Add some clearer error messages

lib/data/dataset.py

@@ -46,7 +46,7 @@ class DataSet(tordata.Dataset):
             data_list.append(_)
         for data in data_list:
             if len(data) != len(data_list[0]):
-                raise AssertionError
+                raise ValueError('Each input data should have the same length.')
 
         return data_list
 

lib/data/sampler.py

@@ -63,7 +63,7 @@ class InferenceSampler(tordata.sampler.Sampler):
         rank = dist.get_rank()
 
         if batch_size % world_size != 0:
-            raise AssertionError("World size({}) need be divisible by batch_size({})".format(
+            raise ValueError("World size({}) is not divisible by batch_size({})".format(
                 world_size, batch_size))
 
         if batch_size != 1:

lib/main.py

@@ -57,7 +57,7 @@ def run_model(cfgs, training):
 if __name__ == '__main__':
     torch.distributed.init_process_group('nccl', init_method='env://')
     if torch.distributed.get_world_size() != torch.cuda.device_count():
-        raise AssertionError("Expect number of availuable GPUs({}) equals to the world size({}).".format(
+        raise ValueError("Expect number of availuable GPUs({}) equals to the world size({}).".format(
             torch.distributed.get_world_size(), torch.cuda.device_count()))
     cfgs = config_loader(opt.cfgs)
     if opt.iter != 0:

lib/modeling/backbones/plain.py

@@ -43,7 +43,7 @@ class Plain(nn.Module):
             cfg = cfg.split('-')
             typ = cfg[0]
             if typ not in ['BC', 'FC']:
-                raise AssertionError
+                raise ValueError('Only support BC or FC, but got {}'.format(typ))
             out_c = int(cfg[1])
 
             if typ == 'BC':