[FasterRCNN] predict boxes inside the graph, and improve performance

examples/FasterRCNN/README.md

@@ -2,7 +2,7 @@
 This example aims to provide a minimal (<1000 lines) multi-GPU implementation of ResNet50-Faster-RCNN on COCO.
 
 ## Dependencies
-+ TensorFlow >= 1.4.0rc0
++ TensorFlow > 1.4.0 (use tf-nightly-gpu for now)
 + Install [pycocotools](https://github.com/pdollar/coco/tree/master/PythonAPI/pycocotools), OpenCV.
 + Pre-trained [ResNet50 model](https://goo.gl/6XjK9V) from tensorpack model zoo.
 + COCO data. It assumes the following directory structure:
@@ -46,10 +46,10 @@ To evaluate the performance (pretrained models can be downloaded in [model zoo](
 
 Mean Average Precision @IoU=0.50:0.95:
 
-+ trainval35k/minival, FASTRCNN_BATCH=256: 33.7. Takes 49h on 8 TitanX.
-+ trainval35k/minival, FASTRCNN_BATCH=64: 32.2. Takes 31h on 8 TitanX.
++ trainval35k/minival, FASTRCNN_BATCH=256: 34.2. Takes 49h on 8 TitanX.
++ trainval35k/minival, FASTRCNN_BATCH=64: 32.7. Takes 31h on 8 TitanX.
 
-The hyperparameters are not carefully tuned. You can probably get better performance by e.g.  training longer.
+The hyperparameters are not carefully tuned. You can probably get better performance by e.g. training longer.
 
 ## Notes
 

examples/FasterRCNN/eval.py

@@ -5,13 +5,13 @@
 import numpy as np
 import tqdm
 import cv2
+import six
 import os
-from collections import namedtuple
+from collections import namedtuple, defaultdict
 
 import tensorflow as tf
 from tensorpack.dataflow import MapDataComponent, TestDataSpeed
 from tensorpack.tfutils import get_default_sess_config
-from tensorpack.utils.argtools import memoized
 from tensorpack.utils.utils import get_tqdm_kwargs
 
 from pycocotools.coco import COCO
@@ -26,59 +26,6 @@ DetectionResult = namedtuple(
     ['class_id', 'boxes', 'scores'])
 
 
-@memoized
-def get_tf_nms(num_output, thresh):
-    """
-    Get a NMS callable.
-    """
-    # create a new graph for it
-    with tf.Graph().as_default(), tf.device('/cpu:0'):
-        boxes = tf.placeholder(tf.float32, shape=[None, 4])
-        scores = tf.placeholder(tf.float32, shape=[None])
-        indices = tf.image.non_max_suppression(
-            boxes, scores, num_output, thresh)
-        sess = tf.Session(config=get_default_sess_config())
-        return sess.make_callable(indices, [boxes, scores])
-
-
-def nms_fastrcnn_results(boxes, probs):
-    """
-    Args:
-        boxes: nx#catx4 floatbox in float32
-        probs: nx#class
-
-    Returns:
-        [DetectionResult]
-    """
-    C = probs.shape[1]
-    boxes = boxes.copy()
-
-    nms_func = get_tf_nms(config.RESULTS_PER_IM, config.FASTRCNN_NMS_THRESH)
-    ret = []
-    for klass in range(1, C):
-        ids = np.where(probs[:, klass] > config.RESULT_SCORE_THRESH)[0]
-        if ids.size == 0:
-            continue
-        probs_k = probs[ids, klass].flatten()
-        boxes_k = boxes[ids, klass - 1, :]
-        selected_ids = nms_func(boxes_k, probs_k)
-        selected_boxes = boxes_k[selected_ids, :].copy()
-        ret.append(DetectionResult(klass, selected_boxes, probs_k[selected_ids]))
-
-    if len(ret):
-        newret = []
-        all_scores = np.hstack([x.scores for x in ret])
-        if len(all_scores) > config.RESULTS_PER_IM:
-            score_thresh = np.sort(all_scores)[-config.RESULTS_PER_IM]
-            for klass, boxes, scores in ret:
-                keep_ids = np.where(scores >= score_thresh)[0]
-                if len(keep_ids):
-                    newret.append(DetectionResult(
-                        klass, boxes[keep_ids, :], scores[keep_ids]))
-            ret = newret
-    return ret
-
-
 def detect_one_image(img, model_func):
     """
     Run detection on one image, using the TF callable.
@@ -91,20 +38,33 @@ def detect_one_image(img, model_func):
     Returns:
         [DetectionResult]
     """
+
+    def group_results_by_class(boxes, probs, labels):
+        dic = defaultdict(list)
+        for box, prob, lab in zip(boxes, probs, labels):
+            dic[lab].append((box, prob))
+
+        def mapf(lab, values):
+            boxes = np.asarray([k[0] for k in values])
+            probs = np.asarray([k[1] for k in values])
+            return DetectionResult(lab, boxes, probs)
+
+        return [mapf(k, v) for k, v in six.iteritems(dic)]
+
     resizer = CustomResize(config.SHORT_EDGE_SIZE, config.MAX_SIZE)
     resized_img = resizer.augment(img)
     scale = (resized_img.shape[0] * 1.0 / img.shape[0] + resized_img.shape[1] * 1.0 / img.shape[1]) / 2
-    fg_probs, fg_boxes = model_func(resized_img)
-    fg_boxes = fg_boxes / scale
-    fg_boxes = clip_boxes(fg_boxes, img.shape[:2])
-    return nms_fastrcnn_results(fg_boxes, fg_probs)
+    boxes, probs, labels = model_func(resized_img)
+    boxes = boxes / scale
+    boxes = clip_boxes(boxes, img.shape[:2])
+    return group_results_by_class(boxes, probs, labels)
 
