training works just slow

examples/train_ngp_nerf_proposal.py

@@ -204,7 +204,7 @@ if __name__ == "__main__":
         subject_id=args.scene,
         root_fp=data_root_fp,
         split=args.train_split,
-        num_rays=target_sample_batch_size // render_n_samples,
+        num_rays=target_sample_batch_size // 32,
         **train_dataset_kwargs,
     )
 
@@ -262,7 +262,7 @@ if __name__ == "__main__":
                 hidden_dim=16,
                 max_res=64,
                 geo_feat_dim=0,
-                n_levels=5,
+                n_levels=2,
                 log2_hashmap_size=17,
             ),
             # NGPradianceField(

nerfacc/ray_marching.py

@@ -195,7 +195,7 @@ def ray_marching(
     if proposal_nets is not None:
         proposal_sample_list = []
         # resample with proposal nets
-        for net, num_samples in zip(proposal_nets, [48]):
+        for net, num_samples in zip(proposal_nets, [32]):
             ray_indices = unpack_info(packed_info)
             with torch.enable_grad():
                 sigmas = sigma_fn(t_starts, t_ends, ray_indices.long(), net=net)