Revert "Merge pull request #1 from John1231983/master"

This reverts commit 6881764, reversing
changes made to 0222d46.

bowl_config.py

@@ -11,29 +11,29 @@ class BowlConfig(Config):
     # Train on 1 GPU and 8 images per GPU. We can put multiple images on each
     # GPU because the images are small. Batch size is 8 (GPUs * images/GPU).
     GPU_COUNT = 1
-    IMAGES_PER_GPU = 2
+    IMAGES_PER_GPU = 1
 
     # Number of classes (including background)
     NUM_CLASSES = 1 + 1 # background + nuclei
 
     # Use small images for faster training. Set the limits of the small side
     # the large side, and that determines the image shape.
-    IMAGE_MIN_DIM = 512
-    IMAGE_MAX_DIM = 512
+    IMAGE_MIN_DIM = 256
+    IMAGE_MAX_DIM = 256
 
     # Use smaller anchors because our image and objects are small
     RPN_ANCHOR_SCALES = (8, 16, 32, 64, 128)  # anchor side in pixels
 
     # Reduce training ROIs per image because the images are small and have
     # few objects. Aim to allow ROI sampling to pick 33% positive ROIs.
-    TRAIN_ROIS_PER_IMAGE = 600
+    TRAIN_ROIS_PER_IMAGE = 1000
 
     STEPS_PER_EPOCH = None
 
     # use small validation steps since the epoch is small
     VALIDATION_STEPS = 5
 
-    USE_MINI_MASK = True
+    USE_MINI_MASK = False
 
 bowl_config = BowlConfig()
-bowl_config.display()
+bowl_config.display()

config.py

@@ -41,7 +41,7 @@ class Config(object):
     # Validation stats are also calculated at each epoch end and they
     # might take a while, so don't set this too small to avoid spending
     # a lot of time on validation stats.
-    STEPS_PER_EPOCH = 670
+    STEPS_PER_EPOCH = 1000
 
     # Number of validation steps to run at the end of every training epoch.
     # A bigger number improves accuracy of validation stats, but slows
@@ -87,8 +87,8 @@ class Config(object):
     # Images are resized such that the smallest side is >= IMAGE_MIN_DIM and
     # the longest side is <= IMAGE_MAX_DIM. In case both conditions can't
     # be satisfied together the IMAGE_MAX_DIM is enforced.
-    IMAGE_MIN_DIM = 512
-    IMAGE_MAX_DIM = 512
+    IMAGE_MIN_DIM = 800
+    IMAGE_MAX_DIM = 1024
     # If True, pad images with zeros such that they're (max_dim by max_dim)
     IMAGE_PADDING = True  # currently, the False option is not supported
 
@@ -111,14 +111,14 @@ class Config(object):
     MASK_SHAPE = [28, 28]
 
     # Maximum number of ground truth instances to use in one image
-    MAX_GT_INSTANCES = 256
+    MAX_GT_INSTANCES = 100
 
     # Bounding box refinement standard deviation for RPN and final detections.
     RPN_BBOX_STD_DEV = np.array([0.1, 0.1, 0.2, 0.2])
     BBOX_STD_DEV = np.array([0.1, 0.1, 0.2, 0.2])
 
     # Max number of final detections
-    DETECTION_MAX_INSTANCES = 512
+    DETECTION_MAX_INSTANCES = 100
 
     # Minimum probability value to accept a detected instance
     # ROIs below this threshold are skipped

inference.py

@@ -7,7 +7,7 @@
 from inference_config import inference_config
 from bowl_dataset import BowlDataset
 from utils import rle_encode, rle_decode, rle_to_string
-import functions as f
+
 ROOT_DIR = os.getcwd()
 MODEL_DIR = os.path.join(ROOT_DIR, "logs")
 
@@ -32,8 +32,7 @@
 
 output = []
 sample_submission = pd.read_csv('stage1_sample_submission.csv')
-ImageId = []
-EncodedPixels = []
+
 for image_id in tqdm(sample_submission.ImageId):
     image_path = os.path.join('stage1_test', image_id, 'images', image_id + '.png')
 
@@ -42,10 +41,25 @@
     r = results[0]
 
     masks = r['masks']
-    ImageId_batch, EncodedPixels_batch = f.numpy2encoding_no_overlap2(masks, image_id, r['scores'])
-    ImageId += ImageId_batch
-    EncodedPixels += EncodedPixels_batch
-
-
-
-f.write2csv('submission_v2.csv', ImageId, EncodedPixels)
+
+    count = masks.shape[-1]
+    occlusion = np.logical_not(masks[:, :, -1]).astype(np.uint8)
+
+    for i in range(count - 2, -1, -1):
+        mask = masks[:, :, i] * occlusion
+        mask_rle = rle_to_string(rle_encode(mask))
+
+        # Sanity check
+        try:
+            rle_decode(mask_rle, original_image.shape[:-1])
+            output.append([image_id, mask_rle])
+            occlusion = np.logical_and(occlusion, np.logical_not(masks[:, :, i]))
+
+        except Exception as e:
+            print(e)
+            print(image_id)
+            print('---')
+
+output_df = pd.DataFrame(output, columns=['ImageId', 'EncodedPixels'])
+output_df.to_csv('submission.csv', index=False, encoding='utf-8')
+

model.py

@@ -1816,7 +1816,7 @@ def build(self, mode, config):
         # Bottom-up Layers
         # Returns a list of the last layers of each stage, 5 in total.
         # Don't create the thead (stage 5), so we pick the 4th item in the list.
-        _, C2, C3, C4, C5 = resnet_graph(input_image, "resnet50", stage5=True)
+        _, C2, C3, C4, C5 = resnet_graph(input_image, "resnet101", stage5=True)
         # Top-down Layers
         # TODO: add assert to varify feature map sizes match what's in config
         P5 = KL.Conv2D(256, (1, 1), name='fpn_c5p5')(C5)

train.py

@@ -28,7 +28,7 @@
                           model_dir=MODEL_DIR)
 
 # Which weights to start with?
-init_with = "imagenet"  # imagenet, coco, or last
+init_with = "coco"  # imagenet, coco, or last
 
 if init_with == "imagenet":
     model.load_weights(model.get_imagenet_weights(), by_name=True)
@@ -57,12 +57,12 @@
 # Passing layers="heads" freezes all layers except the head
 # layers. You can also pass a regular expression to select
 # which layers to train by name pattern.
-#model.train(dataset_train, dataset_val, 
-#            learning_rate=bowl_config.LEARNING_RATE, 
-#            epochs=1, 
-#            layers='heads')
+model.train(dataset_train, dataset_val, 
+            learning_rate=bowl_config.LEARNING_RATE, 
+            epochs=1, 
+            layers='heads')
 
 model.train(dataset_train, dataset_val, 
             learning_rate=bowl_config.LEARNING_RATE / 10,
             epochs=100, 
-            layers="all")
+            layers="all")