Computer Vision News - November 2021

66 import segmentation_models_pytorch as smp self.model = smp.Unet( For the self-teaching network, features are initially learned in a supervised way while only using manually annotated data until it reaches an f-score above a certain threshold. ThenI used the unlabelled data to generate pseudo-labels, while using test-time augmentation to accomplish certainty estimation and majority voting while only using masks with a certainty above a threshold. For the continuation of the training, both manual and pseudo labels are used, and pseudo labels are added/updated iteratively. for iteration in range(1, configs.num_epochs+1): encoder_name="resnet50", # choose encoder, e.g. mobilenet_v2 or efficientnet-b7 encoder_weights="imagenet", # use `imagenet` pre-trained weights for encoder initialization in_channels=1, # model input channels (1 for gray-scale images, 3 for RGB) classes=1, # model output channels (number of classes in your dataset) ) # train initially on labelled ( 1 ) and further on combined dataset ( 3 ) train_mode.reload_data(configs) # labelled images and pseudo labels for unlabelled images fscore = train_mode.train(configs, iteration) if fscore > configs.threshold: # generate pseudo labels ( 2 ) pseudo_mode.reload_data(configs) # unlabelled images pseudo_mode.generate_labels(configs, iteration) # set high certainty pseudo labels for unlabelled images pseudo_mode.update_dataset(configs) My summer internship ...