u

cfgs/config.py

@@ -1 +1 @@
-from .config_normal import cfg
+from .config_debug_1600_vgg19 import cfg

cfgs/config_debug_1600_vgg19.py

@@ -0,0 +1,80 @@
+from easydict import EasyDict as edict
+
+_ = cfg = edict()
+_.train_ann = 'coco/annotations/person_keypoints_train2017.json'
+_.val_ann = 'coco/annotations/person_keypoints_val2017.json'
+
+
+_.train_images_dir = 'coco/train2017'
+_.val_images_dir = 'coco/val2017'
+_.test_images_dir = 'coco/test2017'
+
+_.train_masks_dir = 'coco/train2017_masks'
+_.val_masks_dir = 'coco/val2017_masks'
+
+_.train_labels_dir = 'coco/train2017_labels'
+_.val_labels_dir = 'coco/val2017_labels'
+
+_.img_y = 368
+_.img_x = 368
+
+# whether skip a person if the distance to existing person is too small
+_.skip_adj = False
+
+_.augmentation = True
+_.backbone_grad_scale = 0.25
+
+_.debug = True
+_.debug_sample_num = 1600
+
+# the default oder in coco annotation is:
+# 0:nose          1:left_eye         2:right_eye         3:left_ear
+# 4:right_ear     5:left_shoulder    6:right_shoulder    7:left_elbow
+# 8:right_elbow   9:left_wrist       10:right_wrist      11:left_hip
+# 12:right_hip    13:left_knee       14:right_knee       15:left_ankle
+# 16:right_ankle
+_.from_body_part = [0, 5, 6, 8, 10, 5, 7, 9, 12, 14, 16, 11, 13, 15, 2, 1, 4, 3]
+_.to_body_part =   [0, 6, 6, 8, 10, 5, 7, 9, 12, 14, 16, 11, 13, 15, 2, 1, 4, 3]
+# the oder after adjustment is:
+# 0:nose         1:neck           2:right_shoulder   3:right_elbow
+# 4:right_wrist  5:left_shoulder  6:left_elbow       7:left_wrist
+# 8:right_hip    9:right_knee     10:right_anckle    11:left_hip
+# 12:left_knee   13:left_ankle    14:right_eye       15:left_eye
+# 16:right_ear   17:left_ear
+
+# the limbs include:
+# 0:neck-->right_hip           1:right_hip-->right_knee      2:right_knee-->right_ankle      3:neck-->left_hip
+# 4:left_hip-->left_knee       5:left_knee-->left_ankle      6:neck-->right_shoulder         7:right_shoulder-->right_elbow
+# 8:right_elbow-->right_wrist  9:right_shoulder-->right_ear  10:neck-->left_shoulder         11:left_shoulder-->left_elbow
+# 12:left_elbow-->left_wrist   13:left_shoulder-->left_ear   14:neck-->nose                  15:nose-->right_eye
+# 16:nose-->left_eye           17:right_eye-->right_ear      18:left_eye-->left_ear
+_.limb_from = [1, 8,  9,  1, 11, 12, 1, 2, 3,  2, 1, 5, 6,  5, 1,  0,  0, 14, 15]
+_.limb_to =   [8, 9, 10, 11, 12, 13, 2, 3, 4, 16, 5, 6, 7, 17, 0, 14, 15, 16, 17]
+
+_.limb_seq = []
+_.map_idx = []
+for idx in range(len(_.limb_from)):
+    _.limb_seq.append([_.limb_from[idx], _.limb_to[idx]])
+    _.map_idx.append([2 * idx, 2 * idx + 1])
+
+_.stride = 8
+
+_.sigma = 7
+
+_.grid_y = int(_.img_y / _.stride)
+_.grid_x = int(_.img_x / _.stride)
+
+_.thre = 1
+_.thre1 = 0.1
+_.thre2 = 0.05
+
+_.ch_heats = 18 + 1 # 18个parts 第19个(遍历不会执行到)表示是非人体区域
+_.ch_vectors = 38
+_.stages = 6
+
+_.scale_search = [0.5, 1, 1.5, 2]
+_.pad_value = 128
+
+_.base_lr = 1e-5
+_.momentum = 0.9
+_.weight_decay = 5e-4

modules.py

@@ -5,7 +5,7 @@
 
 # @layer_register(log_shape=True)
 def VGGBlock_official(l):
-    with argscope(Conv2D, kernel_shape=3, nl=tf.nn.relu):
+    with argscope(Conv2D, kernel_shape=3, W_init=tf.random_normal_initializer(stddev=0.01), nl=tf.nn.relu):
         l = (LinearWrap(l)
              .Conv2D('conv1_1', 64)
              .Conv2D('conv1_2', 64)
@@ -28,7 +28,7 @@ def VGGBlock_official(l):
 
