change dark_w32_384x288 batchsize to 32; modify np.pad to support lower numpy version (#3788)

configs/keypoint/hrnet/dark_hrnet_w32_384x288.yml

@@ -111,7 +111,7 @@ TrainReader:
         std: *global_std
         is_scale: true
     - Permute: {}
-  batch_size: 64
+  batch_size: 32
   shuffle: true
   drop_last: false
 

deploy/python/keypoint_postprocess.py

@@ -106,6 +106,7 @@ class HrHRNetPostProcess(object):
             # pad the cost matrix, otherwise new pose are ignored
             if num_valid > num_clusters:
                 cost = np.pad(cost, ((0, 0), (0, num_valid - num_clusters)),
+                              'constant',
                               constant_values=((0, 0), (0, 1e-10)))
             rows, cols = linear_sum_assignment(cost)
             for y, x in zip(rows, cols):

ppdet/modeling/architectures/keypoint_hrhrnet.py

@@ -35,7 +35,7 @@ class HigherHRNet(BaseArch):
 
     def __init__(self,
                  backbone='HRNet',
-                 hrhrnet_head='HigherHRNetHead',
+                 hrhrnet_head='HrHRNetHead',
                  post_process='HrHRNetPostProcess',
                  eval_flip=True,
                  flip_perm=None,
@@ -224,6 +224,7 @@ class HrHRNetPostProcess(object):
             # pad the cost matrix, otherwise new pose are ignored
             if num_valid > num_clusters:
                 cost = np.pad(cost, ((0, 0), (0, num_valid - num_clusters)),
+                              'constant',
                               constant_values=((0, 0), (0, 1e-10)))
             rows, cols = linear_sum_assignment(cost)
             for y, x in zip(rows, cols):