Question about weights update in network of phase test

ssnl / dataset-distillation

Open-source code for paper "Dataset Distillation"

https://ssnl.github.io/dataset_distillation

MIT License

778 stars 115 forks source link

Question about weights update in network of phase test #48

Open Firegreat123 opened 2 years ago

Firegreat123 commented 2 years ago

Hi, after training 10 distilled images, how does these 10 distilled images update the weights of the neural network in phase test?

ssnl commented 2 years ago

the entire testing code is here https://github.com/SsnL/dataset-distillation/blob/master/main.py#L244-L356

Firegreat123 commented 2 years ago

Hi, I have read codes and have a little question. Does the code below used to train test network with distilled images?

ssnl commented 2 years ago

yes that is correct!

Firegreat123 commented 2 years ago

Thank you. After running the codes below, in the output.log I can see that Test accuracy is 79.8% But after I run the test code below, the accuracy will be dropped into 23.1868%. Why is this happen?

python main.py --mode distill_adapt --source_dataset MNIST_RGB --dataset SVHN --arch LeNet --phase test \ --train_nets_type loaded --test_nets_type same_as_train \ --test_distilled_images loaded --test_distilled_lrs loaded