Skip to content

bsc-dd/deeplearning_scratch

 
 

Repository files navigation

This programms is dooing the I/O for deep-learning on large images usage: For I/O simulating only:

mpirun -n 20 python ./run_training.py OUT_DIR

With deep-Learning (using 4 GPUs on proc 19) mpirun -n 20 python run_train_tf.py ./log --num-gpus 4 --num-threads 12

In the current confiuration, it will only run with 20 procs The Procs are dooing the following

0-12: Read data for Training, send to 17 13-16: Read data for Testing, send to 18

17: recv data from 0-12, send to 19 18: recv data from 14-16, send to 19

19: do learing and testing (currently just idle)

can be configured in test_params.py in mpi_params

What else to configure (all in test_params.py)

'iterations': number of iterations 'test_interval':200: how often ist tested. 'test_iter': number of iterations in test

'batch_size': how many crops are read before send 'input_size': size of one crop (or box)

Data is read in the get_crop routines in "batch_iter", the actuall I/O happens in

def crop_from_array(self, arr) in the line crop = arr[y0:y1,x0:x1]

About

IO Simulation of deep learing

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%