search

zoogzog / chexnet

Implementation of the CheXNet network (PyTorch)
211 stars 94 forks source link

Detailed Hardware specification & type of DNN used. #2

Closed Gambitier closed 6 years ago

Gambitier commented 6 years ago

  1. I got that you used Tesla P100 GPU, but I just wanted to know what are its specifications e.g. memory, DDR4/DDR5, no. of cores etc.

  2. second, what was CPU specification?

  3. Which Deep Neural Nets can be used to train images? Did you used ConvNets? and what do you think about using Capsule Neural Nets to train the model? will it improve model accuracy?

zoogzog commented 6 years ago

1&2 Here are the specifications of my machine: 2xXeon CPU E52603, 4xTesla P100 NVLINK (3584 cores), RAM 256GB

The results I got when running on single GPU, by specifying CUDA_VISIBLE_DEVICES for the python script, thinking that it would load the GPU better than spreading in parallel with batch size = 16. My latest tests showed that running in parallel on all 4 GPUs with batch size 256 allowed to increase processing speed (~ 3 times), without significant changes in accuracy.

3 My primary interest was implementing the CheXNet, so I have not done any tests with other networks yet. The authors of the paper, that presented (ChestX-ray)[https://arxiv.org/pdf/1705.02315.pdf] did some tests with ResNet, VGGN, AlexNet. So far, DenseNet architecture showed best results.

Gambitier commented 6 years ago

Hey, thanks for your response, the hardware specs you specified seems costly, so how did you managed to get it? Did you used cloud services like aws, GCP etc? and what was the total cost required to train your model?