[FasterRCNN] include anchors at the boundary

README.md

@@ -9,7 +9,7 @@ Tensorpack is a training interface based on TensorFlow.
 
 ## Features:
 
-It's Yet Another TF wrapper, but different in:
+It's Yet Another TF high-level API, with __speed__, __readability__ and __flexibility__ built together.
 
 1. Focus on __training speed__.
 	+	Speed comes for free with tensorpack -- it uses TensorFlow in the __efficient way__ with no extra overhead.
@@ -33,8 +33,9 @@ See [tutorials](http://tensorpack.readthedocs.io/en/latest/tutorial/index.html)
 ## [Examples](examples):
 
 Instead of showing you 10 random networks trained on toy datasets,
-[tensorpack examples](examples) faithfully replicate papers and care about performance.
-And everything runs on multiple GPUs. Some highlights:
+[tensorpack examples](examples) faithfully replicate papers and care about reproducing numbers,
+demonstrating its flexibility for actual research.
+Some highlights:
 
 ### Vision:
 + [Train ResNet](examples/ResNet) and [other models](examples/ImageNetModels) on ImageNet.

examples/FasterRCNN/data.py

@@ -34,7 +34,7 @@ def get_all_anchors(
     Get all anchors in the largest possible image, shifted, floatbox
 
     Returns:
-        anchors: SxSxNUM_ANCHORx4, where S == MAX_SIZE//STRIDE, floatbox
+        anchors: SxSxNUM_ANCHORx4, where S == ceil(MAX_SIZE/STRIDE), floatbox
         The layout in the NUM_ANCHOR dim is NUM_RATIO x NUM_SCALE.
 
     """
@@ -48,7 +48,7 @@ def get_all_anchors(
     # anchors are intbox here.
     # anchors at featuremap [0,0] are centered at fpcoor (8,8) (half of stride)
 
-    field_size = config.MAX_SIZE // stride
+    field_size = int(np.ceil(config.MAX_SIZE / stride))
     shifts = np.arange(0, field_size) * stride
     shift_x, shift_y = np.meshgrid(shifts, shifts)
     shift_x = shift_x.flatten()
@@ -154,7 +154,7 @@ def get_rpn_anchor_input(im, boxes, is_crowd):
 
     ALL_ANCHORS = get_all_anchors()
     H, W = im.shape[:2]
-    featureH, featureW = H // config.ANCHOR_STRIDE, W // config.ANCHOR_STRIDE
+    anchorH, anchorW = ALL_ANCHORS.shape[:2]
 
     def filter_box_inside(im, boxes):
         h, w = im.shape[:2]
@@ -169,8 +169,7 @@ def get_rpn_anchor_input(im, boxes, is_crowd):
     non_crowd_boxes = boxes[is_crowd == 0]
 
     # fHxfWxAx4
-    featuremap_anchors = ALL_ANCHORS[:featureH, :featureW, :, :]
-    featuremap_anchors_flatten = featuremap_anchors.reshape((-1, 4))
+    featuremap_anchors_flatten = np.copy(ALL_ANCHORS).reshape((-1, 4))
     # only use anchors inside the image
     inside_ind = filter_box_inside(im, featuremap_anchors_flatten)
     inside_anchors = featuremap_anchors_flatten[inside_ind, :]
@@ -178,12 +177,12 @@ def get_rpn_anchor_input(im, boxes, is_crowd):
     anchor_labels, anchor_boxes = get_anchor_labels(inside_anchors, non_crowd_boxes, crowd_boxes)
 
     # Fill them back to original size: fHxfWx1, fHxfWx4
-    featuremap_labels = -np.ones((featureH * featureW * config.NUM_ANCHOR, ), dtype='int32')
+    featuremap_labels = -np.ones((anchorH * anchorW * config.NUM_ANCHOR, ), dtype='int32')
     featuremap_labels[inside_ind] = anchor_labels
-    featuremap_labels = featuremap_labels.reshape((featureH, featureW, config.NUM_ANCHOR))
-    featuremap_boxes = np.zeros((featureH * featureW * config.NUM_ANCHOR, 4), dtype='float32')
+    featuremap_labels = featuremap_labels.reshape((anchorH, anchorW, config.NUM_ANCHOR))
+    featuremap_boxes = np.zeros((anchorH * anchorW * config.NUM_ANCHOR, 4), dtype='float32')
     featuremap_boxes[inside_ind, :] = anchor_boxes
-    featuremap_boxes = featuremap_boxes.reshape((featureH, featureW, config.NUM_ANCHOR, 4))
+    featuremap_boxes = featuremap_boxes.reshape((anchorH, anchorW, config.NUM_ANCHOR, 4))
     return featuremap_labels, featuremap_boxes
 
 

examples/FasterRCNN/model.py

@@ -298,7 +298,7 @@ def sample_fast_rcnn_targets(boxes, gt_boxes, gt_labels):
 
 
 @under_name_scope()
-def crop_and_resize(image, boxes, box_ind, crop_size):
+def crop_and_resize(image, boxes, box_ind, crop_size, pad_border=True):
     """
     Better-aligned version of tf.image.crop_and_resize, following our definition of floating point boxes.
 
