Support groups in de-conv and fix bug. (#939)

* Support groups in de-conv and fix bug.

* Clean code.

fluid/face_detction/image_util.py

@@ -75,11 +75,8 @@ def satisfy_sample_constraint(sampler, sample_bbox, bbox_labels):
     if sampler.min_jaccard_overlap == 0 and sampler.max_jaccard_overlap == 0:
         return True
     for i in range(len(bbox_labels)):
-        object_bbox = bbox(
-            bbox_labels[i][0],
-            bbox_labels[i][1],  # tangxu @ 2018-05-17
-            bbox_labels[i][2],
-            bbox_labels[i][3])
+        object_bbox = bbox(bbox_labels[i][0], bbox_labels[i][1],
+                           bbox_labels[i][2], bbox_labels[i][3])
         overlap = jaccard_overlap(sample_bbox, object_bbox)
         if sampler.min_jaccard_overlap != 0 and \
                 overlap < sampler.min_jaccard_overlap:

fluid/face_detction/pyramidbox.py

@@ -82,8 +82,7 @@ class PyramidBox(object):
         self.conv5 = conv_block(self.conv4, 3, [512] * 3, [3] * 3)
 
         # fc6 and fc7 in paper, priorbox min_size is 128
-        self.conv6 = conv_block(
-            self.conv5, 2, [1024, 1024], [3, 1], with_pool=False)
+        self.conv6 = conv_block(self.conv5, 2, [1024, 1024], [3, 1])
         # conv6_1 and conv6_2 in paper, priorbox min_size is 256
         self.conv7 = conv_block(
             self.conv6, 2, [256, 512], [1, 3], [1, 2], with_pool=False)
@@ -101,9 +100,15 @@ class PyramidBox(object):
             b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
             conv1 = fluid.layers.conv2d(
                 up_from, ch, 1, act='relu', bias_attr=b_attr)
-            # TODO: add group
             conv_trans = fluid.layers.conv2d_transpose(
-                conv1, ch, None, 4, 1, 2, bias_attr=False)
+                conv1,
+                ch,
+                output_size=None,
+                filter_size=4,
+                padding=1,
+                stride=2,
+                groups=ch,
+                bias_attr=False)
             b_attr = ParamAttr(learning_rate=2., regularizer=L2Decay(0.))
             conv2 = fluid.layers.conv2d(
                 up_to, ch, 1, act='relu', bias_attr=b_attr)
@@ -235,8 +240,7 @@ class PyramidBox(object):
         self.prior_boxes = fluid.layers.concat(boxes)
         self.box_vars = fluid.layers.concat(vars)
 
-    def vgg_ssd(self, num_classes, image_shape):  # tangxu
-
+    def vgg_ssd(self, num_classes, image_shape):
         self.conv3_norm = self._l2_norm_scale(self.conv3)
         self.conv4_norm = self._l2_norm_scale(self.conv4)
         self.conv5_norm = self._l2_norm_scale(self.conv5)
@@ -275,7 +279,10 @@ class PyramidBox(object):
         head_loss = fluid.layers.ssd_loss(
             self.head_mbox_loc, self.head_mbox_conf, self.gt_box, self.gt_label,
             self.prior_boxes, self.box_vars)
-        return face_loss, head_loss
+        face_loss = fluid.layers.reduce_sum(face_loss)
+        head_loss = fluid.layers.reduce_sum(head_loss)
+        total_loss = face_loss + head_loss
+        return face_loss, head_loss, total_loss
 
     def test(self):
         test_program = fluid.default_main_program().clone(for_test=True)

fluid/face_detction/train.py

@@ -19,6 +19,7 @@ add_arg('learning_rate', float, 0.0001, "Learning rate.")
 add_arg('batch_size', int, 16, "Minibatch size.")
 add_arg('num_passes', int, 120, "Epoch number.")
 add_arg('use_gpu', bool, True, "Whether use GPU.")
+add_arg('use_pyramidbox', bool, False, "Whether use PyramidBox model.")
 add_arg('dataset', str, 'WIDERFACE', "coco2014, coco2017, and pascalvoc.")
 add_arg('model_save_dir', str, 'model', "The path to save model.")
 add_arg('pretrained_model', str, './vgg_model/', "The init model path.")
@@ -37,8 +38,12 @@ def train(args, data_args, learning_rate, batch_size, pretrained_model,
 
     image_shape = [3, data_args.resize_h, data_args.resize_w]
 
-    network = PyramidBox(image_shape)
-    loss = network.vgg_ssd(num_classes, image_shape)
+    if args.use_pyramidbox:
+        network = PyramidBox(image_shape, sub_network=args.use_pyramidbox)
+        face_loss, head_loss, loss = network.train()
+    else:
+        network = PyramidBox(image_shape, sub_network=args.use_pyramidbox)
+        loss = network.vgg_ssd(num_classes, image_shape)
 
     epocs = 12880 / batch_size
     boundaries = [epocs * 100, epocs * 125, epocs * 150]
@@ -46,6 +51,7 @@ def train(args, data_args, learning_rate, batch_size, pretrained_model,
         learning_rate, learning_rate * 0.1, learning_rate * 0.01,
         learning_rate * 0.001
     ]
+    #print('main program ', fluid.default_main_program())
 
     optimizer = fluid.optimizer.RMSProp(
         learning_rate=fluid.layers.piecewise_decay(boundaries, values),
@@ -60,10 +66,8 @@ def train(args, data_args, learning_rate, batch_size, pretrained_model,
 
     # fluid.io.save_inference_model('./vgg_model/', ['image'], [loss], exe)
     if pretrained_model:
-
         def if_exist(var):
             return os.path.exists(os.path.join(pretrained_model, var.name))
-
         print('Load pre-trained model.')
         fluid.io.load_vars(exe, pretrained_model, predicate=if_exist)
 
@@ -108,7 +112,6 @@ def train(args, data_args, learning_rate, batch_size, pretrained_model,
             if batch_id % 1 == 0:
                 print("Pass {0}, batch {1}, loss {2}, time {3}".format(
                     pass_id, batch_id, loss_v, start_time - prev_start_time))
-        test(pass_id, best_map)
         if pass_id % 10 == 0 or pass_id == num_passes - 1:
             save_model(str(pass_id))
     print("Best test map {0}".format(best_map))