Seems that torch2trt doesn't support input of list of Tensor about torch2trt HOT 6 CLOSED

And then I modified my model with input as :

def forward(self, x1, x2, x3, x4, x5):

And still run:
model_trt = torch2trt(model, input),
here, input is a list of 5 Tensors.

Then when I want to save the model, it gives:

[TensorRT] ERROR: Unused Input: input_0
[TensorRT] ERROR: Unused Input: input_1
[TensorRT] ERROR: Unused Input: input_2
[TensorRT] ERROR: Unused Input: input_3
Traceback (most recent call last):
File "test_torch2trt.py", line 174, in
main(config)
File "test_torch2trt.py", line 68, in main
torch.save(model_resnet_ppm_trt.state_dict(), 'model_resnet_ppm.pth')
File "/data/mawenxin/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 636, in state_dict
hook_result = hook(self, destination, prefix, local_metadata)
File "/data/mawenxin/anaconda3/lib/python3.7/site-packages/torch2trt/torch2trt.py", line 191, in _on_state_dict
state_dict[prefix + 'engine'] = bytes(self.engine.serialize())
AttributeError: 'NoneType' object has no attribute 'serialize'

from torch2trt.

Running into the same issue. Did you figure this out?

from torch2trt.

I am also facing the same problem. Anyone can elaborate?

from torch2trt.

Hi,

Here is errors in my case:

In function call torch2trt:

engine = builder.build_cuda_engine(network)

[TensorRT] ERROR: Unused Input: input_0
[TensorRT] ERROR: Unused Input: input_1

from torch2trt.

Hi,

Here is errors in my case:

In function call torch2trt:

engine = builder.build_cuda_engine(network)

[TensorRT] ERROR: Unused Input: input_0
[TensorRT] ERROR: Unused Input: input_1

Same issue QQ

from torch2trt.

Hi,

I faced same issue and figured out after great struggle. In my case, i was having issue at group normalization. There were two implementations and my code was pointing to incomplete implementation of group normalization code. After fixing this code the invalid input issue ran away.

Thanks,
Muhammad Ajmal Siddiqui

from torch2trt.