 
 def eval_on_dataflow(df, detect_func):
     """
     Args:
         df: a DataFlow which produces (image, image_id)
-        detect_func: a callable, takes [image] and returns a dict
+        detect_func: a callable, takes [image] and returns [DetectionResult]
 
     Returns:
         list of dict, to be dumped to COCO json format

examples/FasterRCNN/model.py

@@ -254,13 +254,13 @@ def sample_fast_rcnn_targets(boxes, gt_boxes, gt_labels):
     def sample_fg_bg(iou):
         fg_mask = tf.reduce_max(iou, axis=1) >= config.FASTRCNN_FG_THRESH
 
-        fg_inds = tf.where(fg_mask)[:, 0]
+        fg_inds = tf.reshape(tf.where(fg_mask), [-1])
         num_fg = tf.minimum(int(
             config.FASTRCNN_BATCH_PER_IM * config.FASTRCNN_FG_RATIO),
             tf.size(fg_inds), name='num_fg')
         fg_inds = tf.random_shuffle(fg_inds)[:num_fg]
 
-        bg_inds = tf.where(tf.logical_not(fg_mask))[:, 0]
+        bg_inds = tf.reshape(tf.where(tf.logical_not(fg_mask)), [-1])
         num_bg = tf.minimum(
             config.FASTRCNN_BATCH_PER_IM - num_fg,
             tf.size(bg_inds), name='num_bg')
@@ -383,27 +383,6 @@ def fastrcnn_head(feature, num_classes):
     return classification, box_regression
 
