update libs

libs/common/img_proc.py

@@ -177,18 +177,18 @@ def jitter_bbox_with_kpts_no_occlu(old_bbox, joints, parameters):
     Similar to the function above, but does not produce occluded joints
     """
     width, height = old_bbox[2] - old_bbox[0], old_bbox[3] - old_bbox[1]
-    old_center = [0.5*(old_bbox[0] + old_bbox[2]), 
-                  0.5*(old_bbox[1] + old_bbox[3])]
-    horizontal_scaling = parameters['scaling'][0]*np.random.rand() + 1
-    vertical_scaling = parameters['scaling'][1]*np.random.rand() + 1
-    horizontal_shift = 0.5*(horizontal_scaling-1)*width*np_random()
-    vertical_shift = 0.5*(vertical_scaling-1)*height*np_random()
+    old_center = [0.5 * (old_bbox[0] + old_bbox[2]), 
+                  0.5 * (old_bbox[1] + old_bbox[3])]
+    horizontal_scaling = parameters['scaling'][0] * np.random.rand() + 1
+    vertical_scaling = parameters['scaling'][1] * np.random.rand() + 1
+    horizontal_shift = 0.5 * (horizontal_scaling - 1) * width * np_random()
+    vertical_shift = 0.5 * (vertical_scaling - 1) * height * np_random()
     new_center = [old_center[0] + horizontal_shift,
                   old_center[1] + vertical_shift]
-    new_width = width*horizontal_scaling
-    new_height = height*vertical_scaling
-    new_bbox = [new_center[0] - 0.5*new_width, new_center[1] - 0.5*new_height,
-                new_center[0] + 0.5*new_width, new_center[1] + 0.5*new_height]
+    new_width = width * horizontal_scaling
+    new_height = height * vertical_scaling
+    new_bbox = [new_center[0] - 0.5 * new_width, new_center[1] - 0.5 * new_height,
+                new_center[0] + 0.5 * new_width, new_center[1] + 0.5 * new_height]
     return new_bbox, joints
 
 def generate_xy_map(bbox, resolution, global_size):

libs/dataset/KITTI/car_instance.py

@@ -264,7 +264,10 @@ def _prepare_key_points_custom(self, style, interp_params, vis_thresh=0.25):
     def _prepare_key_points(self, cfgs):
         self.kpts_style = cfgs['dataset']['2d_kpt_style']
         self._prepare_key_points_custom(self.kpts_style, cfgs['dataset']['interpolate'])
-        self.enlarge_factor = 1.1
+        if 'enlarge_factor' in cfgs['dataset']:
+            self.enlarge_factor = cfgs['dataset']['enlarge_factor']
+        else:
+            self.enlarge_factor = 1.1
         return
 
     def _save_cropped_instances(self):
@@ -1275,13 +1278,15 @@ def __getitem__(self, idx):
             parameters = self.hm_para
             parameters['boxes'] = self.annot_2dpose['boxes'][idx]
             # fs: fully-supervised ss: self-supervised
-            images_fs, heatmaps_fs, weights_fs, meta_fs = lip.get_tensor_from_img(img_path, 
-                                                                   parameters, 
-                                                                   joints=kpts,
-                                                                   pth_trans=self.pth_trans,
-                                                                   rf=parameters['rf'],
-                                                                   sf=parameters['sf'],
-                                                                   generate_hm=True)
+            images_fs, heatmaps_fs, weights_fs, meta_fs = \
+                lip.get_tensor_from_img(img_path, 
+                                        parameters, 
+                                        joints=kpts,
+                                        pth_trans=self.pth_trans,
+                                        rf=parameters['rf'],
+                                        sf=parameters['sf'],
+                                        generate_hm=True
+                                        )
             # use random unlabeled images for data augmentation
             if self.split == 'train' and hasattr(self, 'use_ss') and self.use_ss:
                 images_ss, heatmaps_ss, weights_ss, meta_ss = self.extract_ss_sample(len(images_fs))

libs/metric/criterions.py

@@ -180,13 +180,13 @@ def __init__(self, cfgs, num_joints=None):
         else:
             self.num_joints = cfgs['heatmapModel']['num_joints']
         self.image_size = cfgs['heatmapModel']['input_size']
-	if 'arg_max' in cfgs['testing_settings']:
+        if 'arg_max' in cfgs['testing_settings']:
             self.arg_max = cfgs['testing_settings']['arg_max']
-	else:
-	    self.arg_max = None
-        self.count = 0
-        self.mean = 0.
-        self.PCK_counts = np.zeros(len(PCK_THRES))
+        else:
+            self.arg_max = None
+            self.count = 0
+            self.mean = 0.
+            self.PCK_counts = np.zeros(len(PCK_THRES))
         return
 
     def update(self, prediction, meta_data, ground_truth=None, logger=None):

libs/model/heatmapModel/hrnet.py

@@ -421,7 +421,7 @@ def __init__(self, cfgs, **kwargs):
                 nn.Linear(256, 2)
                 )
         elif self.head_type == 'coordinates':
-            num_chan = 33
+            num_chan = self.num_joints
             self.head1 = nn.Sequential(
                 nn.Conv2d(
                     in_channels=pre_stage_channels[0],

tools/train_IGRs.py

@@ -75,6 +75,7 @@ def train(model, model_settings, GPUs, cfgs, logger, final_output_dir):
     data_cfgs = cfgs['dataset']
     train_dataset, valid_dataset = eval('dataset.' + data_cfgs['name'] + 
                                         '.car_instance').prepare_data(cfgs, logger)
+
     # get the optimizer and learning rate scheduler    
     optim, sche = optimizer.prepare_optim(model, cfgs)