Adds seed to config (and to evaluation tool)

luminoth/datasets/object_detection_dataset.py

@@ -33,7 +33,7 @@ class ObjectDetectionDataset(snt.AbstractModule):
         random_shuffle (bool): To consume the dataset using random shuffle or
             to just use a regular FIFO queue.
     """
-    def __init__(self, config, seed=None, **kwargs):
+    def __init__(self, config, **kwargs):
         """
         Save general purpose attributes for Dataset module.
 
@@ -50,7 +50,7 @@ def __init__(self, config, seed=None, **kwargs):
         self._random_shuffle = config.train.random_shuffle
         # In case no keys are defined, default to empty list.
         self._data_augmentation = config.dataset.data_augmentation or []
-        self._seed = seed
+        self._seed = config.train.seed
 
     def _build():
         pass

luminoth/datasets/object_detection_dataset_test.py

@@ -22,6 +22,7 @@ def setUp(self):
                 'num_epochs': 1,
                 'batch_size': 1,
                 'random_shuffle': False,
+                'seed': None,
             }
         })
 

luminoth/eval.py

@@ -7,7 +7,7 @@
 from luminoth.datasets import TFRecordDataset
 from luminoth.models import get_model
 from luminoth.utils.config import (
-    load_config, merge_into, parse_override
+    get_model_config
 )
 from luminoth.utils.bbox_overlap import bbox_overlap
 
@@ -28,18 +28,12 @@ def evaluate(model_type, dataset_split, config_file, job_dir, watch,
     model_cls = get_model(model_type)
     config = model_cls.base_config
 
-    if config_file:
-        # If we have a custom config file overwritting default settings
-        # then we merge those values to the base_config.
-        custom_config = load_config(config_file)
-        config = merge_into(custom_config, config)
+    config = get_model_config(
+        model_cls.base_config, config_file, override_params
+    )
 
     config.train.job_dir = job_dir or config.train.job_dir
 
-    if override_params:
-        override_config = parse_override(override_params)
-        config = merge_into(override_config, config)
-
     if config.train.debug or config.train.tf_debug:
         tf.logging.set_verbosity(tf.logging.DEBUG)
     else:
@@ -53,6 +47,10 @@ def evaluate(model_type, dataset_split, config_file, job_dir, watch,
     # Only a single run over the dataset to calculate metrics.
     config.train.num_epochs = 1
 
+    # Seed setup
+    if config.train.seed:
+        tf.set_random_seed(config.train.seed)
+
     # Set pretrained as not training
     config.pretrained.trainable = False
 

luminoth/models/fasterrcnn/base_config.yml

@@ -1,6 +1,8 @@
 train:
   # Run on debug mode (which enables more logging)
   debug: False
+  # Seed for random operations
+  seed:
   # Training batch size for images. FasterRCNN currently only supports 1
   batch_size: 1
   # Directory in which model checkpoints & summaries (for Tensorboard) will be saved

luminoth/models/fasterrcnn/network.py

@@ -23,8 +23,7 @@ class FasterRCNN(snt.AbstractModule):
 
     base_config = get_base_config(__file__)
 
-    def __init__(self, config, with_rcnn=True, num_classes=None, debug=False,
-                 seed=None, name='fasterrcnn'):
+    def __init__(self, config, name='fasterrcnn'):
         super(FasterRCNN, self).__init__(name=name)
 
         # Main configuration object, it holds not only the necessary
@@ -43,7 +42,7 @@ def __init__(self, config, with_rcnn=True, num_classes=None, debug=False,
         # Turn on debug mode with returns more Tensors which can be used for
         # better visualization and (of course) debugging.
         self._debug = config.train.debug
-        self._seed = seed
+        self._seed = config.train.seed
 
         # Anchor config, check out the docs of base_config.yml for a better
         # understanding of how anchors work.

luminoth/models/fasterrcnn/network_test.py

@@ -17,7 +17,8 @@ def setUp(self):
                 'with_rcnn': True
             },
             'train': {
-                'debug': True
+                'debug': True,
+                'seed': None,
             },
             'anchors': {
                 'base_size': 256,
@@ -48,8 +49,7 @@ def setUp(self):
                     'type': 'variance_scaling_initializer',
                     'factor': 1.0,
                     'uniform': 'True',
-                    'mode': 'FAN_AVG',
-                    'seed': 0
+                    'mode': 'FAN_AVG'
                 },
                 'roi': {
                     'pooling_mode': 'crop',
@@ -76,8 +76,7 @@ def setUp(self):
                'initializer': {
                  'type': 'truncated_normal_initializer',
                  'mean': 0.0,
-                 'stddev': 0.01,
-                 'seed': 0
+                 'stddev': 0.01
                },
                'activation_function': 'relu6',
                'l2_regularization_scale': 0.0005,
@@ -112,7 +111,7 @@ def _run_network(self):
             tf.float32, shape=self.image.shape)
         gt_boxes = tf.placeholder(
             tf.float32, shape=self.gt_boxes.shape)
-        model = FasterRCNN(self.config, debug=True)
+        model = FasterRCNN(self.config)
 
         results = model(image, gt_boxes)
 
@@ -127,7 +126,7 @@ def _run_network(self):
     def _gen_anchors(self, config, feature_map):
         feature_map_tf = tf.placeholder(
             tf.float32, shape=feature_map.shape)
-        model = FasterRCNN(config, debug=True)
+        model = FasterRCNN(config)
 
         results = model._generate_anchors(feature_map)
 

luminoth/train.py

@@ -15,19 +15,18 @@
 )
 
 
-def run(model_type, config_file, override_params, seed, target='',
-        cluster_spec=None, is_chief=True, job_name=None, task_index=None,
-        **kwargs):
-
-    if seed:
-        tf.set_random_seed(seed)
+def run(model_type, config_file, override_params, target='', cluster_spec=None,
+        is_chief=True, job_name=None, task_index=None, **kwargs):
 
     model_class = get_model(model_type)
 
     config = get_model_config(
         model_class.base_config, config_file, override_params, **kwargs
     )
 
+    if config.train.seed is not None:
+        tf.set_random_seed(config.train.seed)
+
     log_prefix = '[{}-{}] - '.format(job_name, task_index) \
         if job_name is not None and task_index is not None else ''
 
@@ -36,7 +35,7 @@ def run(model_type, config_file, override_params, seed, target='',
     else:
         tf.logging.set_verbosity(tf.logging.INFO)
 
-    model = model_class(config, seed=seed)
+    model = model_class(config)
 
     # Placement of ops on devices using replica device setter
     # which automatically places the parameters on the `ps` server
@@ -45,7 +44,7 @@ def run(model_type, config_file, override_params, seed, target='',
     # See:
     # https://www.tensorflow.org/api_docs/python/tf/train/replica_device_setter
     with tf.device(tf.train.replica_device_setter(cluster=cluster_spec)):
-        dataset = TFRecordDataset(config, seed=seed)
+        dataset = TFRecordDataset(config)
         train_dataset = dataset()
 
         train_image = train_dataset['image']