 # @layer_register(log_shape=True)
 def VGGBlock_ours(l):
-    with argscope(Conv2D, kernel_shape=3, nl=tf.nn.relu):
+    with argscope(Conv2D, kernel_shape=3, W_init=tf.random_normal_initializer(stddev=0.01), nl=tf.nn.relu):
         l = (LinearWrap(l)
              .Conv2D('conv1_1', 64)
              .Conv2D('conv1_2', 64)

predict.py

@@ -41,7 +41,8 @@ def predict(args):
                                    output_names = ['heatmaps', 'pafs'])
     predict_func = OfflinePredictor(predict_config)
 
-    img = cv2.imread(args.input_path)
+    img_bgr = cv2.imread(args.input_path)
+    img = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
 
     h, w, _ = img.shape
 
@@ -74,7 +75,7 @@ def predict(args):
 
     raw_heatmap_shown = np.maximum(0, heatmap_avg[:, :, 0:1] * 255)
     heatmap_shown = cv2.applyColorMap(raw_heatmap_shown.astype(np.uint8), cv2.COLORMAP_JET)
-    img_with_heatmap = cv2.addWeighted(heatmap_shown, 0.5, img, 0.5, 0)
+    img_with_heatmap = cv2.addWeighted(heatmap_shown, 0.5, img_bgr, 0.5, 0)
 
     cv2.imwrite('heatmap_shown.jpg', img_with_heatmap)
 
@@ -296,10 +297,8 @@ def predict(args):
 
 if __name__ == '__main__':
 
-    # img_id = 196283
-    # img_id = 163640
-    img_id = 785
-    img_path = os.path.join('coco/val2017', '%012d.jpg' % img_id)
+    img_id = 524311
+    img_path = os.path.join('coco/train2017', '%012d.jpg' % img_id)
 
     parser = argparse.ArgumentParser()
     parser.add_argument('--model_path', help='path of model', required = True)

reader.py

@@ -228,7 +228,7 @@ def get_data(self):
             # read img, mask, and label data
             img_path = os.path.join(self.images_dir, '%012d.jpg' % img_id)
             img = cv2.imread(img_path)
-            img = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
 
             mask_path = os.path.join(self.masks_dir, "mask_miss_%012d.png" % img_id)
             mask = cv2.imread(mask_path, 0)
@@ -358,8 +358,3 @@ def reset_state(self):
 
     g = ds.get_data()
     sample = next(g)
-    '''
-    for i in g:
-        img, heatmaps, pafs, mask_all = i
-        print(mask_all.shape)
-    '''

train.py

@@ -9,7 +9,7 @@
 from tensorpack.tfutils import optimizer, gradproc
 
 from cfgs.config import cfg
-from modules import VGGBlock_ours as VGGBlock, Stage1Block, StageTBlock
+from modules import VGGBlock_official as VGGBlock, Stage1Block, StageTBlock
 from reader import Data
 
 def apply_mask(t, mask):
@@ -26,7 +26,8 @@ def image_preprocess(image, bgr=True):
         if bgr == False:
             mean = mean[::-1]
         image_mean = tf.constant(mean, dtype=tf.float32)
-        image = image - image_mean
+        image = (image - image_mean) / (255 / 2)
+        # image = image - image_mean
         return image
 
 class Model(ModelDesc):
@@ -101,7 +102,7 @@ def _build_graph(self, inputs):
             wd_cost = tf.constant(0.0)
         loss1_total = tf.add_n(loss1_list)
         loss2_total = tf.add_n(loss2_list)
-        loss_total = loss1 + loss2
+        loss_total = loss1_total + loss2_total
         self.cost = tf.add_n([loss_total, wd_cost], name='cost')
 
 
@@ -113,9 +114,9 @@ def _build_graph(self, inputs):
 
         add_moving_summary(self.cost)
         for idx, loss1 in enumerate(loss1_list):
-            add_moving_summary(tf.identity(loss1_list[idx], name='stage%d_L1_loss' % idx))
+            add_moving_summary(tf.identity(loss1_list[idx], name='stage%d_L1_loss' % (idx+1)))
         for idx, loss2 in enumerate(loss2_list):
-            add_moving_summary(tf.identity(loss2_list[idx], name='stage%d_L2_loss' % idx))
+            add_moving_summary(tf.identity(loss2_list[idx], name='stage%d_L2_loss' % (idx+1)))
         add_moving_summary(tf.identity(loss1_total, name = 'L1_loss'))
         add_moving_summary(tf.identity(loss2_total, name = 'L2_loss'))
 
@@ -181,8 +182,8 @@ def get_config(args):
         dataflow = ds_train,
         callbacks = [
             ModelSaver(),
-            PeriodicTrigger(InferenceRunner(ds_val, [ScalarStats('cost')]),
-                            every_k_epochs=3),
+            # PeriodicTrigger(InferenceRunner(ds_val, [ScalarStats('cost')]),
+            #                 every_k_epochs=3),
             HumanHyperParamSetter('learning_rate'),
         ],
         model = Model(),