search

chenxin061 / pdarts

Codes for our paper "Progressive Differentiable Architecture Search:Bridging the Depth Gap between Search and Evaluation"
Other
359 stars 83 forks source link

How to test? #11

Closed zihaozhang9 closed 5 years ago

zihaozhang9 commented 5 years ago

I run the following code search: python train_search.py \\ --tmp_data_dir /path/to/your/data \\ --save log_path \\ --add_layers 6 \\ --add_layers 12 \\ --dropout_rate 0.1 \\ --dropout_rate 0.4 \\ --dropout_rate 0.7 \\ --note note_of_this_run

I get the weights.pt

I run the following code to test: python test.py --auxiliary --model_path log_pathsearch-note_of_this_run-20190615-081053/weights.pt --data /path/to/your/data

But Error.How to test?

zihaozhang9 commented 5 years ago

Sorry! I forgot the train.

rtrobin commented 5 years ago

@zihaozhang9 @chenxin061 I don't mean to open a duplicated issue. So I'm leaving comments here following this issue.

After the architecture search process, which is provided in _trainsearch.py, training process is required. However, in _traincifar.py script, it doesn't load the previous searched result in args. Here's my questions:

  1. Why not loading the searched result?
  2. Now I want to try PDARTS on other datasets. Is it better to use the network transferred from pre-searched cifar model, or search from the dataset itself? For other NAS methods, maybe only the former option is doable. Since PDARTS is time efficient in searching, maybe I should search from it?

Any comment is appreciated. Thanks. :)

zihaozhang9 commented 5 years ago

train_search.py

  1. Yes, the structure searched in train_search.py is not loaded in train_cifar.py. The network structure loaded in train_cifar.py is the best structure in the paper. I have modified the code for train_cifar.py to load the structure searched in train_search.py. I will upload it to my own github right away. You can also communicate with me first, I can give you the modified code.
  2. I have not tried in other data sets yet. I remember that the paper said that because the network structure was changed, all the initialization models were not used, but the training was re-trained.
rtrobin commented 5 years ago

@zihaozhang9 Thanks for the reply, helping me understand the training part. I will post any issue to your repo, if I have other training question in future.

I will post another issue, discussing network transfer in this repo. Since the question in not related to training anymore.