singa-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From GitBox <...@apache.org>
Subject [GitHub] [singa] chrishkchris edited a comment on issue #651: [WIP ]Simply example APIs
Date Sun, 05 Apr 2020 00:03:13 GMT
chrishkchris edited a comment on issue #651: [WIP ]Simply example APIs
URL: https://github.com/apache/singa/pull/651#issuecomment-609049842
 
 
   > I suggest to merge the examples under `examples/autograd` into the follow structure.
   > 
   > ```
   > autograd
   >     train.py 
   >     train_mpi.py
   >     train_multiprocess.py
   >     data   # define the data loading and preprocessing
   >         cifar10.py
   >         mnist.py
   >     model   # define the model
   >         cnn.py
   >         resnet.py
   >         xception.py
   > ```
   > 
   > The pseudo code of each file:
   > 
   > ```python
   > # train.py
   > def run(max_epoch, worker_id, num_workers, model, data, sgd):
   >     if model == 'resnet':
   >        from model import resnet
   >          model = resnet.create_model()
   >     elif model == 'cnn':
   >          model = cnn.create_model()    
   >     ....
   >     if data == 'cifar10':
   >         from data import cifar10:
   >            train_x, train_y, val_x, val_y = cifar10.load()
   >     elif data == 'mnist':
   >          ....
   >     
   >     train_x, train_y, val_x, val_y = partition(worker_id, num_workers, train_x, train_y,
val_x, val_y)
   >    
   >     # bp and sgd
   > 
   > if __name__ == '__main__':
   >    # use argparse to get command config: max_epoch, model, data, etc. for single gpu
training
   >     sgd = # create sgd
   >     run(0, 1, ..., sgd)
   > 
   > # train_mpi.py
   > if __name__ == '__main__':
   >    # use argparse to get command args: max_epoch, model, data, etc. for multi-gpu
training
   >     sgd = # create sgd
   >     dist_sgd = DistOpt(sgd...)
   >     run(dist_sgd.rank, dist_sgd.world_size, ... dist_sgd)
   > 
   > # train_multiprocess.py
   > def run(rank, num_gpu, ...):
   >     sgd = ...
   >     dist_sgd = DistOpt(sgd)
   >     train.run(rank, num_gpu, ... dist_sgd)
   > 
   > if __name__ == '__main__':
   >    # use argparse to get command args: max_epoch, model, data, etc. for multi-gpu
training
   >     nccl_id = ...
   >     process = []
   >     for worker_id in range(0, gpu_per_node):        
   >         process.append(multiprocessing.Process(target=run, args=(worker_id, ... nccl_id)))

   > 
   >     for p in process:
   >         p.start()
   > ```
   
   thanks! I am working on it. There may be three issues that need to think about: 
   (i) maybe difficult to include graph module (please see resnet_module.py to see the different
between layer model and module model). In this case, we may consider 
   (ii) don't know how to show this example in the doc dist-train.md. So maybe in the doc
we can still use doc_dist_train.py for examples
   
   
   
   

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services

Mime
View raw message