[MaskRCNN] Add GN in backbone

examples/FasterRCNN/basemodel.py

@@ -3,11 +3,11 @@
 
 from contextlib import contextmanager
 import tensorflow as tf
-from tensorpack.tfutils.argscope import argscope, get_arg_scope
+from tensorpack.tfutils import argscope
 from tensorpack.tfutils.scope_utils import auto_reuse_variable_scope
 from tensorpack.tfutils.varreplace import custom_getter_scope
 from tensorpack.models import (
-    Conv2D, MaxPooling, BatchNorm, BNReLU, layer_register)
+    Conv2D, MaxPooling, BatchNorm, layer_register)
 
 from config import config as cfg
 
@@ -56,9 +56,13 @@ def maybe_reverse_pad(topleft, bottomright):
 
 
 @contextmanager
-def resnet_argscope():
+def backbone_argscope():
+    def nonlin(x):
+        x = get_norm()(x)
+        return tf.nn.relu(x)
+
     with argscope([Conv2D, MaxPooling, BatchNorm], data_format='channels_first'), \
-            argscope(Conv2D, use_bias=False), \
+            argscope(Conv2D, use_bias=False, activation=nonlin), \
             argscope(BatchNorm, training=False), \
             custom_getter_scope(maybe_freeze_affine):
         yield
@@ -89,23 +93,23 @@ def image_preprocess(image, bgr=True):
         return image
 
 
-def get_bn(zero_init=False):
-    if zero_init:
-        return lambda x, name=None: BatchNorm('bn', x, gamma_init=tf.zeros_initializer())
+def get_norm(zero_init=False):
+    if cfg.BACKBONE.NORM == 'GN':
+        Norm = GroupNorm
+        layer_name = 'gn'
     else:
-        return lambda x, name=None: BatchNorm('bn', x)
+        Norm = BatchNorm
+        layer_name = 'bn'
+    return lambda x: Norm(layer_name, x, gamma_initializer=tf.zeros_initializer() if zero_init else None)
 
 
 def resnet_shortcut(l, n_out, stride, activation=tf.identity):
-    data_format = get_arg_scope()['Conv2D']['data_format']
-    n_in = l.get_shape().as_list()[1 if data_format in ['NCHW', 'channels_first'] else 3]
+    n_in = l.shape[1]
     if n_in != n_out:   # change dimension when channel is not the same
         # TF's SAME mode output ceil(x/stride), which is NOT what we want when x is odd and stride is 2
+        # In FPN mode, the images are pre-padded already.
         if not cfg.MODE_FPN and stride == 2:
             l = l[:, :, :-1, :-1]
-            return Conv2D('convshortcut', l, n_out, 1,
-                          strides=stride, padding='VALID', activation=activation)
-        else:
         return Conv2D('convshortcut', l, n_out, 1,
                       strides=stride, activation=activation)
     else:
@@ -117,17 +121,17 @@ def resnet_bottleneck(l, ch_out, stride):
     if cfg.BACKBONE.STRIDE_1X1:
         if stride == 2:
             l = l[:, :, :-1, :-1]
-        l = Conv2D('conv1', l, ch_out, 1, strides=stride, activation=BNReLU)
-        l = Conv2D('conv2', l, ch_out, 3, strides=1, activation=BNReLU)
+        l = Conv2D('conv1', l, ch_out, 1, strides=stride)
+        l = Conv2D('conv2', l, ch_out, 3, strides=1)
     else:
-        l = Conv2D('conv1', l, ch_out, 1, strides=1, activation=BNReLU)
+        l = Conv2D('conv1', l, ch_out, 1, strides=1)
         if stride == 2:
             l = tf.pad(l, [[0, 0], [0, 0], maybe_reverse_pad(0, 1), maybe_reverse_pad(0, 1)])
-            l = Conv2D('conv2', l, ch_out, 3, strides=2, activation=BNReLU, padding='VALID')
+            l = Conv2D('conv2', l, ch_out, 3, strides=2, padding='VALID')
         else:
-            l = Conv2D('conv2', l, ch_out, 3, strides=stride, activation=BNReLU)
-    l = Conv2D('conv3', l, ch_out * 4, 1, activation=get_bn(zero_init=True))
-    ret = l + resnet_shortcut(shortcut, ch_out * 4, stride, activation=get_bn(zero_init=False))
+            l = Conv2D('conv2', l, ch_out, 3, strides=stride)
+    l = Conv2D('conv3', l, ch_out * 4, 1, activation=get_norm(zero_init=True))
+    ret = l + resnet_shortcut(shortcut, ch_out * 4, stride, activation=get_norm(zero_init=False))
     return tf.nn.relu(ret, name='output')
 
 
@@ -141,9 +145,9 @@ def resnet_group(name, l, block_func, features, count, stride):
 
 def resnet_c4_backbone(image, num_blocks, freeze_c2=True):
     assert len(num_blocks) == 3
-    with resnet_argscope():
+    with backbone_argscope():
         l = tf.pad(image, [[0, 0], [0, 0], maybe_reverse_pad(2, 3), maybe_reverse_pad(2, 3)])
-        l = Conv2D('conv0', l, 64, 7, strides=2, activation=BNReLU, padding='VALID')
+        l = Conv2D('conv0', l, 64, 7, strides=2, padding='VALID')
         l = tf.pad(l, [[0, 0], [0, 0], maybe_reverse_pad(0, 1), maybe_reverse_pad(0, 1)])
         l = MaxPooling('pool0', l, 3, strides=2, padding='VALID')
         c2 = resnet_group('group0', l, resnet_bottleneck, 64, num_blocks[0], 1)
@@ -159,7 +163,7 @@ def resnet_c4_backbone(image, num_blocks, freeze_c2=True):
 
