The final test accuracy

[ustctf-zz]

Hi, Thanks for this implementation ! I'm wondering how to obtain the quite strong test set result on CIFAR-10, as reported in the original densenet paper (e.g., error rate <=3.5 on C-10+, with depth =190, growth_rate = 40). When I run the script as:

_CUDA_VISIBLEDEVICES=0,1,2,3 python demo.py --depth 190 --efficient False --data ./data --save ./ckpts

The final test error is reported as 0.0535. I'm wondering whether the high error is due to no data augmentation is conducted in the default setting. May I know whether it is C10+ dataset or C10?

Best

[taineleau-zz]

Could you please show the configuration (e.g., pytorch version, python version)?

[gpleiss]

I noticed recently that I was missing the appropriate initialization of the models. Try it again and let me know what the error is.

[gpleiss]

With the new initialization in c609c0c4 I was able to match the errors reported in the original paper. Closing this issue for now. If you're still noticing issues feel free to reopen.

[ustctf-zz]

@gpleiss

Sorry but I got the same test set error rate (>5) using your new initialization.

The command:

CUDA_VISIBLE_DEVICES=0,1,2,3 python demo.py --depth 190 --efficient False --data ./data --save ./ckpts

Would you please show your running script to match the claimed error rate? What is the exact error rate reported finally?

BTW @taineleau the python version: Python 3.6.1 |Anaconda custom (64-bit)| pytorch version: 0.3.1.post2 gpu: 4 titanXp system: Ubuntu 14.04 cuda: 8.0

Thanks.

[gpleiss]

python demo.py --depth 100 --efficient False --data ./data --save ./ckpts --batch_size 64 --valid_size 0

These are the same hyperparameters as in the paper. I got a final test error of 0.047

[ustctf-zz]

Thanks @gpleiss , I can reproduce now and with --depth 190, --growth_rate 40, the test error is 0.035.