 
-@under_name_scope()
-def fastrcnn_predict_boxes(labels, box_logits):
-    """
-    Args:
-        labels: n,
-        box_logits: nx(C-1)x4
-
-    Returns:
-        fg_ind: fg, indices into n
-        fg_box_logits: fgx4
-
-    """
-    fg_ind = tf.reshape(tf.where(labels > 0), [-1])  # nfg,
-    fg_labels = tf.gather(labels, fg_ind)   # nfg,
-
-    ind_2d = tf.stack([fg_ind, fg_labels - 1], axis=1)  # nfgx2
-    # n x c-1 x 4 -> nfgx4
-    fg_box_logits = tf.gather_nd(box_logits, tf.stop_gradient(ind_2d))
-    return fg_ind, fg_box_logits
-
-
 @under_name_scope()
 def fastrcnn_losses(labels, label_logits, fg_boxes, fg_box_logits):
     """
@@ -442,3 +421,63 @@ def fastrcnn_losses(labels, label_logits, fg_boxes, fg_box_logits):
     for k in [label_loss, box_loss, accuracy, fg_accuracy, false_negative]:
         add_moving_summary(k)
     return label_loss, box_loss
+
+
+@under_name_scope()
+def fastrcnn_predictions(boxes, probs):
+    """
+    Generate final results from predictions of all proposals.
+
+    Args:
+        boxes: n#catx4 floatbox in float32
+        probs: nx#class
+    """
+    assert boxes.shape[1] == config.NUM_CLASS - 1
+    assert probs.shape[1] == config.NUM_CLASS
+    N = tf.shape(boxes)[0]
+    boxes = tf.transpose(boxes, [1, 0, 2])  # #catxnx4
+    probs = tf.transpose(probs[:, 1:], [1, 0])  # #catxn
+
+    def f(X):
+        """
+        prob: n probabilities
+        box: nx4 boxes
+
+        Returns: n boolean, the selection
+        """
+        prob, box = X
+        output_shape = tf.shape(prob)
+        # filter by score threshold
+        ids = tf.reshape(tf.where(prob > config.RESULT_SCORE_THRESH), [-1])
+        prob = tf.gather(prob, ids)
+        box = tf.gather(box, ids)
+        # NMS within each class
+        selection = tf.image.non_max_suppression(
+            box, prob, config.RESULTS_PER_IM, config.FASTRCNN_NMS_THRESH)
+        selection = tf.to_int32(tf.gather(ids, selection))
+        # sort available in TF>1.4.0
+        # selection = tf.contrib.framework.sort(selection, direction='ASCENDING')
+        sorted_selection, _ = tf.nn.top_k(-selection, k=tf.size(selection))
+        mask = tf.sparse_to_dense(
+            sparse_indices=-sorted_selection,
+            output_shape=output_shape,
+            sparse_values=True,
+            default_value=False)
+        return mask
+
+    masks = tf.map_fn(f, (probs, boxes), dtype=tf.bool,
+                      parallel_iterations=10)     # #cat x N
+    selected_indices = tf.where(masks)  # #selection x 2, each is (cat_id, box_id)
+    boxes = tf.boolean_mask(boxes, masks)   # #selection x 4
+    probs = tf.boolean_mask(probs, masks)
+    labels = selected_indices[:, 0] + 1
+
+    # filter again by sorting scores
+    topk_probs, topk_indices = tf.nn.top_k(
+        probs,
+        tf.minimum(config.RESULTS_PER_IM, tf.size(probs)),
+        sorted=False)
+    topk_probs = tf.identity(topk_probs, name='probs')
+    topk_boxes = tf.gather(boxes, topk_indices, name='boxes')
+    topk_labels = tf.gather(labels, topk_indices, name='labels')
+    return topk_boxes, topk_probs, topk_labels

examples/FasterRCNN/train.py

@@ -27,10 +27,10 @@ from coco import COCODetection
 from basemodel import (
     image_preprocess, pretrained_resnet_conv4, resnet_conv5)
 from model import (
-    rpn_head, rpn_losses,
     decode_bbox_target, encode_bbox_target,
+    rpn_head, rpn_losses,
     generate_rpn_proposals, sample_fast_rcnn_targets,
-    roi_align, fastrcnn_head, fastrcnn_losses, fastrcnn_predict_boxes)
+    roi_align, fastrcnn_head, fastrcnn_losses, fastrcnn_predictions)
 from data import (
     get_train_dataflow, get_eval_dataflow,
     get_all_anchors)
@@ -39,8 +39,7 @@ from viz import (
     draw_predictions, draw_final_outputs)
 from common import clip_boxes, CustomResize, print_config
 from eval import (
-    eval_on_dataflow, detect_one_image, print_evaluation_scores,
-    nms_fastrcnn_results)
+    eval_on_dataflow, detect_one_image, print_evaluation_scores)
 import config
 
 
@@ -112,7 +111,7 @@ class Model(ModelDesc):
         fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_head('fastrcnn', feature_fastrcnn, config.NUM_CLASS)
 
         if is_training:
-            fg_inds_wrt_sample = tf.where(rcnn_labels > 0)[:, 0]   # fg inds w.r.t all samples
+            fg_inds_wrt_sample = tf.reshape(tf.where(rcnn_labels > 0), [-1])   # fg inds w.r.t all samples
             fg_sampled_boxes = tf.gather(rcnn_sampled_boxes, fg_inds_wrt_sample)
 
             matched_gt_boxes = tf.gather(gt_boxes, fg_inds_wrt_gt)
@@ -143,6 +142,8 @@ class Model(ModelDesc):
                 tf.constant(config.FASTRCNN_BBOX_REG_WEIGHTS), anchors)
             decoded_boxes = tf.identity(decoded_boxes, name='fastrcnn_all_boxes')
 
+            pred_boxes, pred_probs, pred_labels = fastrcnn_predictions(decoded_boxes, label_probs)
+
     def _get_optimizer(self):
         lr = tf.get_variable('learning_rate', initializer=0.003, trainable=False)
         tf.summary.scalar('learning_rate', lr)
@@ -166,8 +167,9 @@ def visualize(model_path, nr_visualize=50, output_dir='output'):
             'generate_rpn_proposals/boxes',
             'generate_rpn_proposals/probs',
             'fastrcnn_all_probs',
-            'fastrcnn_fg_probs',
-            'fastrcnn_fg_boxes',
+            'fastrcnn_predictions/boxes',
+            'fastrcnn_predictions/probs',
+            'fastrcnn_predictions/labels',
         ]))
     df = get_train_dataflow()
     df.reset_state()
@@ -179,21 +181,21 @@ def visualize(model_path, nr_visualize=50, output_dir='output'):
         for idx, dp in itertools.islice(enumerate(df.get_data()), nr_visualize):
             img, _, _, gt_boxes, gt_labels = dp
 
