Removed --limit-classes option from dataset readers.

Luminoth shouldn't have to generate random classes for you. This should
be responsibility of the user.

luminoth/tools/dataset/readers/object_detection/object_detection_reader.py

@@ -23,16 +23,14 @@ class ObjectDetectionReader(BaseReader):
             Iterate over all records.
     """
     def __init__(self, only_classes=None, only_images=None,
-                 limit_examples=None, limit_classes=None, seed=None, **kwargs):
+                 limit_examples=None, **kwargs):
         """
         Args:
             - only_classes: string or list of strings used as a class
                 whitelist.
             - only_images: string or list of strings used as a image_id
                 whitelist.
             - limit_examples: limit number of examples to use.
-            - limit_classes: limit number of classes to use.
-            - seed: seed for random.
         """
         super(ObjectDetectionReader, self).__init__()
         if isinstance(only_classes, six.string_types):
@@ -46,9 +44,6 @@ def __init__(self, only_classes=None, only_images=None,
         self._only_images = only_images
 
         self._limit_examples = limit_examples
-        self._limit_classes = limit_classes
-        random.seed(seed)
-
         self._total = None
         self._classes = None
 
@@ -66,16 +61,19 @@ def classes(self):
 
     @abc.abstractmethod
     def get_total(self):
-        """Returns the total amount of records in the dataset.
+        """
+        Returns the total amount of records in the dataset.
         """
 
     @abc.abstractmethod
     def get_classes(self):
-        """Returns all the classes available in the dataset.
+        """
+        Returns all the classes available in the dataset.
         """
 
     def _filter_total(self, original_total_records):
-        """Filters total number of records in dataset based on reader options
+        """
+        Filters total number of records in dataset based on reader options
         used.
         """
         # Define smaller number of records when limiting examples.
@@ -89,16 +87,11 @@ def _filter_total(self, original_total_records):
         return new_total
 
     def _filter_classes(self, original_classes):
-        """Filters classes based on reader options used.
+        """
+        Filters classes based on reader options used.
         """
         if self._only_classes:  # not None and not empty
             new_classes = sorted(self._only_classes)
-        # Choose random classes when limiting them
-        elif self._limit_classes is not None and self._limit_classes > 0:
-            total_classes = min(len(original_classes), self._limit_classes)
-            new_classes = sorted(
-                random.sample(original_classes, total_classes)
-            )
         else:
             new_classes = list(original_classes) if original_classes else None
 
@@ -121,7 +114,8 @@ def _stop_iteration(self):
 
     @abc.abstractmethod
     def iterate(self):
-        """Iterate over object detection records read from the dataset source.
+        """
+        Iterate over object detection records read from the dataset source.
 
         Returns:
             iterator of records of type `dict` with the following keys:

luminoth/tools/dataset/transform.py

@@ -8,38 +8,30 @@
 from .writers import ObjectDetectionWriter
 
 
-def get_output_subfolder(only_classes, only_images, limit_examples,
-                         limit_classes):
+def get_output_subfolder(only_classes, only_images, limit_examples):
     """
-    Returns: subfolder name for records
+    Returns: subfolder name for records.
     """
     if only_classes is not None:
         return 'classes-{}'.format(only_classes.replace('/', ''))
     elif only_images is not None:
         return 'only-{}'.format(only_images)
-    elif limit_examples is not None and limit_classes is not None:
-        return 'limit-{}-classes-{}'.format(limit_examples, limit_classes)
     elif limit_examples is not None:
         return 'limit-{}'.format(limit_examples)
-    elif limit_classes is not None:
-        return 'classes-{}'.format(limit_classes)
 
 
 @click.command()
 @click.option('dataset_reader', '--type', type=click.Choice(READERS.keys()), required=True)  # noqa
 @click.option('--data-dir', required=True, help='Where to locate the original data.')  # noqa
 @click.option('--output-dir', required=True, help='Where to save the transformed data.')  # noqa
-@click.option('splits', '--split', required=True, multiple=True, help='Which splits to transform.')  # noqa
-@click.option('--only-classes', help='Whitelist of classes.')
-@click.option('--only-images', help='Create dataset with specific examples.')
-@click.option('--limit-examples', type=int, help='Limit dataset with to the first `N` examples.')  # noqa
-@click.option('--limit-classes', type=int, help='Limit dataset with `N` random classes.')  # noqa
-@click.option('--seed', type=int, help='Seed used for picking random classes.')
+@click.option('splits', '--split', required=True, multiple=True, help='The splits to transform (ie. train, test, val).')  # noqa
+@click.option('--only-classes', help='Keep only examples of these classes. Comma separated list.')  # noqa
+@click.option('--only-images', help='Create dataset with specific examples. Useful to test model if your model has the ability to overfit.')  # noqa
+@click.option('--limit-examples', type=int, help='Limit dataset with to the first global `N` examples (not per class).')  # noqa
 @click.option('overrides', '--override', '-o', multiple=True, help='Custom parameters for readers.')  # noqa
 @click.option('--debug', is_flag=True, help='Set level logging to DEBUG.')
 def transform(dataset_reader, data_dir, output_dir, splits, only_classes,
-              only_images, limit_examples, limit_classes, seed, overrides,
-              debug):
+              only_images, limit_examples, overrides, debug):
     """
     Prepares dataset for ingestion.
 
@@ -53,7 +45,7 @@ def transform(dataset_reader, data_dir, output_dir, splits, only_classes,
     # We forcefully save modified datasets into subfolders to avoid
     # overwriting and/or unnecessary clutter.
     output_subfolder = get_output_subfolder(
-        only_classes, only_images, limit_examples, limit_classes
+        only_classes, only_images, limit_examples
     )
     if output_subfolder:
         output_dir = os.path.join(output_dir, output_subfolder)
@@ -75,8 +67,7 @@ def transform(dataset_reader, data_dir, output_dir, splits, only_classes,
             split_reader = reader(
                 data_dir, split,
                 only_classes=only_classes, only_images=only_images,
-                limit_examples=limit_examples, limit_classes=limit_classes,
-                seed=seed, **reader_kwargs
+                limit_examples=limit_examples, **reader_kwargs
             )
 
             if classes is None: