tiny-cuda-nn
tiny-cuda-nn copied to clipboard
NVlabs
backward_backward_input_impl: not implemented error
When I use the tiny-cuda-nn to train a network, I need to take the derivative of one of the intermediate outputs of the network. When I backward the loss function loss.backward(), it reports an error:
RuntimeError: DifferentiableObject::backward_backward_input_impl: not implemented error
The code is as follows:
class Network(torch.nn.Module):
def __init__(self, num_dim=3, base_resolution=16, max_resolution=4096, feat_dim=15):
self.mlp_base = tcnn.NetworkWithInputEncoding(
n_input_dims=num_dim,
n_output_dims=1,
encoding_config={
"otype": "HashGrid",
"n_levels": n_levels,
"n_features_per_level": 2,
"log2_hashmap_size": log2_hashmap_size,
"base_resolution": base_resolution,
"per_level_scale": per_level_scale,
},
network_config={
"otype": "FullyFusedMLP",
"activation": "ReLU",
"output_activation": "None",
"n_neurons": 64,
"n_hidden_layers": 1,
},
)
def forward(self, input_xyz, input_dir=None, input_sun_dir=None):
with torch.enable_grad():
input_xyz.requires_grad_(True)
sigma = self.mlp_base(input_xyz, return_feat=True)
# compute norm
normal_map = torch.autograd.grad(outputs=sigma, inputs=input_xyz, grad_outputs=torch.ones_like(sigma, requires_grad=False),
retain_graph=True, create_graph=True)[0]
......
In the forward, I use normal_map = torch.autograd.grad(outputs=sigma, inputs=input_xyz, grad_outputs=torch.ones_like(sigma, requires_grad=False), retain_graph=True, create_graph=True)[0]
When I backward the loss function loss.backward(), it reports an error:
Traceback (most recent call last):
File "/home/user/Desktop/Nerf_Project/RS_Nerf/train_rsnerf_ngp.py", line 154, in <module>
loss.backward()
File "/home/user/anaconda3/lib/python3.9/site-packages/torch/_tensor.py", line 487, in backward
torch.autograd.backward(
File "/home/user/anaconda3/lib/python3.9/site-packages/torch/autograd/__init__.py", line 197, in backward
Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
File "/home/user/anaconda3/lib/python3.9/site-packages/torch/autograd/function.py", line 267, in apply
return user_fn(self, *args)
File "/home/user/anaconda3/lib/python3.9/site-packages/tinycudann/modules.py", line 136, in backward
doutput_grad, params_grad, input_grad = ctx.ctx_fwd.native_tcnn_module.bwd_bwd_input(
RuntimeError: DifferentiableObject::backward_backward_input_impl: not implemented error
I then replace the self.mlp_base with a simple network as below:
self.mlp_base = torch.nn.Sequential(nn.Linear(num_dim, 64), nn.ReLU(), nn.Linear(64, 64), nn.ReLU(), nn.Linear(64, 1 + self.feat_dim))
The code no longer reports errors as above. How do I fix this?
