Can't replicate some results

[kirk86]

Hi, thanks for open sourcing the implementation. I was wondering if you could provide some insights on things I might be missing during training the dpn. I've tried to use most of the default hyperparams in addition to those specified in the paper but still cannot reproduce some of the results.

More specifically when I train the dpn I used the following hyperparams

epochs=100
lr = 7.5e-4
lrc=70
lr_decay=0.2
target_concentration = 1e-2
concentration = 1.0
gamma = 1.0
weight_decay = 0
augment
rotate
jitter=0.4
clip_norm=10.0

Am I missing somehting, or maybe training it with the wrong hyperparam values?

[1163710124]

Hi, thanks for open sourcing the implementation. I was wondering if you could provide some insights on things I might be missing during training the dpn. I've tried to use most of the default hyperparams in addition to those specified in the paper but still cannot reproduce some of the results.

More specifically when I train the dpn I used the following hyperparams

epochs=100
lr = 7.5e-4
lrc=70
lr_decay=0.2
target_concentration = 1e-2
concentration = 1.0
gamma = 1.0
weight_decay = 0
augment
rotate
jitter=0.4
clip_norm=10.0

Am I missing somehting, or maybe training it with the wrong hyperparam values?

Hello krik86, I want to train the dpn too. But I don't know how to run it. Can u give me some guidance or suggestion? Thank u so much! (no readme seems too hard for me...)

[kirk86]

You need to call the following two files in sequential order

1. setup_dpn.py
2. train_dpn.py

The first will setup the model and the necessary directories and the second will train. Look at each file to read the arguments you can pass.

[1163710124]

Thank you very much! I will try it immediately!!! This is my email: [email protected]. If you want, we can communicate further. Thanks again!

[dongbihe]

You need to call the following two files in sequential order

1. `setup_dpn.py`

2. `train_dpn.py`

The first will setup the model and the necessary directories and the second will train. Look at each file to read the arguments you can pass.

Hey how did you manage to run all ? and which python version did you use ?