-            rpn_boxes, rpn_scores, all_probs, fg_probs, fg_boxes = pred(img, gt_boxes, gt_labels)
+            rpn_boxes, rpn_scores, all_probs, \
+                final_boxes, final_probs, final_labels = pred(img, gt_boxes, gt_labels)
 
+            # draw groundtruth boxes
             gt_viz = draw_annotation(img, gt_boxes, gt_labels)
+            # draw best proposals for each groundtruth, to show recall
             proposal_viz, good_proposals_ind = draw_proposal_recall(img, rpn_boxes, rpn_scores, gt_boxes)
+            # draw the scores for the above proposals
             score_viz = draw_predictions(img, rpn_boxes[good_proposals_ind], all_probs[good_proposals_ind])
 
-            fg_boxes = clip_boxes(fg_boxes, img.shape[:2])
-            fg_viz = draw_predictions(img, fg_boxes, fg_probs)
-
-            results = nms_fastrcnn_results(fg_boxes, fg_probs)
-            final_viz = draw_final_outputs(img, results)
+            final_viz = draw_final_outputs(img, final_boxes, final_probs, final_labels)
 
             viz = tpviz.stack_patches([
-                gt_viz, proposal_viz, score_viz,
-                fg_viz, final_viz], 2, 3)
+                gt_viz, proposal_viz,
+                score_viz, final_viz], 2, 2)
 
             if os.environ.get('DISPLAY', None):
                 tpviz.interactive_imshow(viz)
@@ -207,8 +209,9 @@ def offline_evaluate(model_path, output_file):
         session_init=get_model_loader(model_path),
         input_names=['image'],
         output_names=[
-            'fastrcnn_all_probs',
-            'fastrcnn_all_boxes',
+            'fastrcnn_predictions/boxes',
+            'fastrcnn_predictions/probs',
+            'fastrcnn_predictions/labels',
         ]))
     df = get_eval_dataflow()
     df = PrefetchDataZMQ(df, 1)
@@ -224,8 +227,9 @@ def predict(model_path, input_file):
         session_init=get_model_loader(model_path),
         input_names=['image'],
         output_names=[
-            'fastrcnn_all_probs',
-            'fastrcnn_all_boxes',
+            'fastrcnn_predictions/boxes',
+            'fastrcnn_predictions/probs',
+            'fastrcnn_predictions/labels',
         ]))
     img = cv2.imread(input_file, cv2.IMREAD_COLOR)
     results = detect_one_image(img, pred)
@@ -237,7 +241,10 @@ def predict(model_path, input_file):
 class EvalCallback(Callback):
     def _setup_graph(self):
         self.pred = self.trainer.get_predictor(
-            ['image'], ['fastrcnn_all_probs', 'fastrcnn_all_boxes'])
+            ['image'],
+            ['fastrcnn_predictions/boxes',
+             'fastrcnn_predictions/probs',
+             'fastrcnn_predictions/labels'])
         self.df = PrefetchDataZMQ(get_eval_dataflow(), 1)
 
     def _before_train(self):

examples/FasterRCNN/viz.py

@@ -63,21 +63,19 @@ def draw_predictions(img, boxes, scores):
     return viz.draw_boxes(img, boxes, tags)
 
 
-def draw_final_outputs(img, results):
+def draw_final_outputs(img, final_boxes, final_probs, final_labels):
     """
     Args:
         results: [DetectionResult]
     """
-    all_boxes = []
-    all_tags = []
-    for class_id, boxes, scores in results:
-        all_boxes.extend(boxes)
-        all_tags.extend(
-            ["{},{:.2f}".format(COCOMeta.class_names[class_id], sc) for sc in scores])
-    all_boxes = np.asarray(all_boxes)
-    if all_boxes.shape[0] == 0:
+    if final_boxes.shape[0] == 0:
         return img
-    return viz.draw_boxes(img, all_boxes, all_tags)
+
+    tags = []
+    for prob, label in zip(final_probs, final_labels):
+        tags.append(
+            "{},{:.2f}".format(COCOMeta.class_names[label], prob))
+    return viz.draw_boxes(img, final_boxes, tags)
 
 
 def draw_mask(im, mask, alpha=0.5, color=None):