added threadpoolctl to deal with multiprocessing and threading issues…

…, added missing requirement scikit-learn to requirements.txt

Readme.md

@@ -91,25 +91,10 @@ if __name__ == '__main__':
 
 This is not required on Linux.
 
-## Important!
-Starting from version 1.14.6 numpy has issues with multiprocessing (it's supposed to be a feature...). 
-Mutrix multiplications (which we are using to rotate coordinate systems for data augmentation) now run mutlithreaded on all available threads. 
-This can cause chaos if you are using a multiprocessing pipeline, beacause each background worker will spawn a lot of 
-threads to do the matrix multiplication (8 workers on a 16 Core machine = up to 8*16=256 threads. duh.). There is nothing we (dkfz devs) can do to 
-tackle that problem, but this will only be a real issue in very specific configurations of data augmentation. If you 
-notice unnecessarily high CPU load, there are two things you can do:
-
-- (recommended) run all your experiments with `OMP_NUM_THREADS=1` or (even better) add this to your environment 
-variables (`export OMP_NUM_THREADS=1` in .bashrc on linux)
-- downgrade numpy to 1.14.5 (pip install numpy==1.14.5) to solve the issue
-
-Numpy devs are aware of this problem and trying to find a solution 
-(see https://github.com/numpy/numpy/issues/11826#issuecomment-425087772)
-
 
 ## Release Notes
 
-
+- 0.19.5: fixed OMP_NUM_THREADS issue by using threadpoolctl package
 - 0.19:
    - There is now a complete example for BraTS2017/8 available for both 2D and 3D. Use this if you would like to get 
    some insights on how I (Fabian) do my experiments

batchgenerators/dataloading/multi_threaded_augmenter.py

@@ -28,35 +28,37 @@
 from queue import Empty, Full
 import traceback
 from time import sleep, time
+from threadpoolctl import threadpool_limits
 
 
 def producer(queue, data_loader, transform, thread_id, seed, abort_event):
     try:
-        np.random.seed(seed)
-        data_loader.set_thread_id(thread_id)
-        item = None
+        with threadpool_limits(limits=1):
+            np.random.seed(seed)
+            data_loader.set_thread_id(thread_id)
+            item = None
 
-        while True:
-            # check if abort event was set
-            if not abort_event.is_set():
+            while True:
+                # check if abort event was set
+                if not abort_event.is_set():
 
-                if item is None:
+                    if item is None:
+
+                        try:
+                            item = next(data_loader)
+                            if transform is not None:
+                                item = transform(**item)
+                        except StopIteration:
+                            item = "end"
 
                     try:
-                        item = next(data_loader)
-                        if transform is not None:
-                            item = transform(**item)
-                    except StopIteration:
-                        item = "end"
-
-                try:
-                    queue.put(item, timeout=2)
-                    item = None
-                except Full:
-                    # queue was full because items in it were not consumed. Try again.
-                    pass
-            else:
-                break
+                        queue.put(item, timeout=2)
+                        item = None
+                    except Full:
+                        # queue was full because items in it were not consumed. Try again.
+                        pass
+                else:
+                    break
 
     except KeyboardInterrupt:
         # drain queue, then give 'end', set abort flag and reraise KeyboardInterrupt

requirements.txt

@@ -1,5 +1,6 @@
+threadpoolctl
 scikit-learn
 numpy>=1.10.2
 scipy
 scikit-image
-
+scikit-learn

setup.py

@@ -1,7 +1,7 @@
 from setuptools import setup
 
 setup(name='batchgenerators',
-      version='0.19.4',
+      version='0.19.5',
       description='Data augmentation toolkit',
       url='https://github.com/MIC-DKFZ/batchgenerators',
       author='Division of Medical Image Computing, German Cancer Research Center',
@@ -16,7 +16,8 @@
             "scikit-image",
             "scikit-learn",
             "future",
-            "unittest2"
+            "unittest2",
+            "threadpoolctl"
       ],
       keywords=['deep learning', 'image segmentation', 'image classification', 'medical image analysis',
                   'medical image segmentation', 'data augmentation'],