A question about the performance of this code

Question

A question about the performance of this code

Opened this issue 3 years ago · 5 comments

Answer 1 · 2021-05-29T14:23:11.000Z

Dear author, thank you for work.
I am also trying to reproduce the paper NCDG but get an awful performance. I notice that the computing method of neuron coverage in your implementation is a little different from that described in the paper. So I want to ask you whether you can get the same performance as reported in the paper.

Answer 2 · 2021-08-23T13:01:51.000Z

I also wanted to ask if you reproduced the results as reported in the paper?

Answer 3 · 2021-08-25T06:27:46.000Z

I also wanted to ask if you reproduced the results as reported in the paper?
I didn't reproduced the results

Answer 4 · 2021-08-25T06:42:05.000Z

I also wanted to ask if you reproduced the results as reported in the paper?
I didn't reproduced the results

Thanks for your reply.
How many experiments you have run? What are the gaps?

Answer 5 · 2021-08-27T01:00:58.000Z

Hi , @mousecpn @lingeringlight

I ran the command

python main.py --batch_size 32 --n_classes 7 --learning_rate 0.001 --image_size 256 --nesterov True --min_scale 0.8 --max_scale 1.0 --random_horiz_flip 0.5 --jitter 0.4 --tile_random_grayscale 0.1 --source art_painting cartoon photo --target sketch --epochs 100

And I achieved

If I am not wrong, I am running one experiment from Table 7 of the paper (Source: Photo+Art Painting + Cartoon; Target: Sketch)?

As reported in the paper Photo+Art Painting + Cartoon to Sketch, the test result of ResNet is 82.1.
The results I ran is 80.7. Somehow close.