_loader): # get data (image, label) x, y = image, prob #.argmax(1).long() batch_size = x.size(0) if self.cuda: x = x.cuda() y = y.cuda() # fit (forward) yhat = self.net(x) # measure accuracy and record loss loss = self.criterion(yhat, y.float()) pred = self.accuracy(yhat.data, y.data) f1 = self.f_score(yhat.data, y.data) #self.cnf.add( outputs.argmax(1), y ) # measure elapsed time batch_time.update(time.time() - end) end = time.time() # update self.logger_val.update( {'loss': loss.data[0]}, { 'acc': pred.data[0], 'f1': f1.data[0] }, batch_size, ) if i % self.print_freq == 0: self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=False, bavg=True, bsummary=False, ) #save validation loss self.vallosses = self.logger_val.info['loss']['loss'].avg acc = self.logger_val.info['metrics']['acc'].avg self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=True, bavg=True, bsummary=True, ) #print('Confusion Matriz') #print(self.cnf.value(), flush=True) #print('\n') #self.visheatmap.show('Confusion Matriz', self.cnf.value()) return acc def predict(self, data_loader): Yhat, iDs, Y = [], [], [] # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, (iD, image, prob) in enumerate(tqdm(data_loader)): x = image.cuda() if self.cuda else image yhat = self.net(x) yhat = F.sigmoid(yhat).cpu().numpy() Yhat.append(yhat) iDs.append(iD) Y.append(prob) Yhat = np.concatenate(Yhat, axis=0) iDs = np.concatenate(iDs, axis=0) Y = np.concatenate(Y, axis=0) return iDs, Yhat, Y def __call__(self, image): # switch to evaluate mode self.net.eval() with torch.no_grad(): x = image.cuda() if self.cuda else image yhat = self.net(x) yhat = F.sigmoid(yhat).cpu().numpy() return yhat def _create_model(self, arch, num_output_channels, num_input_channels, pretrained): """ Create model @arch (string): select architecture @num_classes (int) @num_channels (int) @pretrained (bool) """ self.net = None self.size_input = 0 kw = { 'num_classes': num_output_channels, 'num_channels': num_input_channels, 'pretrained': pretrained } self.net = nnmodels.__dict__[arch](**kw) self.s_arch = arch self.size_input = self.net.size_input self.num_output_channels = num_output_channels self.num_input_channels = num_input_channels if self.cuda == True: self.net.cuda() if self.parallel == True and self.cuda == True: self.net = nn.DataParallel(self.net, device_ids=range( torch.cuda.device_count())) def _create_loss(self, loss): # create loss if loss == 'cross': self.criterion = nn.CrossEntropyLoss().cuda() elif loss == 'mse': self.criterion = nn.MSELoss(size_average=True).cuda() elif loss == 'bcewl': self.criterion = nn.BCEWithLogitsLoss(size_average=True).cuda() elif loss == 'l1': self.criterion = nn.L1Loss(size_average=True).cuda() elif loss == 'focal': self.criterion = nloss.FocalLoss(gamma=2).cuda() elif loss == 'dice': self.criterion = nloss.DiceLoss().cuda() elif loss == 'mix': self.criterion = nloss.MixLoss().cuda() else: assert (False) self.s_loss = loss
class NeuralNetClassifier(NeuralNetAbstract): """ Convolutional Neural Net """ def __init__(self, patchproject, nameproject, no_cuda=True, parallel=False, seed=1, print_freq=10, gpu=0): """ Initialization Args: @patchproject (str): path project @nameproject (str): name project @no_cuda (bool): system cuda (default is True) @parallel (bool) @seed (int) @print_freq (int) @gpu (int) """ super(NeuralNetClassifier, self).__init__(patchproject, nameproject, no_cuda, parallel, seed, print_freq, gpu) def create( self, arch, num_output_channels, num_input_channels, loss, lr, momentum, optimizer, lrsch, pretrained=False, topk=(1, ), ): """ Create Args: @arch (string): architecture @num_output_channels, @num_input_channels, @loss (string): @lr (float): learning rate @momentum, @optimizer (string) : @lrsch (string): scheduler learning rate @pretrained (bool) """ super(NeuralNetClassifier, self).create(arch, num_output_channels, num_input_channels, loss, lr, momentum, optimizer, lrsch, pretrained) self.accuracy = nloss.Accuracy(topk) self.cnf = nloss.ConfusionMeter(self.num_output_channels, normalized=True) self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject) # Set the graphic visualization self.metrics_name = ['top{}'.format(k) for k in topk] self.logger_train = Logger('Trn', ['loss'], self.metrics_name, self.plotter) self.logger_val = Logger('Val', ['loss'], self.metrics_name, self.plotter) def training(self, data_loader, epoch=0): self.logger_train.reset() data_time = AverageMeter() batch_time = AverageMeter() # switch to evaluate mode self.net.train() end = time.time() for i, sample in enumerate(data_loader): # measure data loading time data_time.update(time.time() - end) # get data (image, label) inputs, targets = sample['image'], pytutils.argmax(sample['label']) batch_size = inputs.size(0) if