search

jaredleekatzman / DeepSurv

DeepSurv is a deep learning approach to survival analysis.
MIT License
565 stars 166 forks source link

hyperparameter search #60

Open quzhouxiachuan opened 4 years ago

quzhouxiachuan commented 4 years ago

Hi Jared, Thanks for providing deepsurv. I have a question about hyperparameter serach. I am not sure what is the benefit of using docker oputunity vs simply giving a list of parameters options in the parameter variable as shown in DeepSurv_example.ipynb:

hyperparams = { 'L2_reg': 10.0, 'batch_norm': True, 'dropout': 0.4, 'hidden_layers_sizes': [25, 25], 'learning_rate': 1e-05, 'lr_decay': 0.001, 'momentum': 0.9, 'n_in': train_data['x'].shape[1], 'standardize': True }

For example, you can specify learning rate as [0.1,0.001,0.0001] here.

In addition, is it ok to use multiple cores for parallel running using your script?

Much appreciated!

quzhouxiachuan commented 4 years ago

A follow up question on this is, is it possible to use gpu without docker?

saidileep-knv commented 4 years ago

You can use GPU without docker. Usually, for any matrix or vector computations, NumPy uses base C-API if none of the high-level libraries are integrated with it. You can configure CUDA drivers to increase computational speed (if your GPU is compatible with CUDA) or manually build the OpenBLAS or any other libraries and link them to Theano in the environment variables of the system.