Add different random seed for I/O vs training(#27)

You can now specify a seed in the `iotool.sampler` configuration
and a different seed in `training` configuration.
The RandomSequenceSampler should now be using a separate RNG than
PyTorch/Numpy RNG.

.travis.yml

@@ -48,4 +48,4 @@ before_script:
 
 script:
   - singularity exec local.simg pip3 install pytest --user
-  - singularity exec local.simg CUDA_VISIBLE_DEVICES='' pytest -rxXs
+  - singularity exec local.simg pytest -rxXs

contributing.md

@@ -16,7 +16,7 @@ This repository contains a framework to define, train, run, and evaluate machine
 
 Obviously, you should test your code.  Ideally, we would have a unit testing framework that would make it easy for you to prove to others that you at least didn't break something.
 
-Use the command `pytest -rxXs` to run all the tests that are currently available (still work in progress).
+Use the command `CUDA_VISBLE_DEVICES='' pytest -rxXs` to run all the tests that are currently available (still work in progress).
 
 ## Documentation
 

mlreco/iotools/samplers.py

@@ -2,45 +2,46 @@
 from __future__ import division
 from __future__ import print_function
 import numpy as np
-import torch
 from torch.utils.data import Sampler
 
+
 class AbstractBatchSampler(Sampler):
     """
     Samplers that inherit from this class should work out of the box.
     Just define the __iter__ function
     __init__ defines self._data_size and self._batch_size
+    as well as self._random RNG if needed
     """
-    def __init__(self,data_size,batch_size):
-        self._data_size  = int(data_size)
+    def __init__(self, data_size, batch_size, seed=0):
+        self._data_size = int(data_size)
         if self._data_size < 0:
             raise ValueError('%s received negative data size %s', (self.__class__.__name__, str(data_size)))
 
         self._batch_size = int(batch_size)
         if self._batch_size < 0 or self._batch_size > self._data_size:
             raise ValueError('%s received invalid batch size %d for data size %s', (self.__class__.__name__, batch_size, str(self._data_size)))
+        # Use an independent random number generator for random sampling
+        self._random = np.random.RandomState(seed=seed)
 
     def __len__(self):
         return self._data_size
 
 
 class RandomSequenceSampler(AbstractBatchSampler):
     def __iter__(self):
-        starts = torch.randint(high=self._data_size - self._batch_size,
-                               size=(len(self),))
-        return iter(np.concatenate([np.arange(start,start+self._batch_size) for start in starts]))
+        starts = self._random.randint(low=0, high=self._data_size - self._batch_size, size=(len(self),))
+        return iter(np.concatenate([np.arange(start, start+self._batch_size) for start in starts]))
 
     @staticmethod
-    def create(ds,cfg):
-        return RandomSequenceSampler(len(ds),cfg['batch_size'])
-
-
+    def create(ds, cfg):
+        return RandomSequenceSampler(len(ds), cfg['batch_size'], seed=cfg['seed'])
+
 
 class SequentialBatchSampler(AbstractBatchSampler):
     def __iter__(self):
         starts = np.arange(0, self._data_size - self._batch_size, self._batch_size)
-        return iter(np.concatenate([np.arange(start,start+self._batch_size) for start in starts]))
+        return iter(np.concatenate([np.arange(start, start+self._batch_size) for start in starts]))
 
     @staticmethod
-    def create(ds,cfg):
-        return SequentialBatchSampler(len(ds),cfg['batch_size'])
+    def create(ds, cfg):
+        return SequentialBatchSampler(len(ds), cfg['batch_size'])

mlreco/main_funcs.py

@@ -51,6 +51,12 @@ def process_config(cfg):
         cfg['training']['seed'] = int(time.time())
     else:
         cfg['training']['seed'] = int(cfg['training']['seed'])
+    # Update IO seed
+    if 'seed' not in cfg['iotool']['sampler'] or cfg['iotool']['sampler']['seed'] < 0:
+        import time
+        cfg['iotool']['sampler']['seed'] = int(time.time())
+    else:
+        cfg['iotool']['sample']['seed'] = int(cfg['iotool']['sampler']['seed'])
 
     # Batch size checker
     if cfg['iotool']['batch_size'] < 0 and cfg['training']['minibatch_size'] < 0: