[FasterRCNN] clean-up segmentation code

examples/FasterRCNN/README.md

@@ -34,7 +34,7 @@ Otherwise, you probably need different hyperparameters for the same performance.
 
 To predict on an image (and show output in a window):
 ```
-./train.py --predict input.jpg
+./train.py --predict input.jpg --load /path/to/model
 ```
 
 To evaluate the performance (pretrained models can be downloaded in [model zoo](https://drive.google.com/open?id=1J0xuDAuyOWiuJRm2LfGoz5PUv9_dKuxq):

examples/FasterRCNN/coco.py

@@ -15,7 +15,6 @@ from tensorpack.utils.timer import timed_operation
 from tensorpack.utils.argtools import log_once
 
 from pycocotools.coco import COCO
-import pycocotools.mask as cocomask
 
 
 __all__ = ['COCODetection', 'COCOMeta']
@@ -146,8 +145,6 @@ class COCODetection(object):
                             log_once("Image {} has invalid polygons!".format(img['file_name']), 'warn')
 
                         obj['segmentation'] = valid_segs
-                    # rle = segmentation_to_rle(obj['segmentation'], height, width)
-                    # obj['mask_rle'] = rle
 
         # all geometrically-valid boxes are returned
         boxes = np.asarray([obj['bbox'] for obj in valid_objs], dtype='float32')  # (n, 4)
@@ -161,9 +158,7 @@ class COCODetection(object):
         img['class'] = cls          # n, always >0
         img['is_crowd'] = is_crowd  # n,
         if add_mask:
-            # mask_rles = [obj.pop('mask_rle') for obj in valid_objs]
-            # img['mask_rle'] = mask_rles    # list, each is an RLE with full-image coordinate
-            img['segmentation'] = [obj['segmentation'] for obj in valid_segs]
+            img['segmentation'] = [obj['segmentation'] for obj in valid_objs]
 
         del objs
 
@@ -198,22 +193,6 @@ class COCODetection(object):
         return ret
 
 
-def segmentation_to_rle(segm, height, width):
-    if isinstance(segm, list):
-        # polygon -- a single object might consist of multiple parts
-        # we merge all parts into one mask rle code
-        rles = cocomask.frPyObjects(segm, height, width)
-        rle = cocomask.merge(rles)
-    elif isinstance(segm['counts'], list):
-        # uncompressed RLE
-        rle = cocomask.frPyObjects(segm, height, width)
-    else:
-        print("WTF?")
-        import IPython as IP
-        IP.embed()
-    return rle
-
-
 if __name__ == '__main__':
     c = COCODetection('/home/wyx/data/coco', 'train2014')
     gt_boxes = c.load(add_gt=True, add_mask=True)

examples/FasterRCNN/data.py

@@ -10,8 +10,8 @@ import logging
 from tensorpack.utils import logger
 from tensorpack.utils.argtools import memoized, log_once
 from tensorpack.dataflow import (
-    ProxyDataFlow, MapData, imgaug, TestDataSpeed,
-    MapDataComponent)
+    MapData, imgaug, TestDataSpeed,
+    MapDataComponent, DataFromList)
 import tensorpack.utils.viz as tpviz
 from tensorpack.utils.viz import interactive_imshow
 
@@ -198,17 +198,14 @@ def get_train_dataflow():
     # But this filter shall not be applied for testing.
     imgs = list(filter(lambda img: len(img['boxes']) > 0, imgs))    # log invalid training
 
-    ds = DataFromListOfDict(
-        imgs,
-        ['file_name', 'boxes', 'class', 'is_crowd'],  # we need this four keys only
-        shuffle=True)
+    ds = DataFromList(imgs, shuffle=True)
 
     aug = imgaug.AugmentorList(
         [CustomResize(config.SHORT_EDGE_SIZE, config.MAX_SIZE),
          imgaug.Flip(horiz=True)])
 
-    def preprocess(dp):
-        fname, boxes, klass, is_crowd = dp
+    def preprocess(img):
+        fname, boxes, klass, is_crowd = img['file_name'], img['boxes'], img['class'], img['is_crowd']
         im = cv2.imread(fname, cv2.IMREAD_COLOR)
         assert im is not None, fname
         im = im.astype('float32')
@@ -252,6 +249,8 @@ def get_eval_dataflow():
 
 
 if __name__ == '__main__':
+    config.BASEDIR = '/home/wyx/data/coco'
+    config.TRAIN_DATASET = ['train2014']
     from tensorpack.dataflow import PrintData
     ds = get_train_dataflow()
     ds = PrintData(ds, 100)