self.cuda: targets = targets.cuda(non_blocking=True) inputs_var = Variable(inputs.cuda(), requires_grad=False) targets_var = Variable(targets.cuda(), requires_grad=False) else: inputs_var = Variable(inputs, requires_grad=False) targets_var = Variable(targets, requires_grad=False) # fit (forward) outputs = self.net(inputs_var) # measure accuracy and record loss loss = self.criterion(outputs, targets_var) pred = self.accuracy(outputs.data, targets) # optimizer self.optimizer.zero_grad() loss.backward() self.optimizer.step() # update self.logger_train.update( {'loss': loss.data[0]}, dict( zip(self.metrics_name, [pred[p][0] for p in range(len(self.metrics_name))])), batch_size, ) # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % self.print_freq == 0: self.logger_train.logger( epoch, epoch + float(i + 1) / len(data_loader), i, len(data_loader), batch_time, ) def evaluate(self, data_loader, epoch=0): self.logger_val.reset() self.cnf.reset() batch_time = AverageMeter() # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, sample in enumerate(data_loader): # get data (image, label) inputs, targets = sample['image'], pytutils.argmax( sample['label']) batch_size = inputs.size(0) if self.cuda: targets = targets.cuda(non_blocking=True) inputs_var = Variable(inputs.cuda(), requires_grad=False, volatile=True) targets_var = Variable(targets.cuda(), requires_grad=False, volatile=True) else: inputs_var = Variable(inputs, requires_grad=False, volatile=True) targets_var = Variable(targets, requires_grad=False, volatile=True) # fit (forward) outputs = self.net(inputs_var) # measure accuracy and record loss loss = self.criterion(outputs, targets_var) pred = self.accuracy(outputs.data, targets) self.cnf.add(outputs.argmax(1), targets_var) # measure elapsed time batch_time.update(time.time() - end) end = time.time() # update self.logger_val.update( {'loss': loss.data[0]}, dict( zip(self.metrics_name, [ pred[p][0] for p in range(len(self.metrics_name)) ])), batch_size, ) if i % self.print_freq == 0: self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=False, bavg=True, bsummary=False, ) #save validation loss self.vallosses = self.logger_val.info['loss']['loss'].avg acc = self.logger_val.info['metrics']['top1'].avg self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=True, bavg=True, bsummary=True, ) print('Confusion Matriz') print(self.cnf.value(), flush=True) print('\n') self.visheatmap.show('Confusion Matriz', self.cnf.value()) return acc #def _to_end_epoch(self, epoch, epochs, train_loader, val_loader): #print('>> Reset', flush=True ) #w = 1-self.cnf.value().diagonal() #train_loader.dataset.reset(w) def test(self, data_loader): n = len(data_loader) * data_loader.batch_size Yhat = np.zeros((n, self.num_output_channels)) Y = np.zeros((n, 1)) k = 0 # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, sample in enumerate(tqdm(data_loader)): # get data (image, label) inputs = sample['image'] targets = pytutils.argmax(sample['label']) x = inputs.cuda() if self.cuda else inputs x = Variable(x, requires_grad=False, volatile=True) # fit (forward) yhat = self.net(x) yhat = F.softmax(yhat, dim=1) yhat = pytutils.to_np(yhat) for j in range(yhat.shape[0]): Y[k] = targets[j] Yhat[k, :] = yhat[j] k += 1 #print( 'Test:', i , flush=True ) Yhat = Yhat[:k, :] Y = Y[:k] return Yhat, Y def predict(self, data_loader): n = len(data_loader) * data_loader.batch_size Yhat = np.zeros((n, self.num_output_channels)) Ids = np.zeros((n, 1)) k = 0 # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, (Id, inputs) in enumerate(tqdm(data_loader)): # get data (image, label) #inputs = sample['image'] #Id = sample['id'] x = inputs.cuda() if self.cuda else inputs x = Variable(x, requires_grad=False, volatile=True) # fit (forward) yhat = self.net(x) yhat = F.softmax(yhat, dim=1) yhat = pytutils.to_np(yhat) for j in range(yhat.shape[0]): Yhat[k, :] = yhat[j] Ids[k] = Id[j] k += 1 Yhat = Yhat[:k, :] Ids = Ids[:k] return Ids, Yhat def __call__(self, image): # switch to evaluate mode self.net.eval() with torch.no_grad(): x = image.cuda() if self.cuda else image x = Variable(x, requires_grad=False, volatile=True) msoft = nn.Softmax() yhat = msoft(self.net(x)) yhat = pytutils.to_np(yhat) return yhat def representation(self, data_loader): """" Representation -data_loader: simple data loader for image """ # switch to evaluate mode self.net.eval() n = len(data_loader) * data_loader.batch_size k = 0 # embebed features embX = np.zeros([n, self.net.dim]) embY = np.zeros([n, 1]) batch_time = AverageMeter() end = time.time() for i, sample in enumerate(data_loader): # get data (image, label) inputs, targets = sample['image'], pytutils.argmax(sample['label']) inputs_var = pytutils.to_var(inputs, self.cuda, False, True) # representation