@@ -312,9 +312,11 @@ def crop_and_resize(image, boxes, box_ind, crop_size):
     """
     assert isinstance(crop_size, int), crop_size
 
-    # TF's crop_and_resize fails on border
-    image = tf.pad(image, [[0, 0], [0, 0], [1, 1], [1, 1]])
-    boxes = boxes + 1
+    # TF's crop_and_resize produces zeros on border
+    if pad_border:
+        # this can be quite slow
+        image = tf.pad(image, [[0, 0], [0, 0], [1, 1], [1, 1]], mode='SYMMETRIC')
+        boxes = boxes + 1
 
     @under_name_scope()
     def transform_fpcoor_for_tf(boxes, image_shape, crop_shape):

examples/FasterRCNN/train.py

@@ -75,7 +75,7 @@ class Model(ModelDesc):
         image = image_preprocess(image, bgr=True)
         return tf.transpose(image, [0, 3, 1, 2])
 
-    def _get_anchors(self, image):
+    def _get_anchors(self, shape2d):
         """
         Returns:
             FSxFSxNAx4 anchors,
@@ -85,9 +85,7 @@ class Model(ModelDesc):
             all_anchors = tf.constant(get_all_anchors(), name='all_anchors', dtype=tf.float32)
             fm_anchors = tf.slice(
                 all_anchors, [0, 0, 0, 0], tf.stack([
-                    tf.shape(image)[0] // config.ANCHOR_STRIDE,
-                    tf.shape(image)[1] // config.ANCHOR_STRIDE,
-                    -1, -1]), name='fm_anchors')
+                    shape2d[0], shape2d[1], -1, -1]), name='fm_anchors')
             return fm_anchors
 
     def build_graph(self, *inputs):
@@ -96,14 +94,24 @@ class Model(ModelDesc):
             image, anchor_labels, anchor_boxes, gt_boxes, gt_labels, gt_masks = inputs
         else:
             image, anchor_labels, anchor_boxes, gt_boxes, gt_labels = inputs
-        fm_anchors = self._get_anchors(image)
         image = self._preprocess(image)     # 1CHW
-        image_shape2d = tf.shape(image)[2:]
 
-        anchor_boxes_encoded = encode_bbox_target(anchor_boxes, fm_anchors)
         featuremap = pretrained_resnet_conv4(image, config.RESNET_NUM_BLOCK[:3])
         rpn_label_logits, rpn_box_logits = rpn_head('rpn', featuremap, 1024, config.NUM_ANCHOR)
+        fm_shape = tf.shape(featuremap)[2:]    # h,w
+
+        fm_anchors = self._get_anchors(fm_shape)
+        anchor_labels = tf.slice(
+            anchor_labels, [0, 0, 0],
+            tf.stack([fm_shape[0], fm_shape[1], -1]),
+            name='sliced_anchor_labels')
+        anchor_boxes = tf.slice(
+            anchor_boxes, [0, 0, 0, 0],
+            tf.stack([fm_shape[0], fm_shape[1], -1, -1]),
+            name='sliced_anchor_boxes')
+        anchor_boxes_encoded = encode_bbox_target(anchor_boxes, fm_anchors)
 
+        image_shape2d = tf.shape(image)[2:]     # h,w
         decoded_boxes = decode_bbox_target(rpn_box_logits, fm_anchors)  # fHxfWxNAx4, floatbox
         proposal_boxes, proposal_scores = generate_rpn_proposals(
             tf.reshape(decoded_boxes, [-1, 4]),
@@ -170,7 +178,8 @@ class Model(ModelDesc):
                 target_masks_for_fg = crop_and_resize(
                     tf.expand_dims(gt_masks_for_fg, 1),
                     fg_sampled_boxes,
-                    tf.range(tf.size(fg_inds_wrt_gt)), 14)  # nfg x 1x14x14
+                    tf.range(tf.size(fg_inds_wrt_gt)), 14,
+                    pad_border=False)  # nfg x 1x14x14
                 target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1, 'sampled_fg_mask_targets')
                 mrcnn_loss = maskrcnn_loss(mask_logits, fg_labels, target_masks_for_fg)
             else: