Greater Configurability

src/rastervision/label_stores/segmentation_raster_file.py

@@ -79,7 +79,7 @@ def src_to_rv(n: int) -> int:
             if n in src_to_rv_class_map:
                 return src_to_rv_class_map.get(n)
             else:
-                return 0
+                return 0x00
 
         self.src_to_rv = np.vectorize(src_to_rv, otypes=[np.uint8])
 

src/rastervision/label_stores/segmentation_raster_file_test.py

@@ -8,6 +8,7 @@
 
 
 class TestingRasterSource(RasterSource):
+
     def __init__(self, zeros=False):
         self.width = 4
         self.height = 4
@@ -42,6 +43,7 @@ def get_crs_transformer(self, window):
 
 
 class TestSegmentationRasterFile(unittest.TestCase):
+
     def test_clear(self):
         label_store = SegmentationRasterFile(TestingRasterSource(), None)
         extent = label_store.src.get_extent()
@@ -58,7 +60,7 @@ def test_set_labels(self):
         label_store.set_labels(raster_source)
         extent = label_store.src.get_extent()
         rs_data = raster_source._get_chip(extent)
-        ls_data = label_store.get_labels(extent)
+        ls_data = (label_store.get_labels(extent) == 1)
         self.assertEqual(rs_data.sum(), ls_data.sum())
 
 

src/rastervision/ml_backends/tf_deeplab.py

@@ -209,16 +209,15 @@ def create_tf_example(image: np.ndarray,
     """
     class_keys = set(class_map.get_keys())
 
-    def fn(n):
-        return (n if n in class_keys else 0)
-
-    filtered_labels = np.array(np.vectorize(fn)(labels), dtype=np.uint8)
+    def _clean(n):
+        return (n if n in class_keys else 0x00)
+    clean = np.vectorize(_clean, otypes=[np.uint8])
 
     image_encoded = numpy_to_png(image)
     image_filename = chip_id.encode('utf8')
     image_format = 'png'.encode('utf8')
     image_height, image_width, image_channels = image.shape
-    image_segmentation_class_encoded = numpy_to_png(filtered_labels)
+    image_segmentation_class_encoded = numpy_to_png(clean(labels))
     image_segmentation_class_format = 'png'.encode('utf8')
 
     features = tf.train.Features(
@@ -308,6 +307,9 @@ def get_training_args(train_py: str, train_logdir_local: str, tfic_index: str,
     for size in be_options.train_crop_size:
         args.append('--train_crop_size={}'.format(size))
 
+    if len(be_options.dataset) > 0:
+        args.append('--dataset={}'.format(be_options.dataset))
+
     args.append('--train_logdir={}'.format(train_logdir_local))
     args.append('--tf_initial_checkpoint={}'.format(tfic_index))
     args.append('--dataset_dir={}'.format(dataset_dir_local))
@@ -317,14 +319,22 @@ def get_training_args(train_py: str, train_logdir_local: str, tfic_index: str,
     args.append('--decoder_output_stride={}'.format(
         be_options.decoder_output_stride))
     args.append('--train_batch_size={}'.format(be_options.train_batch_size))
-    if len(be_options.dataset) > 0:
-        args.append('--dataset="{}"'.format(be_options.dataset))
     args.append('--save_interval_secs={}'.format(
         be_options.save_interval_secs))
     args.append('--save_summaries_secs={}'.format(
         be_options.save_summaries_secs))
     args.append('--save_summaries_images={}'.format(
         be_options.save_summaries_images))
+    args.append('--last_layer_gradient_multiplier={}'.format(
+        be_options.last_layer_gradient_multiplier))
+    args.append('--initialize_last_layer={}'.format(
+        be_options.initialize_last_layer))
+    args.append('--min_scale_factor={}'.format(be_options.min_scale_factor))
+    args.append('--max_scale_factor={}'.format(be_options.max_scale_factor))
+    args.append('--fine_tune_batch_norm={}'.format(
+        be_options.fine_tune_batch_norm))
+    args.append('--last_layers_contain_logits_only={}'.format(
+        be_options.last_layers_contain_logits_only))
 
     return args
 

src/rastervision/protos/deeplab/train.proto

@@ -11,12 +11,18 @@ message TrainingParameters {
         optional string train_split = 5 [default = "train"];
         required string model_variant = 6;
         repeated int32 atrous_rates = 7;
-        optional int32 output_stride = 8 [default = 1];
-        optional int32 decoder_output_stride = 9 [default = 1];
+        optional int32 output_stride = 8 [default = 8];
+        optional int32 decoder_output_stride = 9 [default = 8];
         repeated int32 train_crop_size = 10;
-        optional int32 train_batch_size = 11 [default = 1];
+        optional int32 train_batch_size = 11 [default = 8];
         optional string dataset = 12 [default = ""];
         optional int32 save_interval_secs = 13 [default = 600];
-        optional int32 save_summaries_secs = 14 [default = 30];
+        optional int32 save_summaries_secs = 14 [default = 5];
         optional bool save_summaries_images = 15 [default = true];
+        optional float last_layer_gradient_multiplier = 16 [default = 1.0];
+        optional bool initialize_last_layer = 17 [default = true];
+        optional float min_scale_factor = 18 [default = 1.0];
+        optional float max_scale_factor = 19 [default = 1.0];
+        optional bool fine_tune_batch_norm = 20 [default = true];
+        optional bool last_layers_contain_logits_only = 21 [default = false];
 }