Closed FFFox-abc closed 10 months ago
When I test the localization method, the RocAUC is different every time. the results are fluctuate within 0.5 roughly. my config is:
experiment_name: 'localization_local_equal_net' dataset_name: mvtec # [mnist, fashionmnist, cifar10, mvtec, retina] last_checkpoint: 600
num_epochs: 601 # mnist/fashionmnist:51, cifar10:201, mvtec:601 batch_size: 32 learning_rate: 1e-3 mvtec_img_size: 128
normal_class: 'toothbrush' # mvtec:'capsule', mnist:3
lamda: 0.5 # mvtec:0.5, Others:0.01
pretrain: True # True:use pre-trained vgg as source network --- False:use random initialize use_bias: False # True:using bias term in neural network layer equal_network_size: False # True:using equal network size for cloner and source network --- False:smaller network for cloner direction_loss_only: False continue_train: True
localization_test: True # True:For Localization Test --- False:For Detection localization_method: 'smooth_grad' # gradients , smooth_grad , gbp
fix random seed
When I test the localization method, the RocAUC is different every time. the results are fluctuate within 0.5 roughly. my config is:
Data parameters
experiment_name: 'localization_local_equal_net' dataset_name: mvtec # [mnist, fashionmnist, cifar10, mvtec, retina] last_checkpoint: 600
Training parameters
num_epochs: 601 # mnist/fashionmnist:51, cifar10:201, mvtec:601 batch_size: 32 learning_rate: 1e-3 mvtec_img_size: 128
normal_class: 'toothbrush' # mvtec:'capsule', mnist:3
lamda: 0.5 # mvtec:0.5, Others:0.01
pretrain: True # True:use pre-trained vgg as source network --- False:use random initialize use_bias: False # True:using bias term in neural network layer equal_network_size: False # True:using equal network size for cloner and source network --- False:smaller network for cloner direction_loss_only: False continue_train: True
Test parameters
localization_test: True # True:For Localization Test --- False:For Detection localization_method: 'smooth_grad' # gradients , smooth_grad , gbp