- use 10 cross validation,customized resnet20/pretrained resnet18/perceptron , weight decay , data augmentation
- The average training accuracy is : 0.9278 , the average test accuracy is : 0.5030 , and the highest accuracy of the entire data is :0.9240.(with customized resnet)
- The average training accuracy is : 0.9992 , the average test accuracy is : 0.5220 The best net is net3 , accuracy of the entire data is :0.9060 (with pretrained resnet on Imagenet)
- With this classifier , find two examples that are similar with clock and crocodile at the same time(it depends on classifier and a hyperparameter "threshold")
clock that looks like crocodile
crocodile that looks like clock
pair1
pair2
pair3
- A DCGAN to generate images which are similar with clock and crocodile at the same time , but i failed , actually time is not enough , otherwise i can try more architectures and hypermeters.
- classifier_pretrained : train_res(k_cross) , set lr , wd in 'Trainer'
- detector : print_similar_image(data_array ,label_target , net , threshold , ctx) / or set the hamming distance between images
- generator(can't generate human recognized images yet)