han-cai/eas

Efficient Architecture Search by Network Transformation, in AAAI 2018

eas's Issues

Questions on replicating the training process

Hi,
I like this work and hope to replicate it, and wish to confirm my understanding regarding this repo:
It seems the repo only provides the trained models and a small example, and there are no visualization code for the curve. In order to replicate the training process in the paper(for example plain CNN), I would also have to implement myself based on this code. Am I right?

default value in argparse in main.py

Hi,

In order to use the default database cifar, you need to add nargs="?" in the main.py as

parser.add_argument("dataset", nargs="?", type=str, default="cifar10")

Cheers,

Reproducing Experiments

I want to replicate the experiment on a single GPU desktop workstation.
Is it possible to run it without the server client configuration ?

May I pre-trained our dataset

hi,
there are some good performance in this work. I want to pre-trained my dataset to test this method and code. may I have the pre-trained code for image classification?

Failed to repeat the training process

Hi,

Thank you for the excellent work and sharing it online. I tried to repeat the experiment but got bad luck.
I tried to use the net2net_cifar10.py to do architecture search and got an error in code/tf_network/expconfig/wider_config.py:

    return self._std_layer.init["kernel"]
KeyError: 'kernel'

The _std_layer.init is initialized as an empty dict.

I used the top_net1/config as the starting point, changed the placeholder folder names, and created my own server_config as:

[
{
"remote":"localhost",
"gpuid":0,
"executive":"RL4AS_NetTrans-master/code/main.py"
}
]

Then I ran it as
python3 net2net_cifar10.py

Do I miss something in the configuration?
Thank you!