emb = self.net.representation(inputs_var) emb = pytutils.to_np(emb) for j in range(emb.shape[0]): embX[k, :] = emb[j, :] embY[k] = targets[j] k += 1 # measure elapsed time batch_time.update(time.time() - end) end = time.time() print( 'Representation: |{:06d}/{:06d}||{batch_time.val:.3f} ({batch_time.avg:.3f})|' .format(i, len(data_loader), batch_time=batch_time)) embX = embX[:k, :] embY = embY[:k] return embX, embY def _create_model(self, arch, num_output_channels, num_input_channels, pretrained): """ Create model @arch (string): select architecture @num_classes (int) @num_channels (int) @pretrained (bool) """ self.net = None self.size_input = 0 kw = { 'num_classes': num_output_channels, 'num_channels': num_input_channels, 'pretrained': pretrained } self.net = nnmodels.__dict__[arch](**kw) self.s_arch = arch self.size_input = self.net.size_input self.num_output_channels = num_output_channels self.num_input_channels = num_input_channels if self.cuda == True: self.net.cuda() if self.parallel == True and self.cuda == True: self.net = nn.DataParallel(self.net, device_ids=range( torch.cuda.device_count())) def _create_loss(self, loss): # create loss if loss == 'cross': self.criterion = nn.CrossEntropyLoss().cuda() elif loss == 'mse': self.criterion = nn.MSELoss(size_average=True).cuda() elif loss == 'l1': self.criterion = nn.L1Loss(size_average=True).cuda() else: assert (False) self.s_loss = loss
class SegmentationNeuralNet(NeuralNetAbstract): """ Segmentation Neural Net """ def __init__(self, patchproject, nameproject, no_cuda=True, parallel=False, seed=1, print_freq=10, gpu=0, view_freq=1): """ Initialization -patchproject (str): path project -nameproject (str): name project -no_cuda (bool): system cuda (default is True) -parallel (bool) -seed (int) -print_freq (int) -gpu (int) -view_freq (in epochs) """ super(SegmentationNeuralNet, self).__init__(patchproject, nameproject, no_cuda, parallel, seed, print_freq, gpu) self.view_freq = view_freq def create( self, arch, num_output_channels, num_input_channels, loss, lr, momentum, optimizer, lrsch, pretrained=False, size_input=388, ): """ Create -arch (string): architecture -loss (string): -lr (float): learning rate -optimizer (string) : -lrsch (string): scheduler learning rate -pretrained (bool) """ super(SegmentationNeuralNet, self).create(arch, num_output_channels, num_input_channels, loss, lr, momentum, optimizer, lrsch, pretrained) self.size_input = size_input self.accuracy = nloss.Accuracy() self.dice = nloss.Dice() # Set the graphic visualization self.logger_train = Logger('Train', ['loss'], ['accs', 'dices'], self.plotter) self.logger_val = Logger('Val ', ['loss'], ['accs', 'dices'], self.plotter) self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject, heatsize=(100, 100)) self.visimshow = gph.ImageVisdom(env_name=self.nameproject, imsize=(100, 100)) def training(self, data_loader, epoch=0): #reset logger self.logger_train.reset() data_time = AverageMeter() batch_time = AverageMeter() # switch to evaluate mode self.net.train() end = time.time() for i, sample in enumerate(data_loader): # measure data loading time data_time.update(time.time() - end) # get data (image, label, weight) inputs, targets, weights = sample['image'], sample[ 'label'], sample['weight'] batch_size = inputs.size(0) if self.cuda: targets = targets.cuda(non_blocking=True) inputs_var = Variable(inputs.cuda(), requires_grad=False) targets_var = Variable(targets.cuda(), requires_grad=False) weights_var = Variable(weights.cuda(), requires_grad=False) else: inputs_var = Variable(inputs, requires_grad=False) targets_var = Variable(targets, requires_grad=False) weights_var = Variable(weights, requires_grad=False) # fit (forward) outputs = self.net(inputs_var) # measure accuracy and record loss loss = self.criterion(outputs, targets_var, weights_var) accs = self.accuracy(outputs, targets_var) dices = self.dice(outputs, targets_var) # optimizer self.optimizer.zero_grad() (loss * batch_size).backward() self.optimizer.step() # update self.logger_train.update( {'loss': loss.data[0]}, { 'accs': accs, 'dices': dices }, batch_size, ) # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % self.print_freq == 0: self.logger_train.logger( epoch, epoch + float(i + 1) / len(data_loader), i, len(data_loader), batch_time, ) def evaluate(self, data_loader, epoch=0): # reset loader self.logger_val.reset() batch_time = AverageMeter() # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, sample in enumerate(data_loader): # get data (image, label) inputs, targets, weights = sample['image'], sample[ 'label'], sample['weight'] batch_size = inputs.size(0) if self.cuda: targets = targets.cuda(non_blocking=True) inputs_var = Variable(inputs.cuda(), requires_grad=False, volatile=True) targets_var = Variable(targets.cuda(), requires_grad=False, volatile=True) weights_var = Variable(weights.cuda(), requires_grad=False, volatile=True) else: inputs_var = Variable(inputs, requires_grad=False, volatile=True) targets_var = Variable(targets, requires_grad=False, volatile=True) weights_var = Variable(weights, requires_grad=False, volatile=True) # fit (forward) outputs = self.net(inputs_var) # measure accuracy and record loss loss = self.criterion(outputs, targets_var, weights_var) accs = self.accuracy(outputs, targets_var) dices = self.dice(outputs, targets_var) # measure elapsed time batch_time.update(time.time() - end) end = time.time() # update self.logger_val.update( {'loss': loss.data[0]}, { 'accs': accs, 'dices': dices }, batch_size, ) if i % self.print_freq == 0: self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=False, bavg=True, bsummary=False, ) #save validation loss self.vallosses = self.logger_val.info['loss']['loss'].avg acc = self.logger_val.info['metrics']['accs'].avg self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=True, bavg=True, bsummary=True, ) #vizual_freq if epoch % self.view_freq == 0: prob = F.softmax(outputs, dim=1) prob = prob.data[0] _, maxprob = torch.max(prob, 0) self.visheatmap.show('Label', targets_var.data.cpu()[0].numpy()[1, :, :]) self.visheatmap.show('Weight map', weights_var.data.cpu()[0].numpy()[0, :, :]) self.visheatmap.show('Image', inputs_var.data.cpu()[0].numpy()[0, :, :]) self.visheatmap.show('Max prob', maxprob.cpu().numpy().astype(np.float32)) for k in range(prob.shape[0]): self.visheatmap.show('Heat map {}'.format(k), prob.cpu()[k].numpy()) return acc def test(self, data_loader): masks = [] ids = [] k = 0 # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, sample in enumerate(tqdm(data_loader)): # get data (image, label) inputs, meta = sample['image'], sample['metadata'] idd = meta[:, 0] x = inputs.cuda() if self.cuda else inputs # fit (forward) yhat = self.net(x) yhat = F.softmax(yhat, dim=1) yhat = pytutils.to_np(yhat) masks.append(yhat) ids.append(idd) return ids, masks def __call__(self, image): # switch to evaluate mode self.net.eval() with torch.no_grad(): x = image.cuda() if self.cuda else image yhat = F.softmax(self.net(x), dim=1) yhat = pytutils.to_np(yhat).transpose(2, 3, 1, 0)[..., 0] return yhat def _create_model(self, arch, num_output_channels, num_input_channels, pretrained): """ Create model -arch (string): select architecture -num_classes (int) -num_channels (int) -pretrained (bool) """ self.net = None #-------------------------------------------------------------------------------------------- # select architecture #-------------------------------------------------------------------------------------------- #kw = {'num_classes': num_output_channels, 'num_channels': num_input_channels, 'pretrained': pretrained} kw = { 'num_classes': num_output_channels, 'in_channels': num_input_channels, 'pretrained': pretrained } self.net = nnmodels.__dict__[arch](**kw) self.s_arch = arch self.num_output_channels = num_output_channels self.num_input_channels = num_input_channels if self.cuda == True: self.net.cuda() if self.parallel == True and self.cuda == True: self.net = nn.DataParallel(self.net, device_ids=range( torch.cuda.device_count())) def _create_loss(self, loss): # create loss if loss == 'wmce': self.criterion = nloss.WeightedMCEloss() elif loss == 'bdice': self.criterion = nloss.BDiceLoss() elif loss == 'wbdice': self.criterion = nloss.WeightedBDiceLoss() elif loss == 'wmcedice': self.criterion = nloss.WeightedMCEDiceLoss() elif loss == 'wfocalmce': self.criterion = nloss.WeightedMCEFocalloss() elif loss == 'mcedice': self.criterion = nloss.MCEDiceLoss() else: assert (False) self.s_loss = loss
class AttentionGMMNeuralNet(AttentionNeuralNetAbstract): """ Attention Neural Net and GMM representation Args: -patchproject (str): path project -nameproject (str): name project -no_cuda (bool): system cuda (default is True) -parallel (bool) -seed (int) -print_freq (int) -gpu (int) -view_freq (in epochs) """ def __init__(self, patchproject, nameproject, no_cuda=True, parallel=False, seed=1, print_freq=10, gpu=0, view_freq=1 ): super(AttentionGMMNeuralNet, self).__init__( patchproject, nameproject, no_cuda, parallel, seed, print_freq, gpu, view_freq ) def create(self, arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum=0.9, weight_decay=5e-4, pretrained=False, size_input=388, num_classes=8, ): """ Create Args: -arch (string): architecture -num_output_channels, -num_input_channels, -loss (string): -lr (float): learning rate -optimizer (string) : -lrsch (string): scheduler learning rate -pretrained (bool) - """ super(AttentionGMMNeuralNet, self).create( arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum, weight_decay, pretrained, size_input, num_classes, ) self.logger_train = Logger( 'Train', ['loss', 'loss_gmm', 'loss_bce', 'loss_att' ], [ 'topk', 'gmm'], self.plotter ) self.logger_val = Logger( 'Val ', ['loss', 'loss_gmm', 'loss_bce', 'loss_att' ], [ 'topk', 'gmm'], self.plotter ) def training(self, data_loader, epoch=0): #reset logger self.logger_train.reset() data_time = AverageMeter() batch_time = AverageMeter() # switch to evaluate mode self.net.train() end = time.time() for i, (x_org, x_img, y_mask, meta ) in enumerate(data_loader): # measure data loading time data_time.update(time.time() - end) batch_size = x_img.shape[0] y_lab = meta[:,0] y_theta = meta[:,1:].view(-1, 2, 3) if self.cuda: x_org = x_org.cuda() x_img = x_img.cuda() y_mask = y_mask.cuda() y_lab = y_lab.cuda() y_theta = y_theta.cuda() # fit (forward) z, y_lab_hat, att, _, _ = self.net( x_img, x_img*y_mask[:,1,...].unsqueeze(dim=1) ) # measure accuracy and record loss loss_bce = self.criterion_bce( y_lab_hat, y_lab.long() ) loss_gmm = self.criterion_gmm( z, y_lab ) loss_att = self.criterion_att( x_org, y_mask, att ) loss = loss_bce + loss_gmm + loss_att topk = self.topk( y_lab_hat, y_lab.long() ) gmm = self.gmm( z, y_lab ) # optimizer self.optimizer.zero_grad() (loss).backward() #batch_size self.optimizer.step() # update self.logger_train.update( {'loss': loss.cpu().item(), 'loss_gmm': loss_gmm.cpu().item(), 'loss_bce': loss_bce.cpu().item(), 'loss_att':loss_att.cpu().item() }, {'topk': topk[0][0].cpu(), 'gmm': gmm.cpu().item() }, batch_size, ) # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % self.print_freq == 0: self.logger_train.logger( epoch, epoch + float(i+1)/len(data_loader), i, len(data_loader), batch_time, ) def evaluate(self, data_loader, epoch=0): # reset loader self.logger_val.reset() batch_time = AverageMeter() # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, (x_org, x_img, y_mask, meta) in enumerate(data_loader): # get data (image, label) batch_size = x_img.shape[0] y_lab = meta[:,0] y_theta = meta[:,1:].view(-1, 2, 3) if self.cuda: x_org = x_org.cuda() x_img = x_img.cuda() y_mask = y_mask.cuda() y_lab = y_lab.cuda() y_theta = y_theta.cuda() # fit (forward) z, y_lab_hat, att, fmap, srf = self.net( x_img ) # measure accuracy and record loss loss_bce = self.criterion_bce( y_lab_hat, y_lab.long() ) loss_gmm = self.criterion_gmm( z, y_lab ) loss_att = self.criterion_att( x_org, y_mask, att ) loss = loss_bce + loss_gmm + loss_att topk = self.topk( y_lab_hat, y_lab.long() ) gmm = self.gmm( z, y_lab ) # measure elapsed time batch_time.update(time.time() - end) end = time.time() # update self.logger_val.update( {'loss': loss.cpu().item(), 'loss_gmm': loss_gmm.cpu().item(), 'loss_bce': loss_bce.cpu().item(), 'loss_att':loss_att.cpu().item() }, {'topk': topk[0][0].cpu(), 'gmm': gmm.cpu().item() }, batch_size, ) if i % self.print_freq == 0: self.logger_val.logger( epoch, epoch, i,len(data_loader), batch_time, bplotter=False, bavg=True, bsummary=False, ) #save validation loss self.vallosses = self.logger_val.info['loss']['loss'].avg acc = self.logger_val.info['metrics']['topk'].avg self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=True, bavg=True, bsummary=True, ) #vizual_freq if epoch % self.view_freq == 0: att = att[0,:,:,:].permute( 1,2,0 ).mean(dim=2) srf = srf[0,:,:,:].permute( 1,2,0 ).sum(dim=2) fmap = fmap[0,:,:,:].permute( 1,2,0 ) self.visheatmap.show('Image', x_img.data.cpu()[0].numpy()[0,:,:]) self.visheatmap.show('Image Attention',att.cpu().numpy().astype(np.float32) ) self.visheatmap.show('Feature Map',srf.cpu().numpy().astype(np.float32) ) self.visheatmap.show('Attention Map',fmap.cpu().numpy().astype(np.float32) ) return acc def representation( self, dataloader, breal=True ): Y_labs = [] Y_lab_hats = [] Zs = [] self.net.eval() with torch.no_grad(): for i_batch, sample in enumerate( tqdm(dataloader) ): if breal: x_img, y_lab = sample['image'], sample['label'] y_lab = y_lab.argmax(dim=1) else: x_org, x_img, y_mask, y_lab = sample y_lab=y_lab[:,0] if self.cuda: x_img = x_img.cuda() z, y_lab_hat, _,_,_ = self.net( x_img ) Y_labs.append(y_lab) Y_lab_hats.append(y_lab_hat.data.cpu()) Zs.append(z.data.cpu()) Y_labs = np.concatenate( Y_labs, axis=0 ) Y_lab_hats = np.concatenate( Y_lab_hats, axis=0 ) Zs = np.concatenate( Zs, axis=0 ) return Y_labs, Y_lab_hats, Zs def __call__(self, image ): # switch to evaluate mode self.net.eval() with torch.no_grad(): x = image.cuda() if self.cuda else image z, y_lab_hat, att, fmap, srf = self.net(x) y_lab_hat = F.softmax( y_lab_hat, dim=1 ) return z, y_lab_hat, att, fmap, srf def _create_loss(self, loss): # create loss if loss == 'attloss': self.criterion_bce = nn.CrossEntropyLoss().cuda() self.criterion_att = nloss.Attloss() self.criterion_gmm = nloss.DGMMLoss( self.num_classes, cuda=self.cuda ) else: assert(False) self.s_loss = loss
def create(self, arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum=0.9, weight_decay=5e-4, pretrained=False, size_input=388, cascade_type='none'): """ Create Args: -arch (string): architecture -num_output_channels, -num_input_channels, -loss (string): -lr (float): learning rate -optimizer (string) : -lrsch (string): scheduler learning rate -pretrained (bool) """ cfg_opt = {'momentum': momentum, 'weight_decay': weight_decay} cfg_scheduler = {'step_size': 100, 'gamma': 0.1} super(SegmentationNeuralNet, self).create(arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, pretrained, cfg_opt=cfg_opt, cfg_scheduler=cfg_scheduler) self.size_input = size_input self.num_output_channels = num_output_channels self.cascade_type = cascade_type self.segs_per_forward = 7 if self.cascade_type == 'none': self.step = self.default_step elif self.cascade_type == 'ransac': self.step = self.ransac_step elif self.cascade_type == 'ransac2': self.step = self.ransac_step2 elif self.cascade_type == 'simple': self.step = self.cascate_step else: raise "Cascada not found" self.accuracy = nloss.Accuracy() if num_output_channels == 2: dice_dim = (1, ) if num_output_channels == 4: dice_dim = (1, 2, 3) self.dice = nloss.Dice(dice_dim) # Set the graphic visualization self.logger_train = Logger('Train', ['loss'], ['accs', 'dices'], self.plotter) self.logger_val = Logger('Val ', ['loss'], ['accs', 'dices', 'PQ'], self.plotter) self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject, heatsize=(256, 256)) self.visimshow = gph.ImageVisdom(env_name=self.nameproject, imsize=(256, 256)) if self.half_precision: self.scaler = torch.cuda.amp.GradScaler()
class SegmentationNeuralNet(NeuralNetAbstract): """ Segmentation Neural Net """ def __init__(self, patchproject, nameproject, no_cuda=True, parallel=False, seed=1, print_freq=10, gpu=0, view_freq=1, half_precision=False): """ Initialization -patchproject (str): path project -nameproject (str): name project -no_cuda (bool): system cuda (default is True) -parallel (bool) -seed (int) -print_freq (int) -gpu (int) -view_freq (in epochs) """ super(SegmentationNeuralNet, self).__init__(patchproject, nameproject, no_cuda, parallel, seed, print_freq, gpu) self.view_freq = view_freq self.half_precision = half_precision def create(self, arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, momentum=0.9, weight_decay=5e-4, pretrained=False, size_input=388, cascade_type='none'): """ Create Args: -arch (string): architecture -num_output_channels, -num_input_channels, -loss (string): -lr (float): learning rate -optimizer (string) : -lrsch (string): scheduler learning rate -pretrained (bool) """ cfg_opt = {'momentum': momentum, 'weight_decay': weight_decay} cfg_scheduler = {'step_size': 100, 'gamma': 0.1} super(SegmentationNeuralNet, self).create(arch, num_output_channels, num_input_channels, loss, lr, optimizer, lrsch, pretrained, cfg_opt=cfg_opt, cfg_scheduler=cfg_scheduler) self.size_input = size_input self.num_output_channels = num_output_channels self.cascade_type = cascade_type self.segs_per_forward = 7 if self.cascade_type == 'none': self.step = self.default_step elif self.cascade_type == 'ransac': self.step = self.ransac_step elif self.cascade_type == 'ransac2': self.step = self.ransac_step2 elif self.cascade_type == 'simple': self.step = self.cascate_step else: raise "Cascada not found" self.accuracy = nloss.Accuracy() if num_output_channels == 2: dice_dim = (1, ) if num_output_channels == 4: dice_dim = (1, 2, 3) self.dice = nloss.Dice(dice_dim) # Set the graphic visualization self.logger_train = Logger('Train', ['loss'], ['accs', 'dices'], self.plotter) self.logger_val = Logger('Val ', ['loss'], ['accs', 'dices', 'PQ'], self.plotter) self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject, heatsize=(256, 256)) self.visimshow = gph.ImageVisdom(env_name=self.nameproject, imsize=(256, 256)) if self.half_precision: self.scaler = torch.cuda.amp.GradScaler() def ransac_step(self, inputs, targets, max_deep=4, segs_per_forward=20, src_c=3, verbose=False): srcs = inputs[:, :src_c] segs = inputs[:, src_c:] lv_segs = segs #.clone() first = True final_loss = 0.0 for lv in range(max_deep): n_segs = segs.shape[1] new_segs = [] actual_c = self.segs_per_forward**(max_deep - lv) if verbose: print(segs.shape, actual_c) actual_seg_ids = np.random.choice(range(n_segs), size=actual_c) step_segs = segs[:, actual_seg_ids] for idx in range(0, actual_c, self.segs_per_forward): mini_inp = torch.cat( (srcs, step_segs[:, idx:idx + self.segs_per_forward]), dim=1) mini_out = self.net(mini_inp) ## calculate loss final_loss += self.criterion(mini_out, targets) * 1 new_segs.append(mini_out.argmax(1, keepdim=True)) if