 @auto_reuse_variable_scope
 def resnet_conv5(image, num_block):
-    with resnet_argscope(), maybe_syncbn_scope():
+    with backbone_argscope(), maybe_syncbn_scope():
         l = resnet_group('group3', image, resnet_bottleneck, 512, num_block, 2)
         return l
 
@@ -170,7 +174,7 @@ def resnet_fpn_backbone(image, num_blocks, freeze_c2=True):
     new_shape2d = tf.to_int32(tf.ceil(tf.to_float(shape2d) / mult) * mult)
     pad_shape2d = new_shape2d - shape2d
     assert len(num_blocks) == 4, num_blocks
-    with resnet_argscope():
+    with backbone_argscope():
         chan = image.shape[1]
         pad_base = maybe_reverse_pad(2, 3)
         l = tf.pad(image, tf.stack(
@@ -178,7 +182,7 @@ def resnet_fpn_backbone(image, num_blocks, freeze_c2=True):
              [pad_base[0], pad_base[1] + pad_shape2d[0]],
              [pad_base[0], pad_base[1] + pad_shape2d[1]]]))
         l.set_shape([None, chan, None, None])
-        l = Conv2D('conv0', l, 64, 7, strides=2, activation=BNReLU, padding='VALID')
+        l = Conv2D('conv0', l, 64, 7, strides=2, padding='VALID')
         l = tf.pad(l, [[0, 0], [0, 0], maybe_reverse_pad(0, 1), maybe_reverse_pad(0, 1)])
         l = MaxPooling('pool0', l, 3, strides=2, padding='VALID')
         c2 = resnet_group('group0', l, resnet_bottleneck, 64, num_blocks[0], 1)

examples/FasterRCNN/config.py

@@ -132,7 +132,8 @@ _C.FPN.PROPOSAL_MODE = 'Level'  # 'Level', 'Joint'
 _C.FPN.NUM_CHANNEL = 256
 # conv head and fc head are only used in FPN.
 # For C4 models, the head is C5
-_C.FPN.FRCNN_HEAD_FUNC = 'fastrcnn_2fc_head'  # choices: fastrcnn_2fc_head, fastrcnn_4conv1fc_head
+_C.FPN.FRCNN_HEAD_FUNC = 'fastrcnn_2fc_head'
+# choices: fastrcnn_2fc_head, fastrcnn_4conv1fc_head, fastrcnn_4conv1fc_gn_head
 _C.FPN.FRCNN_CONV_HEAD_DIM = 256
 _C.FPN.FRCNN_FC_HEAD_DIM = 1024
 
@@ -152,7 +153,7 @@ def finalize_configs(is_training):
     """
     _C.DATA.NUM_CLASS = _C.DATA.NUM_CATEGORY + 1  # +1 background
 
-    assert _C.BACKBONE.NORM in ['FreezeBN', 'SyncBN'], _C.BACKBONE.NORM
+    assert _C.BACKBONE.NORM in ['FreezeBN', 'SyncBN', 'GN'], _C.BACKBONE.NORM
     if _C.BACKBONE.NORM != 'FreezeBN':
         assert not _C.BACKBONE.FREEZE_AFFINE
 

examples/FasterRCNN/model_fpn.py

@@ -56,7 +56,7 @@ def fpn_model(features):
                 lat_sum_5432.append(lat)
         p2345 = [Conv2D('posthoc_3x3_p{}'.format(i + 2), c, num_channel, 3)
                  for i, c in enumerate(lat_sum_5432[::-1])]
-        p6 = MaxPooling('maxpool_p6', p2345[-1], pool_size=1, strides=2, data_format='channels_first')
+        p6 = MaxPooling('maxpool_p6', p2345[-1], pool_size=1, strides=2, data_format='channels_first', padding='VALID')
         return p2345 + [p6]
 
 

examples/FasterRCNN/train.py

@@ -243,7 +243,7 @@ class ResNetC4Model(DetectionModel):
                 wd_cost], 'total_cost')
 
             add_moving_summary(total_cost, wd_cost)
-            return total_cost
+            return total_cost * (1. / cfg.TRAIN.NUM_GPUS)
         else:
             final_boxes, final_labels = self.fastrcnn_inference(
                 image_shape2d, rcnn_boxes, fastrcnn_label_logits, fastrcnn_box_logits)
@@ -379,7 +379,7 @@ class ResNetFPNModel(DetectionModel):
                                    mrcnn_loss, wd_cost], 'total_cost')
 
             add_moving_summary(total_cost, wd_cost)
-            return total_cost
+            return total_cost * (1. / cfg.TRAIN.NUM_GPUS)
         else:
             final_boxes, final_labels = self.fastrcnn_inference(
                 image_shape2d, rcnn_boxes, fastrcnn_label_logits, fastrcnn_box_logits)
@@ -595,8 +595,8 @@ if __name__ == '__main__':
         )
         if is_horovod:
             # horovod mode has the best speed for this model
-            trainer = HorovodTrainer()
+            trainer = HorovodTrainer(average=False)
         else:
             # nccl mode has better speed than cpu mode
-            trainer = SyncMultiGPUTrainerReplicated(cfg.TRAIN.NUM_GPUS, mode='nccl')
+            trainer = SyncMultiGPUTrainerReplicated(cfg.TRAIN.NUM_GPUS, average=False, mode='nccl')
         launch_train_with_config(traincfg, trainer)