verbose: print(mini_inp.shape, idx, idx + self.segs_per_forward, actual_loss.item()) segs = torch.cat(new_segs, dim=1) return final_loss, mini_out def ransac_step2(self, inputs, targets, max_deep=4, n_times=10, segs_per_forward=20, src_c=3, verbose=False): srcs = inputs[:, :src_c] segs = inputs[:, src_c:] first = True final_loss = 0.0 for lv in range(n_times): n_segs = segs.shape[1] actual_seg_ids = np.random.choice(range(n_segs), size=self.segs_per_forward) step_segs = segs[:, actual_seg_ids] mini_inp = torch.cat((srcs, step_segs), dim=1) mini_out = self.net(mini_inp) ## calculate loss final_loss += self.criterion(mini_out, targets) * 1 segs = torch.cat((segs, mini_out.argmax(1, keepdim=True)), dim=1) return final_loss, mini_out def cascate_step(self, inputs, targets, segs_per_forward=20, src_c=3, verbose=False): srcs = inputs[:, :src_c] segs = inputs[:, src_c:] lv_segs = segs.clone() final_loss = 0.0 n_segs = lv_segs.shape[1] actual_seg_ids = np.random.choice(range(n_segs), size=n_segs, replace=False) lv_segs = lv_segs[:, actual_seg_ids] while n_segs > 1: if verbose: print(n_segs) inputs_seg = lv_segs[:, :self.segs_per_forward] inputs_seg_ids = np.random.choice( range(inputs_seg.shape[1]), size=self.segs_per_forward, replace=inputs_seg.shape[1] < self.segs_per_forward) inputs_seg = inputs_seg[:, inputs_seg_ids] mini_inp = torch.cat((srcs, inputs_seg), dim=1) mini_out = self.net(mini_inp) ## calculate loss final_loss += self.criterion(mini_out, targets) if verbose: print(mini_inp.shape, self.segs_per_forward, actual_loss.item()) lv_segs = torch.cat((lv_segs[:, self.segs_per_forward:], mini_out.argmax(1, keepdim=True)), dim=1) n_segs = lv_segs.shape[1] return final_loss, mini_out def default_step(self, inputs, targets): outputs = self.net(inputs) loss = self.criterion(outputs, targets) return loss, outputs def training(self, data_loader, epoch=0): #reset logger self.logger_train.reset() data_time = AverageMeter() batch_time = AverageMeter() # switch to evaluate mode self.net.train() end = time.time() for i, sample in enumerate(data_loader): # measure data loading time data_time.update(time.time() - end) # get data (image, label, weight) inputs, targets = sample['image'], sample['label'] weights = None if 'weight' in sample.keys(): weights = sample['weight'] batch_size = inputs.shape[0] if self.cuda: inputs = inputs.cuda() targets = targets.cuda() if type(weights) is not type(None): weights = weights.cuda() # fit (forward) if self.half_precision: with torch.cuda.amp.autocast(): loss, outputs = self.step(inputs, targets) self.optimizer.zero_grad() self.scaler.scale(loss * batch_size).backward() self.scaler.step(self.optimizer) self.scaler.update() else: loss, outputs = self.step(inputs, targets) self.optimizer.zero_grad() (batch_size * loss).backward() #batch_size self.optimizer.step() accs = self.accuracy(outputs, targets) dices = self.dice(outputs, targets) #pq = metrics.pq_metric(outputs.cpu().detach().numpy(), targets.cpu().detach().numpy()) #pq, n_cells = metrics.pq_metric(targets, outputs) # update self.logger_train.update( {'loss': loss.item()}, { 'accs': accs.item(), #'PQ': pq, 'dices': dices.item() }, batch_size, ) # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % self.print_freq == 0: self.logger_train.logger( epoch, epoch + float(i + 1) / len(data_loader), i, len(data_loader), batch_time, ) def evaluate(self, data_loader, epoch=0): pq_sum = 0 total_cells = 0 self.logger_val.reset() batch_time = AverageMeter() # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, sample in enumerate(data_loader): # get data (image, label) inputs, targets = sample['image'], sample['label'] weights = None if 'weight' in sample.keys(): weights = sample['weight'] #inputs, targets = sample['image'], sample['label'] batch_size = inputs.shape[0] #print(inputs.shape) if self.cuda: inputs = inputs.cuda() targets = targets.cuda() if type(weights) is not type(None): weights = weights.cuda() #print(inputs.shape) # fit (forward) if self.half_precision: with torch.cuda.amp.autocast(): loss, outputs = self.step(inputs, targets) else: loss, outputs = self.step(inputs, targets) # measure accuracy and record loss accs = self.accuracy(outputs, targets) dices = self.dice(outputs, targets) #targets_np = targets[0][1].cpu().numpy().astype(int) if epoch == 0: pq = 0 n_cells = 1 else: #pq, n_cells = metrics.pq_metric(targets, outputs) if False: #self.skip_background: out_shape = outputs.shape zeros = torch.zeros((out_shape[0], 1, out_shape[2], out_shape[3])).cuda() outputs = torch.cat([zeros, outputs], 1) all_metrics, n_cells, _ = metrics.get_metrics( targets, outputs) pq = all_metrics['pq'] pq_sum += pq * n_cells total_cells += n_cells # measure elapsed time batch_time.update(time.time() - end) end = time.time() # update #print(loss.item(), accs, dices, batch_size) self.logger_val.update( {'loss': loss.item()}, { 'accs': accs.item(), 'PQ': (pq_sum / total_cells) if total_cells > 0 else 0, 'dices': dices.item() }, batch_size, ) if i % self.print_freq == 0: self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=False, bavg=True, bsummary=False, ) #save validation loss if total_cells == 0: pq_weight = 0 else: pq_weight = pq_sum / total_cells print(f"PQ: {pq_weight:0.4f}, {pq_sum:0.4f}, {total_cells}") self.vallosses = self.logger_val.info['loss']['loss'].avg acc = self.logger_val.info['metrics']['accs'].avg #pq = pq_weight self.logger_val.logger( epoch, epoch, i, len(data_loader), batch_time, bplotter=True, bavg=True, bsummary=True, ) #vizual_freq if epoch % self.view_freq == 0: prob = F.softmax(outputs.cpu().float(), dim=1) prob = prob.data[0] maxprob = torch.argmax(prob, 0) self.visheatmap.show('Label', targets.data.cpu()[0].numpy()[1, :, :]) #self.visheatmap.show('Weight map', weights.data.cpu()[0].numpy()[0,:,:]) self.visheatmap.show('Image', inputs.data.cpu()[0].numpy()[0, :, :]) self.visheatmap.show('Max prob', maxprob.cpu().numpy().astype(np.float32)) for k in range(prob.shape[0]): self.visheatmap.show('Heat map {}'.format(k), prob.cpu()[k].numpy()) print(f"End Val: wPQ{pq_weight}") return pq_weight def test(self, data_loader): masks = [] ids = [] k = 0 # switch to evaluate mode self.net.eval() with torch.no_grad(): end = time.time() for i, sample in enumerate(tqdm(data_loader)): # get data (image, label) inputs, meta = sample['image'], sample['metadata'] idd = meta[:, 0] x = inputs.cuda() if self.cuda else inputs # fit (forward) yhat = self.net(x) yhat = F.softmax(yhat, dim=1) yhat = pytutils.to_np(yhat) masks.append(yhat) ids.append(idd) return ids, masks def __call__(self, image): # switch to evaluate mode self.net.eval() with torch.no_grad(): x = image.cuda() if self.cuda else image yhat = self.net(x) yhat = F.softmax(yhat, dim=1) #yhat = pytutils.to_np(yhat).transpose(2,3,1,0)[...,0] return yhat def _create_model(self, arch, num_output_channels, num_input_channels, pretrained): """ Create model -arch (string): select architecture -num_classes (int) -num_channels (int) -pretrained (bool) """ self.net = None #-------------------------------------------------------------------------------------------- # select architecture #-------------------------------------------------------------------------------------------- #kw = {'num_classes': num_output_channels, 'num_channels': num_input_channels, 'pretrained': pretrained} kw = { 'num_classes': num_output_channels, 'in_channels': num_input_channels, 'pretrained': pretrained } self.net = nnmodels.__dict__[arch](**kw) self.s_arch = arch self.num_output_channels = num_output_channels self.num_input_channels = num_input_channels if self.cuda == True: self.net.cuda() if self.parallel == True and self.cuda == True: self.net = nn.DataParallel(self.net, device_ids=range( torch.cuda.device_count())) def _create_loss(self, loss): # create loss if loss == 'wmce': # Not tested self.criterion = nloss.WeightedMCEloss() elif loss == 'bdice': # Fail self.criterion = nloss.BDiceLoss() elif loss == 'wbdice': # Fail self.criterion = nloss.WeightedBDiceLoss() elif loss == 'wmcedice': # Fail self.criterion = nloss.WeightedMCEDiceLoss() elif loss == 'mcedice': # Fail self.criterion = nloss.MCEDiceLoss() elif loss == 'bce': # Pass self.criterion = nloss.BCELoss() elif loss == 'bce2c': # Pass self.criterion = nloss.BCELoss2c() elif loss == 'mce': # Pass self.criterion = nloss.MCELoss() elif loss == 'wbce': self.criterion = nloss.WeightedBCELoss() elif loss == 'wce': # Pass self.criterion = nloss.WeightedCrossEntropyLoss() elif loss == 'wfocalce': # Pass self.criterion = nloss.WeightedCEFocalLoss() elif loss == 'focaldice': # Pass self.criterion = nloss.FocalDiceLoss() elif loss == 'wfocaldice': # Pass self.criterion = nloss.WeightedFocalDiceLoss() elif loss == 'dice': # FAIL self.criterion = nloss.DiceLoss() elif loss == 'msedice': # FAIL self.criterion = nloss.MSEDICELoss() elif loss == 'mcc': # FAIL self.criterion = nloss.MCCLoss() elif loss == 'mdice': # FAIL self.criterion = nloss.MDiceLoss() elif loss == 'wcefd': self.criterion = nloss.WeightedCEFocalDice() elif loss == 'jreg': if self.num_output_channels == 2: lambda_dict = { '0': { '0': '1', '1': '0.5' }, '1': { '0': '0.5', '1': '1' } } if self.num_output_channels == 4: lambda_dict = { '0': { '0': '1', '1': '0.5', '2': '0.5', '3': '0.5' }, '1': { '0': '0.5', '1': '1', '2': '0.5', '3': '0.5' }, '2': { '0': '0.5', '1': '0.5', '2': '1', '3': '0.5' }, '3': { '0': '0.5', '1': '0.5', '2': '0.5', '3': '1' } } self.criterion = nloss.WCE_J_SIMPL(lambda_dict=lambda_dict) else: assert (False) self.s_loss = loss